id
stringlengths
33
33
doc_url
stringlengths
33
33
title
stringlengths
13
211
publication_date
timestamp[s]
update_date
timestamp[s]
authors
stringlengths
6
16.6k
category_primary
stringclasses
60 values
category_all
stringlengths
5
132
abstract
stringlengths
123
1.99k
text
stringlengths
190
2.08k
embeddings
sequencelengths
768
768
http://arxiv.org/abs/2406.18312v1
http://arxiv.org/pdf/2406.18312v1
AI-native Memory: A Pathway from LLMs Towards AGI
2024-06-26T12:51:37
2024-06-26T12:51:37
Jingbo Shang, Zai Zheng, Xiang Ying, Felix Tao, Mindverse Team
cs.CL
cs.CL, cs.AI
Large language models (LLMs) have demonstrated the world with the sparks of artificial general intelligence (AGI). One opinion, especially from some startups working on LLMs, argues that an LLM with nearly unlimited context length can realize AGI. However, they might be too optimistic about the long-context capability of (existing) LLMs -- (1) Recent literature has shown that their effective context length is significantly smaller than their claimed context length; and (2) Our reasoning-in-a-haystack experiments further demonstrate that simultaneously finding the relevant information from a long context and conducting (simple) reasoning is nearly impossible. In this paper, we envision a pathway from LLMs to AGI through the integration of \emph{memory}. We believe that AGI should be a system where LLMs serve as core processors. In addition to raw data, the memory in this system would store a large number of important conclusions derived from reasoning processes. Compared with retrieval-augmented generation (RAG) that merely processing raw data, this approach not only connects semantically related information closer, but also simplifies complex inferences at the time of querying. As an intermediate stage, the memory will likely be in the form of natural language descriptions, which can be directly consumed by users too. Ultimately, every agent/person should have its own large personal model, a deep neural network model (thus \emph{AI-native}) that parameterizes and compresses all types of memory, even the ones cannot be described by natural languages. Finally, we discuss the significant potential of AI-native memory as the transformative infrastructure for (proactive) engagement, personalization, distribution, and social in the AGI era, as well as the incurred privacy and security challenges with preliminary solutions.
AI-native Memory: A Pathway from LLMs Towards AGI: Large language models (LLMs) have demonstrated the world with the sparks of artificial general intelligence (AGI). One opinion, especially from some startups working on LLMs, argues that an LLM with nearly unlimited context length can realize AGI. However, they might be too optimistic about the long-context capability of (existing) LLMs -- (1) Recent literature has shown that their effective context length is significantly smaller than their claimed context length; and (2) Our reasoning-in-a-haystack experiments further demonstrate that simultaneously finding the relevant information from a long context and conducting (simple) reasoning is nearly impossible. In this paper, we envision a pathway from LLMs to AGI through the integration of \emph{memory}. We believe that AGI should be a system where LLMs serve as core processors. In addition to raw data, the memory in this system would store a large number of important conclusions derived from reasoning processes. Compared with retrieval-augmented generation (RAG) that merely processing raw data, this approach not only connects semantically related information closer, but also simplifies complex inferences at the time of querying. As an intermediate stage, the memory will likely be in the form of natural language descriptions, which can be directly consumed by users too. Ultimately, every agent/person should have its own large personal model, a deep neural network model (thus \emph{AI-native}) that parameterizes and compresses all types of memory, even the ones cannot be described by natural languages. Finally, we discuss the significant potential of AI-native memory as the transformative infrastructure for (proactive) engagement, personalization, distribution, and social in the AGI era, as well as the incurred privacy and security challenges with preliminary solutions.
[ 0.0641116574, 0.0580641367, -0.0434283502, 0.0353890024, -0.0346714519, -0.0207084864, 0.0206467733, -0.0275378823, 0.0534948967, -0.0236995779, -0.0092641339, -0.0440726206, -0.014428284, 0.0521501824, -0.0002601421, -0.013085587, 0.0365698226, -0.0146437651, 0.0346792452, -0.00170614, -0.0176900066, 0.018166611, 0.030671712, 0.0311543383, -0.0303506963, 0.0007230401, -0.0356683619, -0.0109077059, 0.003849112, -0.0301492792, -0.0321442783, -0.0025736939, 0.0457127914, 0.0325033031, 0.0000023685, -0.0813021436, 0.0030223676, -0.0019789839, -0.0508481227, -0.0159600191, 0.0120483888, 0.0531761348, 0.0057960236, -0.0119539872, -0.0603862554, -0.0149774235, 0.0376050323, 0.0133817038, 0.0166366845, 0.0466303378, -0.0153218666, -0.0787590891, 0.0267626643, -0.0240180604, 0.0495152324, -0.0721336454, 0.012881388, -0.0551332608, 0.0188581496, 0.0015295015, 0.0050878879, 0.0372743309, 0.0514789037, 0.0076353718, 0.0182089116, 0.0159432366, -0.0133721502, -0.0185049959, -0.0089951791, 0.0435839668, 0.0466782972, -0.0130118625, 0.0180559102, 0.0280374624, 0.0253848638, 0.0075168018, -0.0219669603, 0.0764293, -0.0451057777, 0.0050437413, 0.0224982165, 0.0101529509, 0.0365871042, 0.0326884463, -0.0544601567, 0.0742663145, -0.0207518358, -0.0236692652, 0.0063173044, -0.0120621743, 0.0123048862, -0.0453786142, 0.0306069627, 0.0686870366, -0.0053042881, 0.0302097797, -0.0463014729, -0.0040589268, 0.0087801293, -0.044767499, 0.0058607743, 0.0261398535, -0.0146641955, 0.0453122891, -0.000755319, 0.0305930935, -0.0351804495, 0.0000520962, -0.0259332359, -0.0506874584, -0.0438615307, -0.0228284579, -0.0316976644, 0.0985480845, -0.0254338142, -0.023420142, -0.0376705341, 0.0584261417, -0.0231771301, 0.0236543734, -0.0483692065, 0.0531346351, -0.0304386802, -0.0052793482, -0.0676364154, -0.0337598957, -0.0800259039, 0.0013248718, 0.0338089727, -0.0667266026, 0.0171876568, -0.0011703178, 0.0275350437, -0.0050179609, 0.0092746411, 0.0766500905, 0.0229496341, -0.04004886, -0.0835019425, -0.0706228912, 0.0171020254, -0.0743462294, 0.0037983956, -0.0485275351, 0.0353846401, 0.0068902969, -0.0279594269, -0.071605213, -0.00930099, 0.0304063447, -0.0420771949, 0.0327042677, 0.0139571624, -0.0005166855, 0.0602062121, 0.0433039367, 0.0501229167, 0.0121716568, 0.0353671722, 0.0557857044, 0.0514923781, 0.019885499, 0.0227322355, -0.0152717512, 0.0213842299, -0.0325807333, -0.0127873141, 0.0108188819, -0.0404463261, -0.0041286135, -0.0192653481, 0.0648584813, 0.001838588, -0.0188923832, 0.0481692702, 0.1205351204, -0.012953124, 0.0391781181, 0.0480351597, 0.0004535777, 0.0017695784, -0.0162784047, -0.0388930961, 0.0285214279, -0.0589337423, -0.0107091917, 0.0410960577, 0.0305866916, -0.0432227924, -0.0351092741, 0.018612247, -0.0346503295, 0.0458635464, 0.0498153083, 0.0207831841, 0.0074743773, 0.0122078257, -0.0125073278, -0.0041458826, -0.0474612489, 0.0132128354, -0.0783535168, 0.042136997, 0.0956930295, 0.0330963656, -0.0170640238, 0.0009931847, 0.0321344733, -0.0286133345, 0.0314170048, 0.0235896148, 0.0070922673, 0.0128314449, -0.0283208005, -0.0000247423, -0.0312928408, 0.0138008455, 0.0385210626, 0.0161853973, 0.026811637, 0.0266952217, 0.0111058662, 0.0104482016, -0.0154327517, -0.0318153128, -0.0259596668, -0.0038236727, 0.0235150456, 0.0227256007, 0.0409072191, 0.0159669593, 0.0086920904, 0.0332180224, -0.0209805481, -0.0351352505, -0.0018176644, -0.0477313735, 0.0641432032, -0.0024199903, -0.0037550931, 0.034272857, -0.0183357522, -0.0019714267, -0.0570226349, 0.0366269313, -0.0383047089, 0.0570017397, -0.044259198, 0.0035666553, -0.0085296268, 0.0154901771, -0.0146864541, 0.0829657838, -0.0111568244, -0.0643283427, -0.0043145274, 0.0379964449, 0.0272487737, -0.0645878837, -0.0221979953, 0.0075913775, -0.0029563769, 0.0157335252, 0.0542680584, -0.0197264161, -0.0398119576, 0.0169711336, -0.0917288288, 0.0189303178, -0.0027993659, -0.009001215, -0.0171835292, 0.0458377972, 0.0170412604, -0.0115989251, -0.0651669651, 0.0353084467, 0.014816843, 0.02583028, 0.0302935261, 0.0498639792, -0.002090252, -0.0294163655, 0.0002606845, -0.0102228234, 0.0274398308, 0.0217133984, -0.0257595312, -0.0208562668, 0.0098873628, -0.0665372461, 0.0355703793, -0.0180258546, -0.0087488461, 0.0201859009, -0.0399834141, 0.0132874968, 0.1092118472, 0.0046247859, 0.0468138941, -0.0022191431, -0.0126229003, -0.0412287153, 0.0110331131, -0.0131838424, -0.0251405276, -0.0243055634, 0.0095147211, 0.0199838113, 0.0521308705, 0.0162938088, -0.0058836406, -0.0531318486, 0.0514884591, 0.0108284121, -0.0402554013, -0.0256477781, -0.0106255133, -0.0007293736, 0.0596322976, -0.0067944201, -0.0608691275, 0.1023470759, 0.0193191711, 0.0346514843, -0.0019267146, -0.042774789, -0.0350229815, 0.1002544239, -0.0119078336, 0.0404348411, -0.0130089726, -0.0018783734, -0.0249319058, -0.0071612354, -0.0233526304, -0.0402760915, 0.0266504847, -0.056387879, 0.0017940449, 0.0326393731, -0.0525883362, -0.0528700501, -0.0033372766, -0.0446855612, -0.0110631166, -0.004482931, -0.0395402387, -0.0362515338, -0.0322816186, 0.0383767374, 0.0362548977, 0.0013180691, -0.0199503247, -0.0086213807, -0.0230172221, 0.0246948265, -0.0034937612, -0.0125096347, 0.0155431023, 0.0235278346, 0.0728851035, 0.0322081409, -0.0217492208, 0.0175351035, -0.0065614963, 0.0149242319, -0.0117398901, -0.0582511164, 0.0455290228, 0.0214234199, -0.0666066334, -0.0482073091, 0.0836002156, 0.0281389616, -0.0537387058, -0.0060686837, -0.0059605683, 0.0001311203, 0.0122733377, 0.035041213, -0.0790769681, -0.0070911334, 0.0055900263, 0.003123149, -0.0373461209, -0.0388167053, -0.0281994287, -0.1171124652, -0.0064942902, 0.0036731677, -0.0309588313, -0.010239928, -0.0169790257, -0.0345819779, 0.0461240597, -0.0051222369, -0.0010420018, 0.0127486642, -0.0315675028, -0.0441866778, -0.008955718, -0.0056358851, 0.0282657929, -0.0453717932, 0.0365632214, 0.0595404953, 0.0388689525, 0.0237584226, -0.0418929681, 0.0162256006, 0.0101413252, 0.0473693199, -0.0025180562, 0.0480518043, -0.0188740492, 0.0209719352, -0.0203899872, 0.0409166217, -0.0317387693, -0.0161269642, 0.0209748577, 0.0549689084, -0.0553318001, -0.0008901589, -0.0360652097, -0.0196470656, -0.0052977391, 0.0697159618, -0.0186483786, -0.0469797701, -0.0398170836, -0.0372822881, 0.0065023531, -0.0255251396, -0.0155648077, 0.0288070906, 0.0066261874, -0.0258607026, -0.040739622, -0.0502392761, -0.0017934565, 0.0305485558, 0.025102349, -0.0221472736, 0.0137516661, 0.0004877574, -0.059548445, -0.0028820042, -0.0772936046, -0.0221810285, -0.0275602546, -0.0089469673, 0.0328218602, -0.0495888963, 0.0140311103, 0.0260727499, -0.0388370305, -0.0492598452, 0.0094138198, 0.0548666939, 0.0382107645, 0.0476100072, -0.0045580715, -0.0617205128, 0.0509805903, 0.0079122745, -0.0423372686, 0.0276176203, 0.0068159248, 0.0378414057, 0.0090147909, 0.0257693958, -0.0675720945, 0.0098244352, 0.0110759949, -0.0202927478, -0.0092077116, -0.0380212851, 0.0167907514, 0.0285587367, -0.0033375628, 0.0166286472, 0.0177678708, -0.0035096367, 0.0200855825, -0.0354090668, 0.008815092, -0.0142426919, -0.0002311138, 0.0201491378, 0.0168527663, 0.030182926, -0.0196331386, 0.0221826807, 0.0102846893, -0.0190590564, -0.0551451072, 0.0137834344, 0.0180083551, -0.0049418616, 0.0185999963, -0.0351283289, 0.0351301879, -0.0058891037, 0.0701883882, -0.0440425165, 0.0674849898, -0.0042931838, -0.0161514897, -0.0245639011, 0.0388979167, -0.068583414, -0.0839750841, 0.0406030566, 0.0511369891, 0.0217338149, 0.0167614818, -0.0410334133, -0.0453587845, 0.0143499123, -0.0123685589, 0.0074322829, 0.0325271674, -0.0803290978, -0.0500029586, 0.0341203623, -0.0317607485, 0.0271419305, 0.0499606989, -0.0169135239, 0.0405495279, -0.0263947863, -0.0157770794, -0.0449883938, -0.0213260539, -0.0076349443, 0.0659543425, -0.0006587, -0.0151572619, -0.0696310252, -0.0504355542, 0.0183322392, 0.0055072671, 0.0057041934, -0.0276995804, 0.0489612073, 0.0524147749, -0.0032408938, -0.0410855561, -0.024995828, 0.0231666807, -0.0897360593, 0.0046455637, -0.0079310704, -6.640509382e-33, -0.0144628864, -0.035951551, 0.0245263688, -0.0202920698, -0.0041169217, -0.0218163002, -0.0256848708, -0.0101963701, -0.02787452, -0.0273377653, -0.0313893706, 0.0266268179, 0.0026001674, 0.0305132139, -0.0152930468, -0.0254275948, 0.0383696444, 0.0283873416, 0.0056369985, 0.0589154661, 0.0782587454, 0.0159813985, -0.002083763, -0.0944847018, 0.026751101, -0.0018211892, 0.0044057984, 0.027626995, -0.0363934599, 0.049513083, -0.0229591094, 0.0514143817, 0.0179924555, 0.009283537, 0.0192975104, 0.060502205, -0.0802938417, -0.0455374494, 0.0051760203, -0.0247694682, -0.0287857391, -0.070598796, 0.0230281446, -0.0246869996, -0.0059581981, 0.0278489739, 0.0360797271, -0.02315416, -0.0078188935, -0.0393411852, -0.0336677581, -0.0183730591, -0.0013125538, 0.0225219522, -0.0190477613, 0.0034786882, -0.0085131992, 0.0245466921, -0.0418753885, 0.0280929077, -0.0348694846, -0.0298042689, 0.0432167128, 0.0052272547, 0.036333397, 0.0462451205, 0.0561994128, 0.0376487635, -0.0631472766, -0.024647072, 0.0071935742, 0.0353029743, -0.0186078232, 0.0377538726, 0.028612107, -0.0299262442, -0.0542857014, 0.016191531, 0.0321404412, 0.0374685302, 0.0467236117, 0.0023223616, -0.0780986473, -0.0286473483, -0.0253653992, 0.009957673, -0.045871906, -0.0245598126, 0.0369972624, 0.0050938134, -0.040765956, -0.0055321371, 0.0156198014, -0.049173411, 0.0393882729, 0.035272561, 0.0216874331, -0.0220907237, -0.0095945699, 0.019126948, -0.0172017533, -0.0293824058, -0.0272406489, -0.0046980851, 0.0144085269, 0.0382938981, 0.0077593001, 0.0409289747, -0.0198413748, -0.0077556437, -0.0293901395, 0.0184381045, 0.0014679192, -0.0365358815, 0.0320710726, 0.0371757895, -0.0034422362, 0.0658534616, 0.0398868471, 0.0289846919, -0.0403830707, -0.0103259301, -0.0073918705, -0.0129067218, -0.0095653525, -0.0008434963, -0.0502368845, 0.0255847052, 0.0490743555, -0.0590689145, 0.0041468935, -0.0017032949, 3.17e-7, -0.0341869183, -0.0055999598, 0.0298298784, -0.0167308953, 0.0041553448, 0.0244594105, 0.0554982834, 0.0331754349, -0.0002649577, -0.0261256695, -0.0051598535, 0.069637388, 0.0134324078, 0.0109732151, -0.0689481869, 0.0247614421, -0.0195264574, -0.0018204855, -0.0675642043, 0.0484362207, 0.0378419161, 0.0824200958, 0.0550197177, -0.0181590095, 0.0098606376, -0.0795691088, 0.0574872755, -0.0252135377, 0.0123175662, -0.0002851292, 0.0012609483, 0.0852427408, -0.0306866337, -0.0074262433, 0.0122629171, 0.0116908597, -0.0531626418, 0.0613761619, -0.0053638155, -0.0456508994, -0.0220209043, -0.0510707609, -0.0043467553, -0.0197203644, 0.0681848004, -0.0504100919, -0.0525153913, 0.0052687749, 0.044639755, -0.0238491744, -0.0315225199, 0.0006420434, 0.0067096213, -0.0258426815, -0.0063843792, -0.011257234, 0.0296910442, -0.1129518747, 0.0044378252, 0.0076238126, -0.0245624315, -0.0262802206, -0.0421915203, 0.0172567945, 0.0380698815, -0.0417359285, 0.0298971087, 2.793213579e-34, 0.0086173909, 0.0116837565, 0.0010980774, 0.0142007982, 0.0540433191, 0.0049521714, -0.0247392394, 0.031434644, 0.0673209578, -0.0685736388, -0.0318676345 ]
http://arxiv.org/abs/2406.18305v1
http://arxiv.org/pdf/2406.18305v1
S3: A Simple Strong Sample-effective Multimodal Dialog System
2024-06-26T12:45:43
2024-06-26T12:45:43
Elisei Rykov, Egor Malkershin, Alexander Panchenko
cs.CL
cs.CL, cs.AI
In this work, we present a conceptually simple yet powerful baseline for the multimodal dialog task, an S3 model, that achieves near state-of-the-art results on two compelling leaderboards: MMMU and AI Journey Contest 2023. The system is based on a pre-trained large language model, pre-trained modality encoders for image and audio, and a trainable modality projector. The proposed effective data mixture for training such an architecture demonstrates that a multimodal model based on a strong language model and trained on a small amount of multimodal data can perform efficiently in the task of multimodal dialog.
S3: A Simple Strong Sample-effective Multimodal Dialog System: In this work, we present a conceptually simple yet powerful baseline for the multimodal dialog task, an S3 model, that achieves near state-of-the-art results on two compelling leaderboards: MMMU and AI Journey Contest 2023. The system is based on a pre-trained large language model, pre-trained modality encoders for image and audio, and a trainable modality projector. The proposed effective data mixture for training such an architecture demonstrates that a multimodal model based on a strong language model and trained on a small amount of multimodal data can perform efficiently in the task of multimodal dialog.
[ 0.0514015891, 0.092367135, -0.0128529277, 0.0437224209, -0.0014756359, 0.0008982614, 0.0408712812, -0.0293932073, -0.033736378, -0.089324221, -0.0663861856, -0.0138560375, -0.0360643119, 0.0852366686, 0.0502515286, -0.076252237, 0.0277221091, -0.0640073419, -0.0013914033, 0.002553961, 0.0573258102, 0.0100092264, -0.0495383143, 0.0361227468, -0.0583669245, -0.0048455424, -0.046162352, -0.0034835036, -0.0244508348, -0.0530950539, -0.0116926031, 0.0416704565, 0.0689137504, 0.0081642279, 0.0000017465, -0.0254573785, 0.0240191631, -0.0159954131, -0.0198504776, -0.0282195713, 0.0262911748, 0.0023805848, 0.0154178701, 0.0121532576, -0.0217421167, -0.0171581488, 0.0498909689, 0.082956098, 0.0994793698, 0.0652486831, -0.0346482508, -0.0165420063, 0.0164916944, -0.0165896546, 0.0301417597, -0.0424653962, -0.0257351138, 0.0248351172, 0.0146788452, 0.0253082458, -0.0269654021, 0.027694324, 0.0089552188, -0.0336257257, -0.0280359723, 0.0121872611, -0.0011861526, -0.0314113311, -0.051195208, 0.1252182871, 0.0256785862, 0.0479692966, -0.0118974727, 0.0012861881, -0.0042005102, -0.0356384143, -0.060925331, 0.0363492817, 0.0104172565, -0.0162434727, -0.0570873879, 0.0315988623, -0.0010512728, 0.0151061155, 0.0116801308, -0.0178409237, 0.0095427232, -0.0040861052, 0.0074649756, -0.0111641772, -0.0038991338, -0.0606287457, 0.0010911886, -0.0036961637, -0.0610789433, -0.0036506401, -0.035131637, -0.0079331128, -0.0142734479, -0.0203380771, 0.0070709861, 0.0212698355, -0.0271445867, 0.0713465959, 0.0302386489, 0.0230926163, -0.0090660257, 0.0268178359, -0.0052376874, -0.0373216458, -0.0412410162, 0.0060703917, -0.0436240658, -0.0016880848, -0.0660362095, -0.0592779964, 0.0037137151, 0.0296106469, -0.0146737974, 0.007560668, 0.0186102372, 0.0308936369, -0.0382179469, -0.0097489813, -0.0356537998, 0.0573174395, -0.0303206462, 0.0500166751, 0.0088125039, -0.057708282, 0.0175327007, 0.0645606518, 0.0056771664, -0.017514633, -0.0009092904, 0.09775839, -0.0648604333, -0.0389765315, -0.0024609519, -0.0112161441, -0.0404669866, -0.0569283962, 0.0183568913, -0.0215604007, 0.0587221682, -0.0108579127, -0.0022607301, 0.0077072624, 0.0124231176, -0.0200521685, 0.0254007634, 0.0724537298, 0.0172488298, -0.0082189739, 0.0566626117, 0.0181914605, -0.0403340384, 0.0746001154, 0.040110711, 0.0357424095, 0.039234437, 0.0046820212, 0.0171983372, 0.0450671725, -0.0324645415, 0.00816694, 0.0154600907, 0.0032606455, -0.0281046405, 0.031853646, 0.0363364369, 0.0758516416, -0.043879129, -0.0304344837, 0.0720434338, 0.0200675335, 0.0027159338, 0.0058965054, 0.0109012136, 0.0120851165, 0.0044970461, -0.0473116301, -0.0268740766, 0.0485723205, 0.0136500867, -0.0177618377, 0.0468912683, 0.0311674904, -0.0428213179, -0.0101582194, 0.050314974, -0.0288896225, -0.0137468958, 0.0052850097, 0.0366829447, 0.0007382626, -0.0603069812, 0.0355780646, -0.0116119832, -0.0470108688, -0.001054337, 0.0023076704, 0.0214244891, 0.0251446739, 0.0115871457, -0.0033612144, -0.011770145, -0.040615622, 0.0235357434, 0.0315305926, 0.0217196699, -0.031315092, 0.064192161, -0.0797924846, -0.0460496806, -0.0156821143, -0.0126309153, 0.0163062271, -0.0366908871, -0.0020475665, 0.0141096171, -0.0061941287, 0.0689820424, 0.0017921565, 0.0471423082, 0.0142585859, -0.0157383475, -0.0017429979, 0.0652965233, 0.0081133191, -0.0399388336, -0.0165326931, 0.0098991841, -0.0286481027, 0.0204240959, 0.0459116995, -0.0799278691, -0.0465634502, 0.0219814833, -0.0561940745, -0.0008875227, 0.0078295637, -0.0476032346, -0.0436012931, 0.019412268, -0.0579416119, -0.0106503107, -0.0490252674, -0.001612476, 0.0270814206, 0.0126973474, -0.02733678, 0.0173138417, -0.004555285, -0.0365782157, 0.0252151303, 0.0289146043, 0.0376615077, -0.0433627367, -0.040338818, 0.0278714318, -0.0057649002, 0.0303583629, 0.006201786, -0.0127118388, -0.0356373824, -0.0400285311, -0.132952854, -0.031674657, -0.0000605025, -0.0276825372, -0.004707878, 0.0139305238, -0.0031890036, -0.0373163633, -0.0380951762, 0.0225905981, 0.0623694621, 0.0070864786, 0.0129668191, 0.0392455459, 0.0146701755, -0.0417311415, 0.0026263292, 0.0167435277, 0.0514037199, 0.0029882817, -0.0070673642, -0.0148349861, -0.0155844092, -0.0362591483, 0.0663151518, -0.0171483681, -0.0567365587, 0.0249609035, -0.0118158031, 0.0752394423, 0.0102929166, 0.026911974, 0.0180841796, -0.0095631974, -0.0256626569, -0.0340647362, 0.0334612466, 0.0064015999, 0.0158071369, -0.0369400568, 0.0032204119, 0.0137571888, 0.0282541309, 0.0337154344, 0.0487572514, -0.0276393574, 0.0171708353, 0.0357402265, -0.0820448101, -0.0242102351, -0.0065743998, -0.0088234609, 0.0310136806, -0.0277691819, -0.0170830116, 0.023984246, -0.07941062, -0.0354386866, -0.0066459412, 0.0027063712, -0.0110022286, 0.0623084754, -0.018294692, 0.0356281251, -0.0171555262, -0.0119322091, -0.0501944274, -0.0273848902, 0.0308125522, -0.0060115186, 0.0379642732, -0.0289427917, 0.0133486595, -0.0148857012, -0.0829701945, -0.0236324929, -0.0404138602, -0.0318797193, -0.0408352576, -0.0085837757, 0.0029878414, -0.0437447466, 0.0115837855, 0.0344176665, 0.0300149433, 0.0059049684, -0.0037851825, -0.0243679341, 0.0545075573, 0.053890463, 0.0661409423, -0.0445820354, -0.0669925138, 0.0320293494, 0.0309196115, 0.0380707346, 0.0057864566, -0.0602100007, 0.0249127541, -0.0007875569, 0.0602498278, 0.0011119813, 0.0235148445, 0.0516025499, -0.029150771, 0.008700707, 0.0727152005, -0.0181158856, -0.0339345299, -0.0222223271, 0.0146999583, 0.0040799826, 0.0195814651, 0.0004201831, -0.0794737339, -0.0738023371, -0.0123994481, -0.0328144692, 0.0136113092, 0.0121086501, 0.0099598849, -0.0626872703, 0.041253902, -0.0103939082, -0.0175262336, -0.0375105627, 0.0062987674, -0.0115661016, -0.0047587906, 0.0012310199, -0.0133158788, 0.020172799, 0.0054833023, 0.0101095075, 0.0412709117, -0.0424964912, 0.0151647059, -0.0310377758, 0.0581324399, 0.0481288135, 0.0074739424, -0.0175246578, -0.0382401645, 0.0175104327, 0.0286030695, -0.0074040876, 0.0129052652, 0.0578212738, -0.0387018286, -0.0378468446, -0.02996324, 0.003930998, 0.006471076, -0.0197750516, 0.0062809372, 0.0581727251, -0.0375845619, 0.0234794356, -0.0247682314, 0.0248010177, 0.0427887812, 0.0472580381, 0.0064349291, -0.0547073223, -0.0236837473, -0.0136041762, 0.015560789, -0.023903681, 0.0071568047, 0.0109472107, 0.0351539366, 0.0338294245, -0.0599331371, -0.020539796, 0.0123441769, 0.0550006591, -0.0091156894, -0.0255323872, -0.0178089608, 0.0334652252, -0.0188828316, 0.0418224335, -0.0054860045, 0.0090144612, -0.0642976984, 0.0025191149, 0.0512578972, -0.0195150319, -0.0177801158, 0.0301276576, -0.0521996878, -0.0211023148, -0.0507710204, 0.0907416269, 0.0129421148, 0.0324778594, 0.0277935211, -0.0426785834, 0.0228133481, -0.0123100784, -0.0277177021, -0.0092190448, 0.0483781584, 0.0156383123, -0.0159407463, 0.0091417683, -0.0474003442, 0.0106073972, -0.0114534227, 0.0260985345, -0.0085997833, -0.0568879433, -0.0219900161, -0.02368409, 0.0172406435, -0.0054515488, 0.003522034, -0.0391644798, 0.0193002522, 0.0085717225, -0.0090666469, -0.0870158151, 0.0215943754, 0.0038000643, 0.0131771853, 0.0289197657, -0.0005585394, 0.0014830836, 0.0419779271, -0.0141283069, 0.0382314287, 0.0559158213, -0.0249026977, -0.0206011776, 0.0086500226, 0.0010745201, 0.0347135998, 0.0018201793, 0.0098150717, -0.0144954193, 0.1348132342, 0.0069399676, -0.0481240265, 0.0081669576, 0.0534135923, -0.0128026493, -0.0029143423, -0.0261103753, 0.0456288047, 0.0228739586, 0.0274125189, 0.0161688402, 0.0058359508, 0.0097675482, -0.0146473283, 0.0776871666, -0.0126403552, -0.0222185086, 0.0453971736, 0.0203569047, 0.0041955011, 0.0221191011, 0.0466017984, -0.0247075073, 0.0288319997, -0.0052212649, 0.0181566235, -0.0686851963, -0.0231601652, 0.0401021205, 0.04291014, -0.0074094231, 0.0298346765, 0.008464695, -0.0361787751, -0.0195809659, 0.0115515133, -0.0146139888, 0.0181277562, 0.072796613, 0.0706511065, 0.0038965286, -0.014883073, -0.0149458768, -0.0118659344, 0.0170820579, 0.0001132548, -0.0166309401, -5.57196665e-33, -0.0127363587, -0.0143616274, 0.0349584632, -0.0493784733, -0.022305591, -0.012026717, -0.0418228246, 0.0130869411, -0.0054223775, -0.0049987794, 0.0198730361, -0.0068314881, -0.0023426993, 0.0729202107, -0.0137995109, 0.0181874558, 0.0461933725, -0.0127974134, -0.0328929611, 0.0216412656, 0.0373770446, 0.0034379084, 0.02933635, -0.0826886147, -0.0268481467, 0.0307092424, -0.0506423265, 0.0130791832, 0.0195628237, 0.0754337609, -0.0236978885, 0.0365781523, 0.0050314665, -0.0334012061, -0.0131394025, 0.0315151885, -0.044126302, -0.0211331304, 0.0222498979, 0.0297324024, -0.0537693873, 0.0215589907, 0.0048980848, -0.0487531424, 0.0343804322, -0.0008727855, 0.0216461401, 0.0140174069, 0.0094357543, 0.0006701733, -0.0530281067, 0.0232190508, -0.0380111746, -0.0269545205, 0.0206168834, 0.0514177121, 0.0395941585, -0.0307828002, -0.0105069363, 0.0365520194, -0.0174718928, -0.0095313983, 0.0471129604, 0.0301263407, -0.0124184089, 0.0768577084, 0.0355291478, -0.0343180485, -0.0161744338, -0.0132393455, -0.0056293579, 0.063019447, 0.0280103106, 0.0932218805, -0.029421445, 0.011114276, -0.0307183396, 0.0182272401, 0.0216257852, 0.0252466872, 0.0096709076, 0.0251789875, -0.0064842156, 0.0067349961, -0.0586667918, -0.0176502559, -0.0213631485, -0.0477312095, 0.0644248128, 0.0339829102, 0.0143781481, 0.0107595352, 0.0990894139, -0.0177010838, -0.0217341501, 0.0007562385, -0.0056729121, 0.0457631312, -0.0113199027, -0.0125015806, -0.0303257499, -0.0252115261, 0.0852202252, -0.0094127208, 0.0298121125, 0.0064794673, -0.0068385783, 0.0374388173, -0.0375104174, -0.0162477698, -0.0223855302, 0.0097618718, -0.0163553692, 0.053822007, 0.0072198701, -0.0139928851, -0.0831456482, 0.0352779366, -0.0559624694, 0.0046197306, -0.0374694504, -0.0981409103, -0.0172980167, 0.0114519643, -0.0527874157, 0.0071020792, -0.0482017137, -0.0892469212, 0.038451992, -0.0588230416, 0.0013540583, 0.0147120934, 2.446e-7, -0.0197997857, 0.0170676038, 0.0028607182, -0.0008631377, 0.0294010248, 0.0632493421, -0.0226210151, 0.0190074723, -0.0064115385, -0.0404958837, 0.0029903171, 0.0307370443, 0.0102236038, 0.0172332041, -0.0384011343, -0.0170646273, -0.0466772877, -0.0163666345, -0.0464853235, 0.0278403852, 0.0764798597, 0.1483282894, -0.0094371475, -0.0150525533, 0.0021249047, -0.044926174, 0.0730368569, -0.021432735, -0.0126771666, -0.0068868571, 0.0412163511, -0.0178108253, -0.0214300975, 0.0112707419, 0.0075282925, -0.0031108754, 0.033181835, 0.0465027094, -0.0086307647, -0.0515322909, -0.0054195002, 0.0489508361, -0.029014254, -0.0086585116, 0.0575146489, -0.0203684475, -0.0537128337, -0.0636054873, -0.02108174, -0.0243802853, -0.0078014182, -0.0234931875, -0.0419881381, 0.0067372541, 0.0227816589, -0.0207534917, -0.0107601723, -0.0884350091, 0.0080962507, -0.0348877311, -0.0424065217, -0.0343972966, -0.0424902886, -0.0537507385, 0.0458745807, -0.0364497378, -0.0155970259, 1.636315228e-34, 0.0268331803, -0.0145175643, 0.0240225587, 0.010180464, 0.0519500896, 0.0008475478, 0.011261154, -0.0263949577, 0.0041910582, -0.0683220476, -0.0323557667 ]
http://arxiv.org/abs/2406.18301v1
http://arxiv.org/pdf/2406.18301v1
MSR-86K: An Evolving, Multilingual Corpus with 86,300 Hours of Transcribed Audio for Speech Recognition Research
2024-06-26T12:35:12
2024-06-26T12:35:12
Song Li, Yongbin You, Xuezhi Wang, Zhengkun Tian, Ke Ding, Guanglu Wan
eess.AS
eess.AS, cs.CL, cs.SD
Recently, multilingual artificial intelligence assistants, exemplified by ChatGPT, have gained immense popularity. As a crucial gateway to human-computer interaction, multilingual automatic speech recognition (ASR) has also garnered significant attention, as evidenced by systems like Whisper. However, the proprietary nature of the training data has impeded researchers' efforts to study multilingual ASR. This paper introduces MSR-86K, an evolving, large-scale multilingual corpus for speech recognition research. The corpus is derived from publicly accessible videos on YouTube, comprising 15 languages and a total of 86,300 hours of transcribed ASR data. We also introduce how to use the MSR-86K corpus and other open-source corpora to train a robust multilingual ASR model that is competitive with Whisper. MSR-86K will be publicly released on HuggingFace, and we believe that such a large corpus will pave new avenues for research in multilingual ASR.
MSR-86K: An Evolving, Multilingual Corpus with 86,300 Hours of Transcribed Audio for Speech Recognition Research: Recently, multilingual artificial intelligence assistants, exemplified by ChatGPT, have gained immense popularity. As a crucial gateway to human-computer interaction, multilingual automatic speech recognition (ASR) has also garnered significant attention, as evidenced by systems like Whisper. However, the proprietary nature of the training data has impeded researchers' efforts to study multilingual ASR. This paper introduces MSR-86K, an evolving, large-scale multilingual corpus for speech recognition research. The corpus is derived from publicly accessible videos on YouTube, comprising 15 languages and a total of 86,300 hours of transcribed ASR data. We also introduce how to use the MSR-86K corpus and other open-source corpora to train a robust multilingual ASR model that is competitive with Whisper. MSR-86K will be publicly released on HuggingFace, and we believe that such a large corpus will pave new avenues for research in multilingual ASR.
[ -0.0256809425, 0.0503706038, -0.0106096528, 0.0501827635, -0.0319331139, 0.0086256126, 0.025403345, -0.0128059965, 0.0029331294, 0.0122095952, -0.0885717869, -0.0391371772, 0.0248462856, 0.0104136178, 0.0334812365, -0.0665962398, 0.0273681302, -0.0878105089, 0.0163261071, -0.02485043, 0.0188638605, 0.0320702046, -0.0496385433, 0.0344165787, -0.0230099838, 0.0137427812, 0.0137168244, -0.0103786057, 0.0323030055, -0.0008771882, -0.0368678197, 0.0188605506, 0.0231347363, 0.0413729288, 0.0000018314, -0.0308051854, -0.0008887743, 0.0145889567, -0.0331566371, -0.0100002103, -0.0017735731, -0.0169331729, -0.0169688519, 0.055736281, -0.0229457952, -0.0148381237, 0.0372600257, -0.027250262, 0.0232020784, 0.042961739, -0.0189726856, -0.0454194993, 0.0323850326, -0.0139600011, -0.0083018327, -0.0447823331, -0.0279016457, -0.0282180253, -0.003609325, -0.0226826631, -0.0296611898, 0.0624343306, 0.0054666344, -0.0153101906, -0.0420503542, -0.0033305485, -0.0612285733, -0.0802634656, -0.0072794463, 0.0581252538, -0.0238572806, -0.0017562198, 0.0415350385, 0.0207642466, -0.0528666489, -0.0638720393, -0.025611911, 0.0291301869, -0.0066092592, -0.0122516137, 0.0046400372, 0.0155242886, 0.0255114976, -0.0289336368, 0.0414613448, -0.007683333, 0.0256297458, -0.0353825539, 0.0385877825, -0.0217274874, 0.024354307, 0.0026794735, 0.0678397417, -0.0057814955, -0.0157027282, 0.0114902882, -0.0147806639, -0.025222294, 0.0145662017, -0.0655565709, 0.0155772269, 0.0722016692, -0.0558838025, 0.0335019194, -0.0075762155, 0.0523484573, -0.0078936983, -0.0127861444, -0.0204752386, -0.0343351997, -0.0298934579, 0.0131594054, -0.0304504111, -0.0147884209, -0.0312919058, -0.038404163, -0.0230563767, 0.013744778, -0.0011892268, 0.0140592298, 0.0012107696, -0.0378220379, -0.0187818259, -0.005268848, 0.0473367572, 0.022072209, -0.0033943572, 0.0351831615, -0.0207275432, -0.0875933096, -0.0114397956, 0.0351133794, 0.0263360776, -0.0140755353, 0.0246426556, 0.0555456057, -0.0335245468, -0.0347852334, -0.0161817931, 0.0163182244, -0.0209618155, -0.0407694466, 0.0063322922, -0.0420306139, 0.0964975581, 0.013031831, 0.0098923352, -0.0147924237, 0.0007577323, -0.0178158432, 0.0082060723, 0.0699913278, 0.012114333, 0.033340428, 0.0385445245, -0.0151152862, -0.0167660881, 0.0622541755, 0.0569759794, -0.0268629771, 0.0614918396, -0.0240115467, -0.0002974275, 0.0895910263, -0.0423377939, 0.0249888748, 0.0215926692, -0.0236672834, 0.0113562159, 0.0116512952, 0.0184404179, 0.060080044, -0.0089852158, -0.0575439669, 0.0283036027, 0.0574804954, -0.0104820943, 0.0274613071, 0.0693362132, 0.0147764003, 0.0063379696, -0.0039404836, 0.0324716307, 0.0123122828, 0.0149420528, 0.0272302199, 0.0352596603, 0.062119361, -0.0559018366, 0.0158555713, -0.0255501289, -0.0453757867, 0.0007413567, 0.0523478501, 0.0534077697, -0.0103779873, -0.0298133828, 0.031031562, 0.0295691844, -0.0562451854, 0.0307151917, -0.0478251651, -0.0030124278, 0.0448199026, -0.0035704018, -0.0277028661, -0.0391733088, -0.0352289304, -0.0142993722, 0.0524500348, 0.0193713903, -0.0253631081, 0.0793798342, -0.0855712369, 0.0049747117, 0.0375993215, -0.0105578518, 0.0808567405, -0.0320132822, 0.0455837809, 0.0048835701, -0.0198833924, 0.0168706272, -0.0190609004, -0.053960003, 0.0115988692, 0.0175039507, 0.0102130724, 0.0794512555, 0.0554510243, -0.0449437909, -0.0264711175, 0.0334648415, -0.0276810564, 0.0340751596, 0.0783106238, -0.0353525579, -0.0091566835, 0.0551303253, -0.0293665379, 0.0030248398, 0.0329863466, 0.0158666354, -0.0671050623, 0.0671183988, -0.0033818746, 0.048166994, -0.05532033, 0.0529839173, 0.0294596981, 0.0343106501, 0.0542361215, -0.0357806981, 0.0234745108, -0.0091949925, 0.0048282603, -0.030285269, 0.0530094691, -0.0828238577, -0.0325583443, 0.0523699485, -0.0302485973, 0.0108516095, 0.0325807743, 0.0029963448, -0.0088843061, -0.015745027, -0.0770761445, -0.0187326595, -0.0368222222, 0.039997384, -0.045784764, 0.0346176215, -0.0239676833, -0.0445708074, -0.0123957833, -0.0161413606, 0.0152518693, 0.0183225684, -0.0553933345, -0.0167699419, -0.0020814198, -0.0224142242, -0.0230349507, 0.0239509791, -0.0113335149, -0.0229203179, -0.0565383062, -0.0463148244, -0.0173530616, -0.0035443115, 0.0279715601, -0.0246622581, -0.0422414765, -0.0084376158, 0.0247966573, 0.0366100706, 0.0238048416, -0.0324498862, 0.0031999757, 0.0128124347, 0.003410999, -0.0358844325, 0.012270865, 0.0329803787, -0.0145236719, -0.0230047926, 0.0161910709, -0.0257123727, 0.0108895628, 0.0182753447, 0.0315221287, -0.049398832, -0.066274032, -0.0324184336, -0.0354590677, -0.0061331671, -0.0092049036, -0.0467155315, 0.028499458, 0.0177414343, -0.0046742694, 0.0410711803, -0.0754235312, -0.0555116273, -0.0373845287, -0.039818082, -0.0394370072, 0.0249462221, -0.0310318936, 0.0219522808, -0.0257437583, -0.0527890846, -0.0089405831, 0.0212180614, 0.0203475188, -0.0230445173, 0.0161771514, -0.0497009121, 0.0002244815, -0.0091871209, -0.0705123469, 0.0039296327, 0.0163000301, -0.0338352248, 0.0226338804, 0.0138711939, -0.0164482109, -0.0449798554, 0.029033713, 0.037152268, 0.0612801425, 0.0160415936, 0.0598591454, -0.0105121545, 0.0335417874, 0.0160374995, -0.0100911334, -0.0404961705, -0.039770864, 0.0436941981, -0.0172755066, 0.0657810643, -0.0219443422, -0.0248771813, 0.000716728, 0.0668275356, 0.0105263609, -0.0147241009, -0.0214555934, 0.0405828767, -0.0028326495, 0.0357319415, 0.0884969831, -0.0121083716, -0.0540127866, -0.0245577693, -0.0189912971, 0.0014934728, 0.0150740556, 0.011469231, -0.0888561308, -0.0061666933, 0.050773222, 0.0142764132, 0.0411800183, 0.0233047716, -0.0138930734, -0.0835212469, 0.0132398363, 0.0074047623, -0.0534051396, -0.0352012925, -0.055138126, -0.0370968916, -0.0685604513, 0.0348443985, -0.0302451029, 0.0101124942, -0.0330030993, 0.0511750244, -0.044414524, -0.0700554177, 0.0298788957, -0.0404827967, 0.080723308, 0.0294075087, 0.0437513255, -0.04544615, -0.0450079031, 0.0070262114, -0.005158097, 0.000634503, -0.0159865562, 0.0013766548, 0.0153869884, -0.0382163897, 0.0354527347, 0.0167252012, -0.0122516099, 0.0112267761, 0.0386450738, 0.0633418038, -0.016561076, 0.0031154309, 0.0305291452, -0.0105777048, 0.0228415504, 0.0449945703, -0.0111752301, -0.0368001945, 0.0390435047, 0.0184930842, 0.0037324291, -0.0140034761, 0.0122662298, -0.0085939383, -0.0127102816, -0.0344158895, -0.0269632749, -0.020757135, 0.0190688111, -0.0090179127, 0.0318310186, -0.0132528981, 0.0177203733, 0.0425793342, -0.0187652446, 0.0127744749, 0.0178314596, 0.018948786, -0.0794191062, -0.0447236486, 0.0264023114, -0.0561850145, -0.0149871111, 0.0086748023, -0.0361034833, -0.0777201802, 0.0126748653, 0.0414722338, 0.0179640017, 0.0650037453, -0.0510062762, -0.0172382928, 0.021562025, -0.0058603743, -0.0443723798, -0.0105931489, 0.0574473031, 0.0039711739, -0.0296666473, 0.020185085, -0.0462924801, 0.0216686614, -0.0329937302, 0.0032484063, -0.0450730547, -0.0525388569, -0.0454570055, 0.0275154058, 0.0458795987, -0.0411434285, -0.0158771668, -0.0715137795, -0.0024201039, 0.0039198264, 0.0644376501, -0.0439696163, 0.0164555311, 0.0375786759, 0.0191071797, -0.0276120305, 0.0199877732, -0.0136187989, -0.0644449592, 0.0009378426, 0.0586260483, 0.0401698574, -0.0062728953, 0.0030930443, 0.0311416127, -0.0190699268, 0.0754966065, 0.0084460331, 0.0127256727, -0.004914131, 0.0176236015, -0.0311660264, 0.0092800148, 0.05995268, 0.0427151509, -0.0140992925, -0.0186428744, 0.0108018834, 0.0565554649, 0.0671361536, -0.0048089097, 0.0960024521, -0.004661432, 0.0107440073, -0.0336679891, 0.0487199686, -0.0232328027, -0.029347267, 0.001702221, -0.0067195864, -0.031197181, 0.0437019467, 0.0793590695, -0.0335133448, 0.0231553875, 0.0164533295, 0.0300466307, -0.0224651918, -0.0327813886, 0.0053684958, -0.0096064676, -0.038336128, -0.0010506416, 0.0099352282, -0.0812224001, 0.0070004668, -0.0078427959, 0.0167820826, 0.0016645093, -0.001804604, 0.0582271852, -0.0313400328, 0.0223021004, -0.0271537192, -0.0802426413, -0.0390675962, -0.0257609449, 0.0357747227, -5.827407082e-33, 0.0151878083, 0.0515062101, -0.0055983299, 0.0380501188, -0.0215671416, -0.0201222971, -0.0103754308, 0.009888933, -0.0208895914, -0.0160901006, -0.0273860916, -0.0135004427, 0.0365346447, 0.000038994, 0.0557392575, -0.0035346467, 0.0658567473, -0.0014604754, 0.0261863861, 0.016674364, 0.0400063992, 0.0636951625, 0.0777778849, -0.0629165322, -0.0042388239, 0.023312293, 0.0033888884, -0.0044292486, 0.0154969329, 0.0841938555, -0.0170765053, 0.0466721542, -0.0132848611, -0.0098174969, -0.0188364312, 0.0005233408, -0.0619382635, 0.0106124431, -0.0014927591, 0.0935868472, 0.019795686, -0.0113070579, -0.0152873946, -0.0327241495, -0.0185839199, -0.0015045895, 0.0257441718, -0.0110882167, -0.0043975487, -0.03226199, -0.0355301164, 0.0417158529, -0.0262628347, -0.0474303961, 0.1002355218, 0.0546385013, -0.0176876821, -0.0471435525, -0.053109616, 0.0342266709, -0.0618849993, 0.0201950166, 0.0670711026, 0.0100491112, -0.0083970642, 0.0122081256, 0.0351481214, -0.0940238833, 0.0554066785, -0.0187657718, 0.019121768, 0.0301548373, 0.0003403508, 0.0329673961, 0.0139621841, -0.0654482618, -0.0257908888, -0.0513048284, 0.0701415613, 0.0196681824, 0.0096683297, 0.0100828111, -0.0303727388, 0.0062876721, -0.0474023186, -0.0544239283, -0.0215629395, -0.0265176408, 0.0229400285, 0.0105662951, 0.0370670892, 0.0228327513, 0.0759697482, 0.0009850294, 0.0195087921, 0.0018884703, 0.0016690862, -0.0179090686, 0.0062102424, 0.0578815527, -0.0614708886, -0.0231735427, 0.0027256315, 0.008660269, 0.0172943883, -0.0191378146, 0.0117469663, 0.0822617635, -0.058146853, 0.0114167463, -0.0142050255, 0.0215130616, 0.0111436099, 0.0359697305, 0.0389668494, -0.0013700663, -0.0235776398, 0.0544201806, 0.0122503359, 0.0506718867, -0.0095046172, -0.0355714783, -0.0395367555, 0.0081792837, -0.0422510803, 0.0198629629, -0.018358713, -0.0346718766, 0.012151625, -0.0300178025, -0.0054844576, 0.0137084322, 2.718e-7, -0.0440277047, 0.0244997442, 0.0113446387, -0.0004684378, 0.0170417726, 0.0117360111, -0.0473661982, 0.0140552027, -0.0097528119, 0.0385704488, -0.0106614893, -0.0052047474, 0.1000409499, 0.0384058096, -0.1042164937, -0.0549531318, -0.0794412345, 0.0040760119, -0.0232182313, 0.0069901259, -0.0213244725, 0.1011574119, -0.0121693956, -0.0056859343, -0.0107091339, -0.0250344388, 0.0353610367, -0.0086049205, -0.0386796966, -0.0307963472, 0.019556867, -0.013176119, -0.0198464263, -0.0165769737, -0.041773323, 0.0401131846, 0.0414371789, 0.0548347495, -0.003494198, -0.0552037321, 0.0186201297, 0.045129545, -0.0634267926, -0.0148440944, 0.0470685884, 0.0445039496, -0.032154005, -0.03570351, -0.0084999939, 0.0030662576, 0.0476726219, -0.0335994363, 0.0141880838, 0.0420602709, 0.0284855645, -0.0424797721, -0.0179614071, -0.0379089117, 0.0229572542, -0.0209175237, -0.0354109257, 0.0095447209, 0.0112288967, -0.0050149621, -0.0076519875, 0.0007101455, 0.0046858997, 1.492925006e-34, 0.0469464697, -0.023661349, 0.0178090483, -0.0181720201, 0.013197585, -0.0449946672, 0.0114433113, 0.0265649986, -0.0198758319, -0.0103485677, -0.0405579805 ]
http://arxiv.org/abs/2406.18297v1
http://arxiv.org/pdf/2406.18297v1
FactFinders at CheckThat! 2024: Refining Check-worthy Statement Detection with LLMs through Data Pruning
2024-06-26T12:31:31
2024-06-26T12:31:31
Yufeng Li, Rrubaa Panchendrarajan, Arkaitz Zubiaga
cs.CL
cs.CL
The rapid dissemination of information through social media and the Internet has posed a significant challenge for fact-checking, among others in identifying check-worthy claims that fact-checkers should pay attention to, i.e. filtering claims needing fact-checking from a large pool of sentences. This challenge has stressed the need to focus on determining the priority of claims, specifically which claims are worth to be fact-checked. Despite advancements in this area in recent years, the application of large language models (LLMs), such as GPT, has only recently drawn attention in studies. However, many open-source LLMs remain underexplored. Therefore, this study investigates the application of eight prominent open-source LLMs with fine-tuning and prompt engineering to identify check-worthy statements from political transcriptions. Further, we propose a two-step data pruning approach to automatically identify high-quality training data instances for effective learning. The efficiency of our approach is demonstrated through evaluations on the English language dataset as part of the check-worthiness estimation task of CheckThat! 2024. Further, the experiments conducted with data pruning demonstrate that competitive performance can be achieved with only about 44\% of the training data. Our team ranked first in the check-worthiness estimation task in the English language.
FactFinders at CheckThat! 2024: Refining Check-worthy Statement Detection with LLMs through Data Pruning: The rapid dissemination of information through social media and the Internet has posed a significant challenge for fact-checking, among others in identifying check-worthy claims that fact-checkers should pay attention to, i.e. filtering claims needing fact-checking from a large pool of sentences. This challenge has stressed the need to focus on determining the priority of claims, specifically which claims are worth to be fact-checked. Despite advancements in this area in recent years, the application of large language models (LLMs), such as GPT, has only recently drawn attention in studies. However, many open-source LLMs remain underexplored. Therefore, this study investigates the application of eight prominent open-source LLMs with fine-tuning and prompt engineering to identify check-worthy statements from political transcriptions. Further, we propose a two-step data pruning approach to automatically identify high-quality training data instances for effective learning. The efficiency of our approach is demonstrated through evaluations on the English language dataset as part of the check-worthiness estimation task of CheckThat! 2024. Further, the experiments conducted with data pruning demonstrate that competitive performance can be achieved with only about 44\% of the training data. Our team ranked first in the check-worthiness estimation task in the English language.
[ 0.0281821247, 0.0563504472, 0.0181530528, 0.0407276414, -0.0616867691, 0.0081377029, -0.0172804426, 0.0133102005, 0.0214996208, -0.0469894595, 0.0185855329, -0.0153538482, -0.0102797775, 0.0154339094, 0.0037098352, 0.0173772816, 0.0813762248, 0.0109061105, 0.0757082105, -0.0260097142, 0.0110038556, 0.0266959015, -0.0464721136, 0.0097716898, -0.0414648801, 0.0140846707, 0.0455171727, -0.004435665, 0.0554458909, -0.0230648853, 0.0024350467, 0.048005417, -0.0069637639, 0.0678254068, 0.0000018085, -0.0474208631, -0.0167495999, 0.0256809723, -0.0391665287, 0.0460973494, 0.0635452196, 0.0170284323, -0.0390117541, 0.0207442753, -0.0067049195, -0.0102363927, 0.029557161, 0.0357639715, 0.0943155885, 0.100048773, 0.000915698, -0.0010561384, 0.0286871716, -0.0095447507, -0.0133955311, -0.0439614914, 0.0646856874, -0.0102082901, -0.0294461288, 0.0239332877, 0.0188617799, 0.0460145101, 0.0071356297, -0.0067374669, -0.0452286601, 0.0113769751, -0.085010998, -0.0141694508, -0.0178028215, 0.0288750622, 0.0197185036, -0.0565830395, 0.0152605409, -0.0331416801, 0.0058204425, 0.0264525805, -0.0055298018, 0.0368984304, -0.0425357074, 0.0124056935, 0.0134839574, -0.0250451192, 0.0230494812, 0.0425468236, 0.009402914, 0.0782741308, -0.0135870231, -0.0652851313, 0.0489088073, 0.0073005119, 0.0017688399, -0.0431606211, 0.0189740043, -0.0169422887, 0.0003137097, 0.0152273541, -0.0202235281, 0.0122591155, 0.0685978383, -0.0442312956, -0.0186542645, 0.0084394533, -0.0565817878, 0.0812968761, -0.0118816402, 0.0503482558, 0.0211724564, -0.0039568269, 0.022303097, -0.0234285817, -0.0245585442, 0.0135198953, -0.0020722302, 0.0441726446, -0.0174535848, 0.0208087061, -0.0560321547, -0.0392903537, -0.0275780559, 0.0405974165, 0.0158011168, -0.0390597284, 0.0119873229, 0.0065060989, -0.0672631413, 0.030496832, 0.0075991652, 0.014267765, -0.041749876, -0.0789755732, 0.0050151097, 0.0271357168, 0.0113759348, -0.0201985873, 0.0083049778, -0.0282983854, 0.0095455712, -0.0340492167, -0.0427495204, -0.0057486575, -0.0381694399, -0.0656240582, -0.0007809863, -0.0572101697, 0.0062544523, -0.0364141092, 0.0234951712, -0.0159459822, 0.0012999737, 0.0488920063, -0.0250281785, 0.0245964155, 0.0127609326, 0.0217489786, 0.0062847962, -0.006975905, 0.0421851687, 0.0646308064, 0.0448996462, 0.0270254109, 0.0734021589, 0.0037434467, 0.0247243494, 0.032491792, 0.0268966127, 0.0356562734, 0.0348008014, 0.0140621196, -0.0222619958, -0.0201275274, -0.007960475, 0.0598417521, 0.0000092767, -0.0077383579, 0.0814854875, 0.0096133584, 0.0171389915, -0.0248118173, 0.0386050679, 0.0454702266, 0.0281078853, 0.0079574641, -0.0370293893, -0.002096893, -0.0401609689, -0.0281312894, 0.0308788586, 0.0788901597, -0.0075681792, -0.0391652547, -0.0400068015, 0.0155823641, -0.0131341545, 0.0337607674, 0.0048792977, -0.0171952192, -0.0564433858, 0.0540816262, 0.0457333885, -0.080615215, -0.0143163083, -0.0707645193, 0.0378174223, 0.0046293531, 0.0309827626, -0.0223681126, 0.0041490509, -0.0362933353, 0.0500862151, 0.045868434, 0.0804374963, -0.0068231262, -0.0043878281, -0.0190518238, -0.020781273, 0.0152326925, 0.0456606634, -0.0384288393, 0.0664085522, 0.032174632, 0.0021719011, -0.0479231887, -0.0609848239, 0.0375225879, 0.0233321916, 0.0110093616, 0.0097241467, 0.018304076, 0.0258914661, 0.0171367861, -0.005093094, -0.0323820375, 0.0391411595, 0.0021677504, 0.008262693, -0.0291701499, 0.0204112232, -0.0158100631, 0.0301310159, -0.0387225151, -0.0306501742, 0.0698778555, -0.0580008216, -0.0109778373, -0.0489870608, -0.0356066488, -0.0241209418, -0.0493725799, -0.0075701037, -0.0389656536, 0.0136368806, -0.003928469, 0.0706711933, -0.0239292309, 0.0140710715, -0.0187802557, -0.0225649402, 0.0340428464, -0.0339666121, -0.0400706045, 0.0195421223, -0.0112020038, 0.0255998373, 0.0320590697, 0.0022239147, -0.0372799002, -0.0264039077, -0.0528137796, 0.0081406971, -0.018999625, 0.0203105025, 0.0115205254, -0.0390623845, -0.0131682698, 0.0169257484, -0.0125248143, -0.0247378889, 0.0429594256, 0.0065742242, 0.0445879549, -0.0335063189, -0.0150041124, -0.0050991457, 0.0173721332, -0.0252474342, 0.0146823032, 0.0389176942, -0.0501488596, -0.0149102965, 0.0455853008, -0.0609585978, -0.0033929392, -0.0095815929, -0.0220949296, -0.004110116, 0.0354350954, -0.0171191022, 0.0312653482, 0.021730626, 0.0042260671, 0.0356376693, 0.0218785126, -0.015049289, 0.0117864572, -0.0264148954, -0.0193611477, -0.0062250164, -0.0558917411, 0.0559761785, 0.0186087266, 0.023747284, -0.0211085472, -0.0487774387, 0.020864144, -0.0306370296, -0.0055710757, 0.0096824542, 0.0109270848, -0.0200582836, 0.0473264791, -0.0328146107, -0.0305706672, 0.0669826269, -0.0324235223, 0.0422210135, -0.039992284, -0.0133884409, -0.0269326065, 0.0184346791, 0.0000999704, 0.0047097309, 0.0193024911, 0.0138902003, 0.0091680922, -0.0364233591, 0.0508715771, -0.0495106056, 0.0212341808, -0.0364894904, 0.0386533812, 0.0023208039, -0.0480422825, -0.0315577388, -0.0918938443, -0.0325482972, -0.0008034918, -0.0401552022, -0.0150918113, -0.0652790293, -0.0109354034, 0.0462482162, 0.0398434661, -0.0010039861, 0.0259556454, 0.0060631065, 0.0153138498, 0.0145529294, 0.003827268, 0.0483662449, -0.0010452195, -0.0231360756, 0.0481341146, -0.0133501943, 0.0249619912, -0.0038311931, -0.0498225391, 0.0111124478, -0.047114186, 0.0024644437, -0.0104540512, 0.0389211699, -0.048477903, 0.0058332672, 0.091252394, 0.0107733021, 0.0071877502, -0.0059527475, 0.0392762311, 0.0015729779, 0.028041508, -0.0009250938, -0.1272762418, -0.0157207064, 0.0046332139, -0.0055722003, -0.0392417163, -0.0182983857, -0.050798282, -0.1246545613, 0.0724790916, -0.0144222081, 0.009724318, -0.040028356, -0.0547376275, 0.0186566152, -0.042112343, -0.0192516278, 0.0246410538, 0.0019259447, -0.0073853252, -0.0018590239, -0.0586236268, -0.0847878084, 0.057321433, -0.0007310089, 0.0117096603, 0.0387960859, 0.0480475314, -0.0277581755, -0.0653613359, 0.0308308937, -0.0169033986, -0.0553039089, -0.0048708939, 0.0653480217, -0.0304058082, 0.0056272461, -0.0122133791, 0.0149109941, 0.035721533, -0.003332023, 0.0213081744, 0.0767911747, -0.0023927155, -0.0392202511, 0.0018941602, 0.0424924754, 0.0228118058, 0.0999977961, 0.0075442954, 0.0183757991, -0.0083532883, -0.0197457951, -0.0245257542, -0.0388804264, -0.0317523144, 0.0581443943, -0.0082434183, -0.0290711429, -0.0579292141, 0.0058714794, 0.0025962899, -0.0042312467, -0.028472513, 0.0083847903, 0.0477449372, 0.0221335273, -0.0094431685, 0.0201812182, -0.0704877749, 0.0321715511, -0.0466163307, -0.0054106684, 0.0457321815, -0.0064064572, -0.0258988664, 0.0434945151, -0.0935158059, -0.0435984842, -0.0117640086, -0.0021550988, 0.0265772212, 0.0292453133, -0.036872644, -0.0183503497, 0.0231788587, 0.0736471415, -0.0501570292, 0.0293391477, 0.0186423007, -0.0220540483, 0.0016950313, 0.0099927904, -0.015938269, 0.0319497474, 0.0215253904, -0.0033049211, -0.0218291506, -0.0297345351, 0.0288456269, 0.0727301314, 0.0458302833, 0.0025633131, -0.0378602296, -0.0555933528, 0.0030893479, 0.0215676054, -0.0428484045, -0.0008861679, 0.0462318212, 0.0029327208, -0.0297629107, -0.0458257385, 0.0106683858, -0.0333154537, 0.0397845283, -0.0514737591, 0.0233142991, -0.0096614622, 0.013236546, 0.0049023889, -0.0534565076, -0.0391166769, 0.0194789357, 0.0091958782, 0.0147390179, 0.0131868832, 0.0070033171, 0.0116319256, 0.0019372239, 0.0435493812, -0.0190845616, -0.0644894019, -0.0074670822, -0.0443148054, 0.0166585054, 0.0435506552, 0.0161793493, 0.0454803556, -0.0110004935, -0.0425764732, -0.0217303224, 0.000005264, 0.013796214, -0.0101716891, 0.035988424, 0.0010331565, -0.0440169051, 0.0186017249, 0.0394912846, 0.008141906, 0.0491676144, 0.0382584967, -0.0537693091, -0.0646066964, -0.0793144107, -0.0589114614, -0.0100932177, -0.0086525474, -0.005363218, -0.0766615495, -0.0372204743, 0.0405713245, 0.0309545379, 0.0319497772, -0.000851333, 0.0015019306, 0.0682732612, -0.0037097873, 0.0072022635, -0.0545894392, -0.0354161225, 0.0059938799, -0.001854014, -0.001800721, -6.237267065e-33, -0.0178272855, -0.0370385982, 0.0182056725, -0.0051122522, -0.0233086515, 0.0349080339, -0.0258057527, -0.0451856703, -0.0496324599, 0.040903084, 0.0005721689, 0.0343862139, 0.0237961151, -0.0017792328, -0.0408867747, -0.0073401919, 0.0653696358, 0.0031102621, -0.0525663309, 0.0203193724, 0.0276181288, -0.011240934, 0.0600147247, -0.0440073647, 0.0284636505, 0.0182123482, -0.0590358898, 0.0266137067, -0.0594843775, 0.0525707975, -0.006005425, 0.0387908369, 0.0291065164, -0.0386698097, 0.0232702158, 0.0166404545, -0.0656718686, 0.0090714647, 0.041245643, -0.0207404848, 0.0294474736, 0.0137669481, 0.0370742306, -0.0195927918, -0.041953627, 0.0107844761, 0.0268638786, -0.0384608358, 0.0083232559, -0.0333153531, -0.0350787006, 0.0636359155, 0.0051740622, 0.0085701821, -0.0280275606, 0.0530957095, 0.0021317946, 0.0460435897, -0.0543541387, 0.0562431291, 0.0240027308, 0.0030994685, 0.0473420657, -0.0719284415, -0.0309307911, 0.0673486441, -0.0176743623, 0.0114702499, -0.0301482286, -0.0165243298, 0.004845663, 0.0268561933, 0.0491539165, 0.1279854476, 0.0141342217, -0.0672993511, -0.0983195305, -0.0057676816, 0.0430698916, 0.0173647627, 0.0319065265, -0.039313104, -0.0328009799, -0.0116311284, -0.037704438, 0.0092191156, -0.0366849974, 0.0435635671, 0.006406629, 0.0310195275, -0.0322182067, 0.0309538115, 0.0071575241, -0.0350518972, 0.006228447, 0.0465829186, -0.0457943007, 0.000136698, -0.0076836287, 0.0229663681, -0.0744867027, 0.0176267531, -0.0140983975, 0.0169225559, 0.0073831584, -0.0110336952, 0.0528712682, 0.0049889181, -0.0797439516, 0.0406333543, -0.0181935597, 0.0385790281, -0.0307082906, -0.0023779103, 0.0292007849, 0.007820488, 0.0111197503, 0.0144043071, 0.0096117118, -0.0140633518, -0.0219350625, 0.0075346273, -0.0062918747, 0.0381052494, -0.0139964474, -0.0070182444, -0.0109087313, 0.0335689522, 0.0081980983, -0.0499244668, 0.0244352594, 0.0161720421, 2.596e-7, 0.0153247174, 0.0571617633, -0.0288346987, -0.0074009835, -0.0009048248, 0.0243325364, -0.0019225343, -0.0148845259, -0.0171330944, -0.0466040075, -0.0199474543, 0.0256847199, 0.0548981242, 0.0395183302, -0.0746222734, -0.0028413171, -0.0691868886, -0.0306086186, -0.0563910417, -0.0093272384, -0.0077229417, 0.0436734371, 0.0397356153, -0.0265534315, -0.0160613731, -0.0249081179, 0.0277508646, -0.0045272373, -0.0623065494, 0.027399024, 0.0702435523, -0.0090808477, 0.017954668, 0.047997985, -0.0281639975, 0.0131371683, -0.0158467777, 0.0935845748, 0.0472714603, 0.0013669407, -0.0323364586, 0.0023657889, -0.010070744, -0.0486335792, 0.0751217753, -0.0042340634, -0.0292637348, -0.0637289733, -0.0281999502, -0.031387195, 0.039292112, -0.0306016114, -0.0224333294, -0.0122860251, -0.0248344652, -0.0020956867, -0.0209613945, -0.0211918298, 0.019723976, -0.0090445783, -0.0428217202, -0.0966328606, 0.0352000967, -0.0136113279, 0.0197325181, -0.0686886981, 0.0244400464, 2.388836635e-34, 0.0405286662, -0.0146803427, 0.0032511631, -0.0654714555, 0.0670780763, 0.0513810627, -0.0131110419, -0.0163492858, -0.0168021396, -0.0232400093, 0.0045168763 ]
http://arxiv.org/abs/2406.18294v2
http://arxiv.org/pdf/2406.18294v2
Hierarchical Context Pruning: Optimizing Real-World Code Completion with Repository-Level Pretrained Code LLMs
2024-06-26T12:26:16
2024-06-27T04:40:52
Lei Zhang, Yunshui Li, Jiaming Li, Xiaobo Xia, Jiaxi Yang, Run Luo, Minzheng Wang, Longze Chen, Junhao Liu, Min Yang
cs.CL
cs.CL
Some recently developed code large language models (Code LLMs) have been pre-trained on repository-level code data (Repo-Code LLMs), enabling these models to recognize repository structures and utilize cross-file information for code completion. However, in real-world development scenarios, simply concatenating the entire code repository often exceeds the context window limits of these Repo-Code LLMs, leading to significant performance degradation. In this study, we conducted extensive preliminary experiments and analyses on six Repo-Code LLMs. The results indicate that maintaining the topological dependencies of files and increasing the code file content in the completion prompts can improve completion accuracy; pruning the specific implementations of functions in all dependent files does not significantly reduce the accuracy of completions. Based on these findings, we proposed a strategy named Hierarchical Context Pruning (HCP) to construct completion prompts with high informational code content. The HCP models the code repository at the function level, maintaining the topological dependencies between code files while removing a large amount of irrelevant code content, significantly reduces the input length for repository-level code completion. We applied the HCP strategy in experiments with six Repo-Code LLMs, and the results demonstrate that our proposed method can significantly enhance completion accuracy while substantially reducing the length of input. Our code and data are available at https://github.com/Hambaobao/HCP-Coder.
Hierarchical Context Pruning: Optimizing Real-World Code Completion with Repository-Level Pretrained Code LLMs: Some recently developed code large language models (Code LLMs) have been pre-trained on repository-level code data (Repo-Code LLMs), enabling these models to recognize repository structures and utilize cross-file information for code completion. However, in real-world development scenarios, simply concatenating the entire code repository often exceeds the context window limits of these Repo-Code LLMs, leading to significant performance degradation. In this study, we conducted extensive preliminary experiments and analyses on six Repo-Code LLMs. The results indicate that maintaining the topological dependencies of files and increasing the code file content in the completion prompts can improve completion accuracy; pruning the specific implementations of functions in all dependent files does not significantly reduce the accuracy of completions. Based on these findings, we proposed a strategy named Hierarchical Context Pruning (HCP) to construct completion prompts with high informational code content. The HCP models the code repository at the function level, maintaining the topological dependencies between code files while removing a large amount of irrelevant code content, significantly reduces the input length for repository-level code completion. We applied the HCP strategy in experiments with six Repo-Code LLMs, and the results demonstrate that our proposed method can significantly enhance completion accuracy while substantially reducing the length of input. Our code and data are available at https://github.com/Hambaobao/HCP-Coder.
[ 0.009151618, 0.0930035636, -0.0421765707, 0.0023219844, 0.0005203794, 0.0099442806, -0.0244821459, 0.0360025242, 0.0008118962, 0.044878263, -0.0539044179, 0.036867056, 0.0325152613, -0.0171447378, 0.0050364141, -0.0106877154, 0.0390214324, -0.0172366817, -0.0425020568, -0.0134036103, -0.0376481265, 0.016569959, 0.0383396819, 0.0538668856, -0.0006641299, -0.0537356883, 0.004087097, 0.0367514826, -0.032597106, -0.0026002955, 0.0312989801, 0.0410772003, 0.00017789, 0.0187371317, 0.0000023954, -0.0256574601, -0.0094623612, -0.0141774109, -0.0268712081, 0.0290145744, -0.0172330551, -0.0181936678, 0.0065940763, 0.0008659941, -0.0087354593, 0.0016471804, 0.0332803242, -0.0121439286, 0.0227420889, 0.0195697639, 0.0292358696, -0.0178082772, 0.0059877597, -0.0126361297, 0.0735821277, -0.0150760291, -0.0017036148, 0.0487116352, 0.0424267836, 0.0075675012, 0.0010779225, 0.0187809393, 0.0466339774, 0.0166344494, 0.0297282021, 0.0546765178, 0.0041979137, -0.0291826073, 0.0315041468, 0.0544964969, -0.0741818473, -0.0335021205, 0.007005536, -0.0343503654, -0.0535774194, -0.0501559973, -0.0288534183, -0.0029879787, -0.0232741814, -0.0020094814, -0.038893234, -0.031054955, 0.0226715393, 0.0082451273, -0.035797786, 0.0581395812, -0.0026887706, -0.0034877369, 0.0046952148, -0.0273447484, -0.0132872509, -0.0629024357, 0.0016526926, 0.0414311923, -0.0284805335, 0.0183510445, -0.0160055924, -0.0093858251, 0.0631138384, -0.0304901004, -0.0543748029, 0.0518911369, -0.0686581507, 0.0134052597, -0.0474140234, 0.0752784759, -0.0304368753, -0.0420114808, -0.0376223698, -0.0069043026, -0.0232528336, 0.0006502028, -0.056245517, 0.0393816568, -0.0409874283, 0.0204741955, -0.0368130915, 0.0242018588, 0.0076953089, 0.0578393601, -0.0328986458, -0.0152940117, -0.0499700084, -0.001817262, 0.0279471911, -0.0298502836, -0.0415617153, -0.0299467947, -0.0480532087, -0.0149916047, 0.0071150712, 0.0062218956, -0.0457462333, -0.0546760485, 0.0326849967, 0.0433564112, 0.0504329875, 0.0084048957, -0.0267169308, -0.0339403674, 0.0080948118, -0.0247823596, -0.0311531536, -0.0915891007, -0.0427791961, 0.0274547692, -0.0007486705, -0.0819944069, -0.0253704581, -0.0010521875, -0.0049440316, 0.0175636057, -0.0645753816, 0.0282474682, 0.061634738, 0.0297494959, -0.0369215608, 0.0354537703, 0.0169497281, 0.0543758795, -0.0006830316, -0.0630902648, 0.0256834626, 0.0152809899, -0.0177792199, 0.0077695423, 0.0123432809, 0.0325683057, 0.0322433785, -0.0221911799, 0.0604358651, 0.0137811285, -0.0093030771, 0.0012393002, 0.0078065312, 0.0191295892, 0.0832698718, 0.0272278655, 0.0241456497, 0.0589572005, 0.0076672472, 0.0031639922, -0.0669889003, 0.038321808, -0.0231833886, 0.0218687057, -0.018207334, 0.0360998251, 0.0041643209, -0.0517293476, 0.0096321478, -0.051942315, -0.0436185785, 0.034760464, 0.034208253, -0.0169900395, -0.0243066568, 0.011632666, 0.0479571, -0.0351743326, -0.0037481717, -0.008017811, 0.053159181, 0.0455503948, 0.0352483988, -0.0679286569, -0.0196137615, -0.0045070658, 0.0174077637, -0.0698098466, -0.0084908912, -0.0133046415, 0.0236767512, -0.0370275751, -0.0468475111, -0.0311211515, -0.030038666, 0.0492209308, 0.0193704814, -0.0426741019, 0.0143593745, 0.0099399937, -0.031342037, 0.0470483303, 0.0138785057, 0.032810349, 0.0264639799, 0.0154844886, 0.0212917347, -0.0354121737, -0.0382828414, -0.035310775, 0.0589543395, -0.0031935056, -0.0188795663, -0.0507116579, -0.0486136377, -0.0158593338, 0.015909614, 0.0001229325, -0.0159986969, -0.0191055574, -0.0181364678, 0.0035620388, -0.023288466, -0.0230971798, 0.0588762648, 0.0102391783, 0.0286090653, -0.0433880128, 0.0063167131, -0.0494279452, 0.0539103895, 0.0243527237, 0.050943993, 0.0513135195, 0.0802915469, 0.0532523729, 0.0178838689, -0.0173402578, 0.021604307, -0.0331542157, -0.0013236743, 0.0182358734, 0.0445583574, -0.0180248339, -0.0483118854, -0.0858510286, 0.0074544228, -0.0088451765, -0.0014231468, -0.0054585277, 0.0300780013, -0.0107230628, -0.0420712605, 0.0284196418, 0.0215958133, -0.0299864579, -0.0018727698, 0.0340783074, -0.004695585, -0.0452953577, -0.0292601306, -0.0183711573, 0.0036762774, 0.0439969376, -0.0286250953, -0.0471795164, -0.0499113463, 0.1017275453, 0.0224750228, 0.032133311, -0.0157669913, -0.0299855098, 0.0064297891, 0.0154413218, 0.097277984, 0.0350382999, -0.0172647778, -0.0499085076, 0.0851113796, 0.0315089896, -0.0686892197, 0.0184419341, 0.0400730968, -0.0575389415, -0.0262643509, -0.0005980924, 0.020927893, 0.0345987044, 0.0047897543, -0.0073192278, -0.021814717, -0.0494038388, -0.0108614191, 0.0024486603, 0.0149920909, -0.050190784, -0.0131953442, 0.0763168111, -0.0296798833, -0.1108090207, 0.0376652926, -0.0614958704, -0.047559835, 0.0049666008, -0.0153398942, -0.0528683998, 0.0689400285, 0.0312455352, -0.0021609075, -0.0025583266, -0.0634046718, -0.0237415005, 0.0348649882, -0.0637686551, 0.0123904711, 0.0402964316, -0.0311714597, 0.0044136797, -0.01216779, -0.0271077044, -0.0273449477, -0.0020392819, 0.0006682642, -0.0075489753, 0.0169356596, -0.006933569, 0.0589618795, 0.0081604943, 0.0262217969, 0.0609307624, 0.0676585585, -0.0448681489, -0.0164051689, -0.0485901497, -0.0107556116, -0.0295365769, 0.0067573786, 0.0270660948, -0.0095172906, 0.0009480233, 0.0760862827, 0.0619276315, -0.0592328236, -0.012460527, 0.0503593907, -0.0229154266, -0.0964626372, 0.0633729771, 0.0499046817, -0.0387277752, 0.0127303638, 0.0911342353, 0.0415625796, -0.0658199638, 0.0114598908, 0.0076363538, 0.041166842, 0.0764091611, 0.0548008457, -0.0922297686, 0.0059818113, 0.0034674627, -0.0156410076, 0.021642914, 0.002549883, 0.0040371409, -0.0474917777, 0.0295290612, 0.006893306, -0.0096860984, -0.0171452332, -0.0286219623, 0.0180060416, 0.0267299954, -0.046338141, 0.0357921571, -0.0061088502, -0.0057324446, -0.007672179, -0.0755273402, 0.035853412, 0.0610010549, -0.0223953705, -0.0076919259, 0.1039014608, 0.0068334728, 0.0169307534, 0.02332443, 0.0279928613, -0.0170158241, 0.0281176623, -0.0398848392, 0.031181626, 0.0168916658, 0.0074773729, 0.0602866076, -0.0026228514, -0.0252920445, -0.0362527519, 0.0369237922, 0.0725501701, -0.0406135768, -0.0220245328, -0.0075564631, -0.0153069571, -0.0125662852, 0.0147450268, 0.0641442612, -0.0247489829, -0.0141626177, -0.0407464579, 0.0035054623, -0.0161343589, 0.0572615601, -0.0182067342, -0.0243726969, -0.0107682599, -0.0132889086, -0.0074310862, 0.0358141996, 0.0275205094, 0.0161992218, 0.0036791936, -0.014539741, 0.0074174646, -0.0301253498, 0.0172091816, -0.0123079019, 0.029046176, -0.0488992408, -0.0206585806, 0.0296721309, -0.0330599956, -0.0240632948, -0.0275330413, -0.05946986, -0.0176137649, -0.0453106277, 0.0273855738, 0.0125931744, 0.0286508258, -0.048244372, -0.1193911359, -0.0329658575, -0.0083570378, -0.0819741711, 0.0412500091, 0.0318095572, 0.0022803429, 0.0074496102, 0.0021044104, -0.0724110827, -0.0001528235, 0.0311226621, -0.0380547084, -0.022341365, -0.007739245, 0.0191943329, 0.0017084673, 0.0329868048, 0.0200806763, -0.0188309532, -0.0145295225, 0.0331678502, 0.033572007, 0.0012699923, -0.0105868429, 0.0109973606, 0.0256475322, -0.0008051258, 0.0481176972, -0.003651286, -0.012637849, 0.0297722351, -0.0133376643, 0.0345365852, 0.0160295554, 0.0417696945, 0.0989737883, 0.0260416828, -0.0233845133, -0.0117056388, 0.007470456, 0.0199891683, -0.033357203, 0.019367205, -0.0041137077, 0.0296256803, -0.0001071673, -0.0142114712, -0.0250367355, -0.057163246, -0.0249664895, 0.0267676804, 0.0451350324, 0.0426514149, -0.006410805, -0.0079771541, 0.0193765555, -0.0455889404, 0.031321235, 0.0113830604, -0.07018134, -0.0013011139, 0.0117195845, -0.0382960774, -0.011087765, 0.0891797021, -0.0055518821, 0.0317430161, -0.0348640904, -0.0011215252, -0.0151457135, 0.0206815619, 0.018637158, 0.0187584199, -0.0616945401, -0.0263566896, -0.0169232991, -0.0407763124, 0.055215735, 0.0073758345, -0.0151143735, 0.00524254, 0.0239042193, 0.1005618423, -0.0020001396, -0.050299339, 0.0024723548, -0.0148759913, 0.029967647, -0.036924012, -0.067915611, -5.282744346e-33, -0.0278729461, -0.0458274782, 0.0137474118, 0.0103111621, 0.0043513207, -0.0016983333, -0.0350428671, -0.0121217705, 0.0064327717, -0.0692132264, -0.0134902624, 0.0071327533, -0.015218663, -0.0491413102, 0.0616134293, 0.0632462651, 0.0623268597, 0.0417560413, -0.0363588594, 0.0478143655, -0.0631062388, 0.0639642924, -0.0021833756, -0.0443734787, -0.0570874736, 0.0070043704, 0.0601305887, -0.0176150631, 0.0217060801, 0.0541188084, -0.0167704746, 0.0096665183, -0.003529212, 0.0005367557, -0.0458891094, -0.0010737871, -0.0596166588, -0.0405913964, 0.0066659306, 0.0309932847, 0.0279552657, -0.0714199468, 0.0763738006, 0.0087433131, -0.0006349681, 0.0037259914, 0.0285596829, -0.0114900777, -0.0425976478, -0.0172365848, -0.0530139655, 0.0372441225, -0.0104140723, 0.0207271855, 0.0166844446, 0.0214182734, 0.0148507357, 0.0726803616, 0.0145424884, -0.0226361789, -0.0040620742, 0.0779952109, 0.0593870021, -0.013114403, 0.0053172363, -0.0114619434, 0.0523904152, 0.0160980541, -0.0062440005, 0.0045451811, 0.0459259115, 0.0448399037, 0.0043410594, 0.0005104874, 0.0289241578, -0.0321797654, -0.0254464857, 0.0153714437, -0.0146682551, 0.0619635582, -0.01898643, 0.0706709027, -0.0169784594, -0.0012790543, -0.035110347, 0.0235770978, -0.0159625933, -0.0611099452, 0.0208393112, -0.0047252281, 0.0074595576, 0.0289533846, 0.0043008239, -0.0299552623, 0.001190725, 0.0307807978, -0.0473973379, 0.0548056252, 0.0020530608, 0.0289264005, -0.0171141867, -0.011948796, 0.0154194729, -0.0464761443, 0.0704770237, -0.0010271757, 0.0312617421, 0.0175558366, -0.0590693951, 0.0031897626, -0.0365245491, -0.0550010204, 0.0002190132, 0.0084195379, 0.0046869991, -0.0096905334, 0.0240747705, 0.0325167961, -0.0308134165, -0.0259018857, -0.0881396011, -0.0062568681, -0.0020542424, 0.010549197, -0.0257771164, -0.0600907095, -0.0272536017, 0.056103617, 0.0515852198, 0.0080409683, -0.0566326082, 0.0382085443, 2.974e-7, -0.012296496, 0.028214816, 0.0053759357, 0.0169066302, -0.0270044934, -0.019073857, -0.0224782471, -0.0360540226, -0.0003164972, -0.0156463366, 0.0062155668, -0.0222432278, 0.0467858873, 0.0266295709, -0.0808581784, -0.0118579781, 0.0059282319, -0.0101788342, -0.0929445475, -0.008269093, -0.0015549943, 0.0755584836, 0.0043484229, 0.0011209452, 0.0345342308, 0.001791165, -0.0079528941, -0.0255544521, 0.0376747996, 0.0584527068, -0.057904806, -0.0080748592, -0.0168838836, -0.0215038117, 0.0106074708, -0.023420047, -0.0066126641, 0.0482951738, 0.0088881971, 0.0310589857, 0.0248995963, -0.0113187404, -0.0344247594, -0.0481328927, 0.0292991046, -0.0339578353, 0.0463081524, 0.004950881, -0.006488251, -0.0340320244, 0.0122453533, -0.0037656203, 0.0324502513, -0.0050612823, 0.0024850185, -0.0510960743, -0.0438016206, -0.0471300557, -0.0095602591, 0.0370275192, -0.0522015207, -0.0239745416, -0.0655533895, 0.0048579229, 0.0261820853, -0.0126508391, 0.0170486774, 2.707112521e-34, 0.0053278646, -0.0539368317, 0.0093819527, 0.018392561, 0.0404624678, -0.018670354, -0.0103111882, 0.0440037549, -0.0060844584, -0.01791583, -0.0253803153 ]
http://arxiv.org/abs/2406.18276v1
http://arxiv.org/pdf/2406.18276v1
Sanskrit Knowledge-based Systems: Annotation and Computational Tools
2024-06-26T12:00:10
2024-06-26T12:00:10
Hrishikesh Terdalkar
cs.CL
cs.CL, cs.SE
We address the challenges and opportunities in the development of knowledge systems for Sanskrit, with a focus on question answering. By proposing a framework for the automated construction of knowledge graphs, introducing annotation tools for ontology-driven and general-purpose tasks, and offering a diverse collection of web-interfaces, tools, and software libraries, we have made significant contributions to the field of computational Sanskrit. These contributions not only enhance the accessibility and accuracy of Sanskrit text analysis but also pave the way for further advancements in knowledge representation and language processing. Ultimately, this research contributes to the preservation, understanding, and utilization of the rich linguistic information embodied in Sanskrit texts.
Sanskrit Knowledge-based Systems: Annotation and Computational Tools: We address the challenges and opportunities in the development of knowledge systems for Sanskrit, with a focus on question answering. By proposing a framework for the automated construction of knowledge graphs, introducing annotation tools for ontology-driven and general-purpose tasks, and offering a diverse collection of web-interfaces, tools, and software libraries, we have made significant contributions to the field of computational Sanskrit. These contributions not only enhance the accessibility and accuracy of Sanskrit text analysis but also pave the way for further advancements in knowledge representation and language processing. Ultimately, this research contributes to the preservation, understanding, and utilization of the rich linguistic information embodied in Sanskrit texts.
[ 0.0374806188, -0.0555517785, -0.0492420867, 0.0893947929, -0.0203987937, -0.0056908894, -0.0570229031, 0.0323709995, 0.0248383135, -0.0393331535, 0.0095771905, 0.0126093002, -0.0233046748, -0.0165108219, 0.0134451166, -0.0335220844, 0.0256767608, -0.0227236897, 0.0515659973, 0.0094050076, 0.0738159791, 0.0378773585, -0.0277514402, 0.0066583599, -0.0419301391, -0.0327237919, 0.0065386649, -0.0443221703, 0.0182664674, -0.013302274, -0.0405422114, 0.0339095928, -0.0214024223, 0.0251187906, 0.0000014437, -0.0727265552, -0.0474889912, -0.0000166395, -0.0279673841, -0.0239826273, 0.0101089897, 0.0167155396, 0.0461658016, 0.0008052278, -0.056624949, 0.0339762159, 0.0763986483, -0.0535216145, 0.0442803651, 0.03268224, -0.0236512665, -0.0009813331, 0.0312248692, 0.0345246308, -0.0084281974, 0.022665387, 0.0075222887, 0.061139334, 0.014085955, -0.0286603589, -0.0273714233, 0.0449441858, -0.0255409274, -0.0180667769, 0.0343234241, -0.0369536988, -0.0611806139, -0.0058478578, 0.0254265741, -0.0024746577, 0.066718027, -0.031848561, -0.0013834352, 0.0275464393, -0.0504600331, -0.0055416967, -0.0381612442, 0.013185055, -0.0589049272, 0.0196882822, 0.0567808561, -0.0064790794, -0.0380625837, 0.0152898645, 0.0585685521, 0.0227443781, -0.003198318, -0.0130981272, -0.0054689953, -0.0248546377, -0.0600289591, -0.0351692252, 0.0279782917, 0.0826626644, -0.0037900172, 0.0694046319, 0.0052825832, -0.0796995312, 0.0317103676, -0.0231191386, -0.0077286833, 0.0234380662, 0.0026386238, 0.000241653, 0.0067600207, 0.0316130109, 0.0390226804, 0.011707915, -0.0273450874, -0.0426844321, 0.0164747424, -0.0717444941, -0.0758230537, 0.014091, 0.004399918, 0.0362250134, -0.0337780304, 0.0229944345, -0.0216796007, -0.0087056896, 0.0167411882, -0.0235188901, 0.0095511535, -0.0052264961, -0.0409697667, 0.007133841, -0.0162080564, -0.0116330376, -0.0200937781, -0.0272364896, 0.0259507131, 0.0059530591, 0.013536986, -0.0060473206, -0.0367542319, 0.0217798427, 0.0235927068, -0.0225996431, -0.0386159681, -0.0109311761, -0.0247813724, -0.0310371537, -0.0126788411, -0.0790922642, 0.0577849075, 0.0604084507, -0.0332675166, -0.0591455773, 0.0465743318, 0.0040364284, -0.0169549026, 0.0663111731, -0.057415247, -0.0188238993, 0.0154760415, 0.0043115299, 0.0185600612, -0.0236380007, 0.0362139791, -0.0075454777, 0.0900139585, -0.0041227685, 0.0131474268, 0.0161901861, -0.0045055202, -0.032506492, 0.00470777, 0.0666217878, -0.0194572024, 0.0265197866, -0.0494407974, 0.046049159, 0.0196662229, 0.0284222048, -0.0136310784, 0.0753256157, -0.0047522285, 0.0623265877, 0.0023353202, -0.0021212883, 0.058252383, -0.0190958157, -0.0157130677, 0.0354157351, 0.0101781003, 0.0154654766, -0.0141481971, -0.055514507, 0.0109176654, -0.0460522287, -0.0650985315, -0.0339218639, -0.027856987, -0.0269367043, 0.0504165106, 0.0019685505, 0.0047665946, 0.0121156443, 0.0540968403, -0.0563940778, 0.0460282639, -0.0283643249, 0.0663557053, 0.0588318408, 0.0224545151, -0.0733867735, -0.0118078217, 0.0694968924, -0.0064895051, 0.0361306779, 0.0079559823, -0.0354381241, -0.0610456392, -0.0159485377, 0.0381769091, 0.0218934771, -0.0199671462, 0.0126633458, -0.0129277641, 0.01514214, -0.0090728831, -0.0450415239, -0.0188462231, 0.025731273, 0.0362183414, 0.0025047385, -0.0459871665, 0.0113811539, 0.0389079712, 0.0154541861, 0.0192212779, 0.0264808219, 0.0413318351, 0.0057550892, -0.0158856567, 0.0003859033, 0.0573732443, -0.0159808062, 0.0281443801, -0.0369644724, 0.0149010746, 0.010993707, 0.0179798882, -0.0482948348, -0.0046412055, -0.0399041437, 0.0508459732, -0.0339488387, 0.0025420287, 0.0461014509, -0.0237722509, -0.049250789, 0.0404977016, 0.0462654792, 0.0258536469, -0.0185423028, -0.0703071877, 0.0434257276, -0.0443507768, -0.024801543, -0.0004125651, 0.0249076467, 0.0069402209, -0.0065582097, -0.0283293314, -0.0400409177, 0.0202496089, -0.1237587929, 0.0042510862, 0.0017690253, 0.0672950894, -0.0094077727, 0.0286312159, -0.0116293291, -0.0260932464, -0.018146785, 0.0372324735, 0.070584394, -0.0244220849, 0.0266673453, -0.0002364834, 0.0148990126, -0.0560897589, -0.0122832041, -0.031960547, 0.0190419853, 0.0078396928, -0.0754878893, -0.00912691, -0.0108326031, -0.0744023696, 0.0281258672, -0.0014023564, -0.0081839645, -0.0345210992, -0.0109174894, -0.0199861489, -0.0200513303, -0.0044654068, 0.0084772334, 0.0245629288, -0.0277247299, 0.0033575606, -0.0092404196, 0.0125045432, 0.0195529275, 0.0071591046, 0.006112319, -0.007463797, 0.005845719, 0.0197525565, -0.0173981059, 0.0046380353, 0.0060444381, -0.0691128075, 0.046644561, -0.0126542877, 0.0111062275, -0.0390933417, 0.0630212203, 0.0076630055, -0.0283568259, 0.0666260794, -0.0178817939, 0.0064152344, -0.0620737001, 0.0326088108, -0.0328461267, 0.0534823053, -0.0076416368, 0.0190284252, 0.0043336395, 0.0070034065, -0.025263967, 0.0023744272, 0.0376936086, -0.0524325222, 0.010898686, -0.0467617698, 0.0504340641, -0.0407426134, 0.0287097655, -0.0110794259, 0.0331183262, -0.0180724598, -0.0029011904, 0.0462256931, 0.0207498427, 0.0753957927, 0.0640898943, -0.0278817955, -0.0302121658, -0.0128985895, 0.014308366, 0.0050693569, 0.053490594, 0.0526393279, 0.0311284736, 0.0156165836, 0.0799176022, 0.0124040702, 0.0094117429, 0.0111357439, -0.0318381451, 0.0314447768, -0.0283783097, 0.0352019817, -0.0765891671, -0.0049681794, -0.0170325059, -0.0196306463, -0.0346815027, 0.0340081453, 0.0594972149, 0.0844406188, 0.0025554916, -0.0098589212, -0.0780525655, -0.0487320386, 0.0110078007, 0.0408668816, -0.1365463585, -0.0009116434, -0.025881175, 0.0010328564, 0.0103048021, -0.0187473055, -0.0062352545, -0.064235203, 0.000693984, -0.0423556268, -0.0397670716, 0.0096396552, 0.01540446, -0.024433149, -0.0099264449, -0.0395474881, 0.0277066175, -0.0408868231, 0.019099839, 0.0247842651, 0.0074163168, 0.0559747778, 0.0070418688, -0.0238711797, -0.0055782683, 0.0128755374, 0.0524294227, -0.0097067049, -0.006487919, -0.0343411043, -0.0115023879, -0.0052768281, -0.0359557122, 0.0287431944, -0.0038111401, -0.0142081399, -0.0466039963, 0.0017846643, -0.0181194395, -0.0256853215, 0.0810954869, 0.0881498083, -0.0092615932, 0.0046366476, -0.0016837625, -0.007247753, -0.024614092, 0.021978559, -0.0194036644, 0.0447448529, -0.0195965823, -0.0401866175, 0.0304091908, -0.0103906477, -0.0538744703, 0.0096366499, -0.0471206941, -0.0013260838, -0.0628346875, -0.041944623, 0.0726151019, -0.0111270072, 0.0329688229, 0.0431085601, 0.0328102484, 0.0579857826, -0.0776206478, -0.049147442, -0.0406544171, 0.0083032642, 0.0047029816, -0.0145564554, 0.0208313074, 0.0014785908, 0.0001262371, 0.0119201085, 0.0363475494, -0.0250274763, -0.013419616, 0.0386030152, 0.0201483257, -0.0128451074, -0.0435111821, -0.0261463001, -0.0049606576, -0.014099081, -0.0553146116, -0.0052349074, 0.0214857925, 0.0069477088, -0.024069868, 0.0326583274, -0.0386138111, -0.0265395977, -0.0479412153, -0.0028408123, 0.0050859586, -0.0111863557, 0.0618369915, 0.0583967231, -0.0169354994, -0.0686655566, -0.004414707, -0.0116225472, 0.002630719, -0.0255434066, -0.0422177911, -0.0160338841, 0.0149530843, -0.0284912139, 0.0259912536, -0.0278057959, 0.021639647, 0.0086955912, -0.0055226777, 0.034119226, 0.0534948669, 0.0736834779, 0.031849727, -0.0130106295, 0.0041850405, 0.0003373485, -0.0069616721, -0.0167010296, 0.0013186316, -0.0106391096, 0.0519111007, -0.0193791222, -0.0298345406, 0.0311446935, 0.0058747754, 0.0421096981, -0.0008046622, 0.0163128898, 0.022596309, 0.0549202338, 0.0134458411, 0.0465398952, -0.0632930845, 0.0035528655, 0.012985711, 0.0054007, 0.0179560483, 0.0077300584, 0.0331975855, 0.016742928, 0.0244037304, -0.0116616627, 0.0671116188, -0.0537633263, 0.0141187645, 0.0293411631, -0.0895549282, -0.0266386364, -0.0612328388, -0.0395469218, 0.0241258983, 0.036678385, -0.0203159507, -0.0701980069, -0.0596841425, -0.0131427916, 0.0145792235, -0.0078549264, -0.0233174972, 0.0502773672, 0.0301250499, 0.0645209253, -0.0330124237, -0.0481527559, 0.0484942943, -0.1135489792, -0.0066241054, 0.0231623519, -4.187087368e-33, -0.0661089197, 0.0013814901, 0.0089493515, 0.0729545578, 0.0183907486, -0.0074633225, -0.0481511839, 0.0015481896, -0.0233023744, 0.021082947, -0.0239651371, 0.0020225225, 0.0093544861, -0.0297147296, 0.0108179236, -0.0179261919, -0.049370978, 0.0136099653, 0.0072282283, -0.0153076947, 0.0784664825, -0.0357931964, 0.0824938789, -0.0099503072, 0.0797383934, -0.028983159, -0.0150878578, 0.0428898782, -0.0334654972, 0.0537370145, -0.0548403971, 0.0808848515, 0.018819876, 0.001398716, 0.0093434975, 0.0124270646, -0.0910294279, -0.044132106, 0.0727662668, 0.038404882, -0.0371513851, -0.0222614612, -0.0195333455, 0.0057763848, 0.0186993647, 0.0667394325, 0.0556155369, -0.0136078261, 0.0089741126, -0.0729848593, -0.0619603097, 0.0209653694, -0.031040011, -0.0229137838, -0.0032080477, -0.0151734836, 0.0192342494, -0.0300178807, -0.0326306336, -0.0095343124, 0.0022664883, 0.0260463431, 0.0294170119, 0.0041765398, -0.0510275327, -0.0415325463, -0.0128007969, -0.086021319, 0.0439694896, 0.0649092644, -0.0193756241, 0.0443652794, 0.0610445663, 0.0324957706, -0.0032910933, -0.0131661482, -0.0344441272, -0.013387247, -0.0281131919, 0.0447687469, -0.0177641753, -0.0143118668, -0.0164105333, -0.0292081703, 0.0118791629, -0.0284840334, -0.0276745167, -0.0240481868, 0.0232532043, -0.054356344, -0.0264126845, 0.0454220511, -0.0084216874, -0.0470137484, 0.0346252173, 0.0391708091, 0.0116393548, 0.0308819842, -0.0213809218, 0.0286062919, -0.030965725, 0.0181869939, -0.0221610777, -0.0285685547, -0.0199450087, -0.0174289364, 0.0125236074, -0.0149729438, -0.025589535, 0.0153214885, 0.007285614, 0.0487031005, -0.0410740301, -0.0100992834, -0.008479449, 0.0425485, 0.0287205577, -0.0063062753, 0.065503791, 0.0380943678, -0.0528482758, 0.0177492145, 0.0150656607, -0.0510360934, 0.0035505674, 0.0183881037, -0.0253822319, 0.0336679332, -0.0127398288, -0.0184225943, 0.0394153967, 0.0245455001, 2.217e-7, 0.0838552415, 0.0265676603, 0.0455732718, -0.0163311176, -0.0008082524, -0.0275945179, 0.0097838454, -0.0221121404, -0.009739629, 0.0157917328, 0.0132110659, 0.0358639695, -0.0047150571, 0.0803923309, -0.0810831785, -0.0316743664, -0.0275994968, -0.0907611027, -0.0544121005, 0.0522694141, 0.0159656312, 0.0188396759, 0.0458027385, 0.0029741144, -0.0611873306, -0.033510115, 0.0251445603, 0.0299473293, 0.0345485173, 0.0453384183, 0.0089969197, -0.0559406541, -0.0008106875, 0.0226126574, 0.0150844352, 0.0296666864, 0.0181919057, 0.0306592043, -0.0179434735, 0.014641243, 0.0123708537, -0.016005436, 0.0537353307, -0.0346870013, 0.0701540932, 0.01629664, 0.004463478, 0.02777211, 0.0157460924, -0.0018440377, -0.0336162709, -0.0031727867, -0.019695038, -0.0046447758, 0.0026612463, -0.0590496771, -0.0121328849, -0.0345280059, -0.0097150123, 0.0642307922, 0.0254807957, -0.0680825859, 0.007382934, -0.0129120359, 0.0264673568, -0.0771039724, 0.005474587, 1.29838425e-34, 0.0150066856, 0.0004457294, 0.0164234042, 0.005812957, 0.0055018817, -0.0129040349, 0.0008362889, 0.0351813398, 0.0187310409, -0.0519995205, -0.0225061011 ]
http://arxiv.org/abs/2406.18266v2
http://arxiv.org/pdf/2406.18266v2
"Vorbeşti Româneşte?" A Recipe to Train Powerful Romanian LLMs with English Instructions
2024-06-26T11:39:51
2024-06-27T20:30:47
Mihai Masala, Denis C. Ilie-Ablachim, Alexandru Dima, Dragos Corlatescu, Miruna Zavelca, Ovio Olaru, Simina Terian, Andrei Terian, Marius Leordeanu, Horia Velicu, Marius Popescu, Mihai Dascalu, Traian Rebedea
cs.CL
cs.CL
In recent years, Large Language Models (LLMs) have achieved almost human-like performance on various tasks. While some LLMs have been trained on multilingual data, most of the training data is in English; hence, their performance in English greatly exceeds other languages. To our knowledge, we are the first to collect and translate a large collection of texts, instructions, and benchmarks and train, evaluate, and release open-source LLMs tailored for Romanian. We evaluate our methods on four different categories, including academic benchmarks, MT-Bench (manually translated), and a professionally built historical, cultural, and social benchmark adapted to Romanian. We argue for the usefulness and high performance of RoLLMs by obtaining state-of-the-art results across the board. We publicly release all resources (i.e., data, training and evaluation code, models) to support and encourage research on Romanian LLMs while concurrently creating a generalizable recipe, adequate for other low or less-resourced languages.
"Vorbeşti Româneşte?" A Recipe to Train Powerful Romanian LLMs with English Instructions: In recent years, Large Language Models (LLMs) have achieved almost human-like performance on various tasks. While some LLMs have been trained on multilingual data, most of the training data is in English; hence, their performance in English greatly exceeds other languages. To our knowledge, we are the first to collect and translate a large collection of texts, instructions, and benchmarks and train, evaluate, and release open-source LLMs tailored for Romanian. We evaluate our methods on four different categories, including academic benchmarks, MT-Bench (manually translated), and a professionally built historical, cultural, and social benchmark adapted to Romanian. We argue for the usefulness and high performance of RoLLMs by obtaining state-of-the-art results across the board. We publicly release all resources (i.e., data, training and evaluation code, models) to support and encourage research on Romanian LLMs while concurrently creating a generalizable recipe, adequate for other low or less-resourced languages.
[ 0.0031976947, 0.0522010326, -0.0010537537, 0.0552376546, -0.0214062408, 0.0245311316, -0.0012293118, 0.0130275255, 0.0128840115, -0.0343461521, -0.0570191517, -0.0613981336, -0.0058697597, -0.0346041098, 0.0522052534, -0.0311663914, 0.0614646114, -0.0387083329, -0.0529650711, -0.0106310332, -0.0019838361, -0.0064543094, -0.033778131, -0.001651578, -0.0001278678, 0.0163105931, 0.0149642369, 0.0181981605, 0.0267073289, -0.0109435916, 0.0117750578, -0.0026221077, 0.0325000994, 0.036290504, 0.0000017314, -0.0569058061, -0.0005906658, -0.0247526728, 0.0161115248, 0.0279533938, 0.0525358021, -0.0495702848, -0.0332397819, 0.0142027792, -0.0362966694, -0.0147208525, 0.0233462714, 0.0247625168, 0.0366403423, 0.0992462784, -0.004663635, -0.0372040905, 0.0180677343, -0.0566941388, -0.0283639114, -0.0268770624, 0.0254046582, 0.0128458664, -0.0011582564, 0.013717832, -0.0035330751, 0.054819908, -0.0034889563, 0.0037311094, -0.0501652099, 0.0126020256, -0.0357299596, -0.0172414873, -0.0456072651, 0.053890802, 0.0217535738, 0.0015971912, 0.0335063636, 0.0335936137, -0.0421268493, -0.0268910322, -0.0585408919, 0.055507794, 0.0241894126, 0.022706436, 0.0208504293, 0.001076901, 0.0292960275, 0.0115650138, 0.0272733346, 0.0568187721, 0.0388440229, -0.0316908956, 0.050269261, 0.0003775753, 0.0277200956, -0.0228210706, 0.0216466468, 0.0752629787, -0.0111931954, 0.0291114543, -0.0163194593, -0.0037263103, 0.0561592467, -0.050775826, 0.0296230782, 0.0412995927, -0.0237743221, 0.0522725433, -0.0410134085, 0.0692037195, -0.0213895384, 0.0219702646, -0.0398176946, -0.0176735986, -0.028540954, 0.026566755, -0.0410140045, -0.0187885221, -0.0306680072, 0.007455098, -0.0780941918, 0.0158079527, 0.0491490886, 0.0338518284, 0.0220562462, -0.0289967973, -0.0420142785, -0.0017108367, -0.0297638923, 0.0381453894, -0.0192828998, 0.0093147662, -0.0509446636, -0.0473883748, -0.0085125146, 0.0352511071, -0.0316772833, -0.0256872959, 0.0561410673, 0.0646630824, -0.0345016755, -0.0219769999, -0.0377274416, 0.0105642229, -0.0042253798, -0.098799862, -0.0159317628, -0.0664705113, 0.0302866995, -0.0080171786, 0.0214844067, -0.0668220297, 0.0169040933, 0.0419018343, -0.0374510027, 0.034466017, -0.0127186077, -0.000373871, 0.0036819119, 0.0157587361, 0.011962614, 0.0632486492, 0.0293999631, 0.0498746596, 0.0444689505, 0.0161624178, 0.0022440059, 0.076937072, 0.0055553303, 0.0278223269, -0.041164428, 0.0515923016, 0.0163374878, 0.0355461501, 0.0616236776, 0.0923552364, -0.0008634555, 0.0043835081, 0.0594211482, 0.0374061055, 0.0323192738, 0.0373601131, 0.0644102097, -0.0302966684, 0.0158373155, -0.022745356, -0.0473762453, -0.024182884, 0.0126547841, 0.0137300426, -0.019660525, 0.0267163087, -0.0340261646, -0.0236852467, -0.0147246551, -0.0439575166, 0.0177815016, -0.0237777419, 0.0682027042, 0.0075591635, -0.0466072038, 0.0638906583, 0.0498065241, -0.0671940669, 0.0115357423, -0.0250736605, 0.0656936541, 0.0482244752, 0.0218375381, -0.0418023951, 0.0006347821, -0.0095827496, 0.0521028936, 0.0465238094, 0.0577177256, -0.0285969023, 0.0035156368, -0.091942355, 0.0230255481, 0.0049250461, 0.0079377871, 0.0374650955, -0.0053053005, 0.040873453, 0.0267325994, -0.0044094352, -0.0374339297, 0.0430266336, -0.0541762337, -0.0113377459, -0.0231464785, -0.0161094312, -0.0039242026, 0.0243506655, 0.0113284355, -0.0218643341, 0.0622050948, -0.051074788, 0.0266811065, -0.0112417806, -0.0054548527, -0.0127959484, 0.0739343986, -0.0303272158, -0.0344103761, 0.0214694552, -0.033196222, -0.0338893607, 0.0354783572, -0.0333016291, -0.0069128424, -0.0363784842, 0.036454346, 0.0079867253, 0.0104691042, -0.0400196612, 0.0578214042, -0.0246812645, 0.0123068318, 0.0611744896, -0.028010888, 0.0190102272, -0.0387357064, -0.0453346893, 0.0177159794, 0.0039702039, 0.0004467223, 0.0007532858, 0.0525952354, -0.0178502928, -0.0259642228, -0.101761274, 0.054092709, -0.0530484356, 0.0354814306, 0.004433346, -0.0045807785, -0.0283379238, -0.034819141, -0.0652526617, -0.0205269977, -0.012230779, -0.0118726091, 0.0278243832, 0.0201579928, -0.0194871686, -0.0482696146, 0.0080222553, 0.0094137248, -0.0140262209, -0.0208027195, -0.0152537739, -0.061032705, 0.0496047214, -0.0511986502, 0.0322890803, -0.0593655221, -0.0505407825, -0.0132180713, 0.0090541681, 0.0728742778, 0.0296775289, -0.0173338186, 0.0151688233, 0.047142487, 0.0053079328, -0.0481477901, 0.0154257957, -0.0190972127, -0.0230630357, -0.0986391231, -0.0482205041, -0.0218207687, 0.0524485856, -0.0272921994, 0.0222874135, -0.0272876117, -0.0247234292, -0.0281555988, -0.0117527805, -0.0079724835, 0.0051555787, -0.0251860712, 0.0395389311, 0.01637212, -0.0326453857, 0.0273445472, -0.0537683517, -0.0187946074, -0.0237885918, -0.0431109853, -0.0784107074, 0.0183876697, -0.004491115, 0.0044249799, -0.0501698405, -0.015300421, -0.0417838618, -0.0194612481, 0.0042156829, -0.0399296433, 0.0920988396, -0.0550246499, -0.0041215029, -0.0276414528, -0.0274683014, -0.0345017873, -0.0164391082, -0.0342069902, 0.0121758142, -0.0018014901, -0.0230135769, -0.0251059234, -0.0028791171, 0.0138067473, 0.052736681, -0.0149843078, 0.0258041862, -0.0233057775, 0.0028462114, 0.037356779, 0.0176614709, 0.0233097523, -0.0407510027, -0.0065575908, 0.0453748144, 0.0529292934, -0.0444832034, -0.0372668803, -0.0153566366, 0.041127529, -0.0513085201, -0.0214414503, -0.0232710447, 0.0237954985, -0.0339709669, -0.019073125, 0.0941758528, 0.015501976, -0.0605883934, -0.0047698864, -0.018037444, 0.021886982, 0.0298141185, 0.037182875, -0.0845518783, -0.0126283839, -0.0099018933, 0.0159468241, 0.0117217842, 0.0077133179, -0.0162950791, -0.0969794691, 0.0103383372, 0.0049436674, -0.0027272988, -0.0229324866, -0.0224641487, -0.0088241724, -0.0514351353, 0.0272123665, 0.0141670648, -0.0593421869, -0.0332765393, 0.0262538996, -0.0247880407, -0.0513371155, -0.0103077544, -0.0290747881, 0.0718296245, 0.0428026021, 0.0261063445, -0.0372015834, -0.0293778703, 0.0726491585, -0.0121067995, 0.0092049604, -0.0338808373, 0.0456043929, 0.0135987131, -0.0156307798, 0.0093947593, 0.0097739641, -0.019266719, -0.0038358842, 0.0489543341, 0.0537768491, -0.0121651078, -0.0151361423, 0.017546365, -0.0020225157, 0.0328585692, 0.0591458566, 0.0128566185, -0.0036813822, -0.0121816201, -0.0198823344, -0.0432228781, -0.0282316897, -0.0767048597, 0.0063956408, -0.0400535055, -0.0251504947, -0.0636107549, -0.0547103211, 0.032524921, 0.0229985528, 0.0308328532, -0.0046713469, 0.0032077772, -0.0015904496, -0.0356782973, 0.0444243997, -0.0407572612, 0.0226299912, -0.0782355219, -0.0208851639, 0.0573938526, -0.0163589697, -0.0120352227, 0.0223811492, -0.097077392, -0.0643773973, -0.0078800479, 0.0695101395, 0.0223334078, 0.0701694861, -0.0065404596, -0.059525989, 0.0313225202, -0.0247094873, -0.0122699076, 0.0177999288, 0.0423293337, 0.0077478471, 0.0029768392, 0.044394251, -0.0309808571, 0.0281123854, 0.0169650987, 0.0414724201, -0.0453802273, -0.012981547, -0.0046938793, 0.0152788544, 0.0177848879, -0.0527963713, -0.0127911568, -0.025082672, 0.0457099825, -0.0542344339, 0.008422832, -0.0090407664, 0.0232582595, 0.0332068875, -0.0073989145, -0.0312427785, -0.0214278009, -0.0216949489, -0.0059540817, -0.0453750938, 0.0312211048, 0.0016004094, 0.0345387235, 0.0350734256, 0.0238020103, -0.0148109524, 0.0670856386, -0.0014648669, 0.0550147183, -0.0188265722, 0.0312803052, 0.0083206911, -0.0161408279, 0.0155969355, 0.0221820604, -0.0158562753, -0.0116556687, 0.017732881, 0.0108781699, 0.0890425965, 0.003137693, 0.1189477593, -0.0370397717, -0.0014552635, -0.0166583732, 0.0291445944, -0.0085497489, 0.0023950813, 0.0139597282, 0.007163649, -0.0258993562, 0.0334221497, 0.0499601215, -0.0006396774, 0.0207736921, 0.0222159121, 0.0065326551, -0.0380081311, -0.0449665897, -0.0018111625, -0.0500679798, -0.0394659825, -0.0352520794, -0.0499104373, -0.0479089171, 0.0355119109, 0.042571269, -0.0367284566, -0.010452956, 0.0378142186, 0.0227187462, 0.0107854363, 0.0325760357, -0.0738123879, 0.0229647104, -0.0041170367, -0.0384398885, -0.0422103368, -5.347071071e-33, -0.0352815427, 0.0090782987, 0.0577030629, -0.0005857535, -0.0372052118, -0.0300999992, -0.0201326925, -0.0047408519, 0.0019719056, 0.01839705, -0.0060323677, -0.0124702798, 0.0315035172, -0.0002062728, -0.0081565715, -0.0173852704, 0.0146076977, 0.009868715, -0.004560682, 0.0259496737, 0.0067248386, 0.0380185992, 0.0926551148, -0.0700233057, 0.0064163269, -0.0327691846, -0.0161850043, 0.0540127568, 0.0345440991, 0.08714284, -0.0236726701, 0.0181330051, 0.0073993225, -0.0180717614, -0.0063981526, -0.0007866613, -0.0561704822, -0.0159439351, 0.0491377078, 0.0486633889, -0.0017069598, -0.0403371677, 0.0162837431, -0.0229612943, -0.0313145779, 0.0171478856, 0.0357358009, -0.041346807, -0.0156186847, -0.0329276808, -0.0566936135, 0.0276785735, -0.0036600688, 0.0049444111, 0.0278927144, 0.0373445004, 0.0077455905, 0.0066942517, -0.0603654273, 0.0230002105, -0.0235780273, 0.0403537787, 0.0635391474, -0.0238663014, -0.0161149912, 0.0375266671, 0.0056356415, -0.0159207936, -0.0038780472, -0.0094035231, 0.0426798984, 0.0282965098, 0.0415363349, 0.0544391908, -0.0103437575, -0.0106618172, -0.0970611945, 0.0054220785, 0.05894088, 0.0362121426, -0.0079542156, -0.0264800303, -0.0374024846, -0.0031916813, -0.0096293138, 0.008173706, 0.0040453663, -0.0288483221, 0.0387738347, 0.004982864, 0.0123000154, 0.0055458806, 0.0097152861, -0.0332055911, -0.041147247, 0.0031425718, -0.0079151103, 0.0167364068, -0.0188317318, 0.049729377, -0.0537035763, -0.0383524485, -0.0168503914, 0.0024282702, 0.0346538089, -0.010485718, 0.0440336876, 0.0241211038, -0.0668812022, 0.0080720214, -0.0169537012, 0.0227517188, -0.0005069114, 0.0391965136, 0.0102641387, -0.0028680356, -0.0063380529, 0.052257143, 0.0132796243, -0.0323269814, -0.0178052671, -0.0417632014, 0.0457763039, 0.0355122648, -0.0374827832, 0.0244854502, -0.0203310549, 0.0597333871, 0.0391355343, -0.0229467526, 0.012138023, 0.0142559083, 2.348e-7, 0.0613150671, 0.051244285, -0.0006271905, 0.0411643945, -0.0164540261, -0.0116172619, -0.0766431615, 0.0215091985, -0.0205584224, 0.0291893315, 0.0345274024, 0.0406844057, 0.0707655102, 0.0147848306, -0.0983678252, -0.0041702571, -0.0529437214, -0.0222360175, -0.0535767488, -0.0068458943, -0.0221083388, 0.0863674432, 0.017315438, -0.018912524, -0.0492791981, -0.0218379423, 0.0667302758, -0.0360925123, -0.0321288183, 0.0578939542, 0.032369744, 0.0336380862, -0.0391906388, -0.022276368, -0.0017250641, 0.0291691646, -0.0005752353, 0.1025223583, 0.0127580091, -0.0260341708, 0.047995165, -0.0313717015, -0.0219884422, 0.0016612842, 0.035450168, 0.0092816083, 0.0177081302, -0.031320747, -0.0100904964, 0.0158044305, 0.0140941124, -0.0190293137, -0.0038263733, 0.0364454389, 0.0481308699, -0.0316049978, -0.0436481908, -0.0842869654, -0.0290871318, -0.0112178857, -0.0342172273, -0.0501433536, -0.0068649985, 0.0134813786, -0.0478735641, -0.0559349917, -0.0143805742, 2.107817032e-34, 0.04804774, -0.019566115, -0.0160674546, 0.0532114245, 0.025248358, 0.000521772, 0.000685146, 0.0075468379, -0.0052411482, -0.0655375123, 0.0057192487 ]
http://arxiv.org/abs/2406.18259v1
http://arxiv.org/pdf/2406.18259v1
Detecting Machine-Generated Texts: Not Just "AI vs Humans" and Explainability is Complicated
2024-06-26T11:11:47
2024-06-26T11:11:47
Jiazhou Ji, Ruizhe Li, Shujun Li, Jie Guo, Weidong Qiu, Zheng Huang, Chiyu Chen, Xiaoyu Jiang, Xinru Lu
cs.CL
cs.CL, cs.AI
As LLMs rapidly advance, increasing concerns arise regarding risks about actual authorship of texts we see online and in real world. The task of distinguishing LLM-authored texts is complicated by the nuanced and overlapping behaviors of both machines and humans. In this paper, we challenge the current practice of considering LLM-generated text detection a binary classification task of differentiating human from AI. Instead, we introduce a novel ternary text classification scheme, adding an "undecided" category for texts that could be attributed to either source, and we show that this new category is crucial to understand how to make the detection result more explainable to lay users. This research shifts the paradigm from merely classifying to explaining machine-generated texts, emphasizing need for detectors to provide clear and understandable explanations to users. Our study involves creating four new datasets comprised of texts from various LLMs and human authors. Based on new datasets, we performed binary classification tests to ascertain the most effective SOTA detection methods and identified SOTA LLMs capable of producing harder-to-detect texts. We constructed a new dataset of texts generated by two top-performing LLMs and human authors, and asked three human annotators to produce ternary labels with explanation notes. This dataset was used to investigate how three top-performing SOTA detectors behave in new ternary classification context. Our results highlight why "undecided" category is much needed from the viewpoint of explainability. Additionally, we conducted an analysis of explainability of the three best-performing detectors and the explanation notes of the human annotators, revealing insights about the complexity of explainable detection of machine-generated texts. Finally, we propose guidelines for developing future detection systems with improved explanatory power.
Detecting Machine-Generated Texts: Not Just "AI vs Humans" and Explainability is Complicated: As LLMs rapidly advance, increasing concerns arise regarding risks about actual authorship of texts we see online and in real world. The task of distinguishing LLM-authored texts is complicated by the nuanced and overlapping behaviors of both machines and humans. In this paper, we challenge the current practice of considering LLM-generated text detection a binary classification task of differentiating human from AI. Instead, we introduce a novel ternary text classification scheme, adding an "undecided" category for texts that could be attributed to either source, and we show that this new category is crucial to understand how to make the detection result more explainable to lay users. This research shifts the paradigm from merely classifying to explaining machine-generated texts, emphasizing need for detectors to provide clear and understandable explanations to users. Our study involves creating four new datasets comprised of texts from various LLMs and human authors. Based on new datasets, we performed binary classification tests to ascertain the most effective SOTA detection methods and identified SOTA LLMs capable of producing harder-to-detect texts. We constructed a new dataset of texts generated by two top-performing LLMs and human authors, and asked three human annotators to produce ternary labels with explanation notes. This dataset was used to investigate how three top-performing SOTA detectors behave in new ternary classification context. Our results highlight why "undecided" category is much needed from the viewpoint of explainability. Additionally, we conducted an analysis of explainability of the three best-performing detectors and the explanation notes of the human annotators, revealing insights about the complexity of explainable detection of machine-generated texts. Finally, we propose guidelines for developing future detection systems with improved explanatory power.
[ 0.0640406236, 0.016996596, -0.0218053926, 0.0356285498, -0.0176689923, 0.034294717, 0.0652968287, 0.0346772559, 0.0154058421, -0.0237192474, 0.0187262371, 0.0126809068, -0.0231263544, 0.0393744446, 0.0410520919, 0.0123817949, -0.0072776154, -0.0083458759, 0.0489247963, 0.0310173128, 0.0250811204, 0.0479256064, -0.0027195495, 0.0194764808, -0.0169432685, -0.0305418428, -0.0043859193, 0.00814512, 0.0180846192, -0.0064501627, -0.0116191534, 0.0381470099, 0.0059348368, 0.085103564, 0.0000019018, -0.0368628167, -0.0304349679, -0.0005774388, -0.0692522749, -0.0199632831, 0.0274004806, 0.0099297594, -0.0217512585, 0.024112545, -0.0178562962, 0.0602467172, 0.0331230164, 0.0585476756, 0.0448689908, 0.0690785944, -0.0272532962, -0.0125013608, 0.0339813456, 0.0315898098, 0.0259209741, 0.0109946439, -0.0014582981, -0.0438770391, -0.0207481962, 0.0382041559, 0.0301851239, 0.0139731113, 0.0006500442, -0.0533017665, 0.0061853137, 0.0167017225, -0.0040477966, -0.039086815, 0.0043104906, 0.0247019306, 0.0134981479, -0.0466636121, 0.065140374, -0.0134170102, 0.0034778072, 0.0463829674, 0.0140907792, 0.0555629954, -0.0492548645, -0.0032505293, 0.0225712098, 0.0094500585, 0.0117961708, 0.0460356884, -0.0103079015, 0.0356726311, -0.0356177054, -0.0259650964, -0.0248431135, -0.002630831, 0.019699635, -0.0607474633, 0.0438737758, 0.0397670679, 0.0037960946, -0.0041994499, -0.0365628526, -0.0392314605, 0.0263805725, -0.0454428084, 0.0459742509, -0.0169882774, -0.0411261618, 0.0525396131, -0.0593032688, 0.0505796112, -0.0028036144, 0.0425712876, 0.0012801616, 0.0039662328, -0.0234512649, -0.0341607407, -0.0076433932, 0.0028610667, 0.0191898476, 0.0170958079, -0.0211497229, 0.0408793464, 0.0064663119, 0.0186624881, -0.0209427327, -0.0250914544, -0.0017423742, -0.014400214, -0.0328527093, -0.0237120874, -0.0293955989, 0.0246156454, -0.0118716648, -0.0477087796, 0.0147712296, 0.0852990523, 0.0421509333, -0.0050564506, 0.0137230568, 0.0504558645, -0.0073186397, -0.0266846158, -0.0435132347, -0.0799618885, -0.0124338176, -0.047783453, -0.0294697192, -0.0111095943, -0.0331922621, 0.0191997327, -0.0110355159, -0.0193490703, -0.0136868674, 0.0202031806, 0.0031717513, 0.0536189266, -0.0725865364, -0.003720843, 0.0529114455, -0.0041740341, -0.0440476201, -0.0055812299, 0.0479218513, 0.0963158458, 0.0642043874, -0.0028450277, 0.0746296272, -0.0014334281, 0.005626386, 0.0032243468, 0.0292118602, -0.0188178755, -0.0839037299, 0.0013528115, 0.0255168788, 0.0302844513, -0.0056218076, -0.0096045621, 0.0754293054, 0.0582633391, 0.0083622979, 0.029375365, 0.0417222008, 0.0293312091, 0.0007112178, 0.0468503647, -0.0035121418, 0.0232536569, -0.0428833105, -0.0233439226, 0.001458531, -0.0235988963, 0.0055955644, -0.0644113272, -0.0079175131, -0.0188218039, 0.0191318952, -0.0176417269, 0.0261097904, 0.0001036223, -0.0385621563, -0.0031255081, -0.0341569036, -0.0735302418, -0.0016882181, -0.0894059241, 0.0898436457, 0.0785232633, 0.0308019239, -0.0320213027, 0.0372127816, -0.0464010648, 0.0395805351, 0.0491923504, -0.0282988399, 0.0176278967, -0.0014479393, -0.030119665, -0.0289076008, -0.0156071661, 0.006487722, 0.0038886357, -0.004714089, -0.0223470647, 0.0131204268, -0.0249322932, -0.0472461097, 0.0472210757, -0.0016308038, -0.0391698703, -0.0076702083, 0.0403728746, 0.0077810767, 0.0376987495, 0.0037565429, 0.0055863056, 0.0276785884, -0.0298255123, -0.0451072641, 0.0034846282, 0.0367535129, 0.0265168548, -0.0017024269, -0.0221776143, -0.0353321359, 0.0065560238, -0.0214501247, -0.0253040008, -0.0289125368, -0.0617712215, -0.0180530883, -0.0661634132, 0.0012380519, -0.0094168782, 0.0076113795, -0.0219129212, 0.0712494776, 0.001600804, -0.0009823228, -0.0248651206, -0.0321808681, 0.0600876287, 0.0154965511, 0.0304952618, -0.0098303026, -0.0351709574, 0.0051516523, 0.0126939742, 0.0046107797, -0.0091837626, -0.0326071642, -0.0560304262, -0.018714007, -0.0176248103, 0.0338895954, 0.021078134, -0.0433584452, 0.0070910561, 0.0045628115, -0.0327689983, -0.0141296256, -0.0015121066, 0.0091786245, 0.0888791978, -0.0286895335, -0.0108466893, -0.0087964293, 0.0141068455, -0.0289660431, 0.0488733351, 0.0739738569, -0.0496154651, -0.0791875571, 0.0308299623, -0.0084142182, -0.0574957766, -0.0162956771, -0.0151848905, 0.0198525153, -0.0079014078, -0.0399439894, 0.059712667, 0.0292036384, 0.0574139655, 0.0093552442, 0.0594778545, -0.0005579657, -0.0311158765, -0.0012277954, 0.0128116338, 0.0059939688, -0.0705370307, 0.0346402936, 0.0304278601, -0.0175674912, -0.0019024318, -0.0226103347, 0.0164892003, 0.0171797536, -0.0132798366, -0.0071580051, -0.0671237931, -0.0195719432, 0.0953535438, -0.0096175065, -0.0784749761, 0.0639725327, -0.0369426683, -0.0042047948, -0.0119651593, -0.0319337398, -0.0303167496, 0.0381788835, -0.007363148, 0.0290796682, 0.0299829096, -0.0252799969, -0.013314764, -0.0043957122, 0.0705139562, 0.0124522196, 0.0091175959, -0.0129257357, 0.0022928326, -0.0102048134, -0.0167628955, -0.0142961685, -0.0672895759, -0.0109589659, -0.0144094732, 0.0306853764, -0.0433603413, -0.0142963929, -0.0246223994, 0.0434928462, 0.0411771536, -0.0105045317, -0.0253827609, -0.0394288339, 0.0361399278, -0.0032595079, 0.0050524301, 0.0247097928, -0.0400417261, -0.0040544048, 0.0448373891, -0.0155794732, 0.0110174334, -0.004927949, -0.021401599, -0.0178634562, -0.0389398523, -0.0223231725, 0.0024487828, 0.0315413885, -0.0462484658, 0.0393770337, 0.0953055546, 0.0242658556, -0.015577253, -0.0092113307, -0.0206082109, -0.0323492363, -0.0107027991, -0.0025674871, -0.0897487029, -0.0158029459, -0.0057469313, 0.0229835752, -0.0502010323, -0.0079567693, -0.0158191845, -0.0184303857, 0.0328679755, -0.0112809185, -0.0326427519, 0.0064634588, -0.0622971505, -0.0461590737, 0.0086684749, -0.0554430559, -0.0124471169, 0.0000430975, 0.0152693493, 0.0115519194, -0.0305319168, -0.0092175556, -0.0128740398, -0.0264370367, -0.0024109194, 0.0447525568, 0.0732372031, -0.0125126913, -0.0277954061, -0.0001537716, 0.0025653993, 0.0210904554, -0.0120093599, 0.1078340635, -0.0122981677, 0.0150460918, -0.0617817566, 0.0384096801, -0.0129319588, 0.0044663427, 0.0041017937, 0.0476279147, -0.022300927, -0.0186901707, 0.0040380904, -0.0100937495, 0.0083356034, 0.0530721359, 0.0312889814, -0.02738446, -0.0652084425, -0.0414066873, 0.0398976766, -0.0527164638, -0.0039151562, -0.0409954749, -0.0306015667, -0.051627636, -0.0731678158, -0.035426341, -0.045231577, -0.009236251, 0.0165960211, 0.0284442585, -0.0056156456, -0.0026017593, -0.0387772955, -0.0297581758, 0.0001590744, 0.0386524275, -0.0328182764, 0.0012253789, 0.0143843489, -0.0470036231, 0.0083358912, -0.0024097469, -0.0211020038, -0.0070443233, -0.0317421257, -0.012880628, 0.0259940997, -0.0059307651, -0.0073613646, -0.0839981735, 0.0604890361, 0.0610539541, -0.0241511129, -0.0013009659, 0.0427075773, 0.0181588437, -0.0221472774, 0.0235934816, -0.015461158, 0.0135565046, -0.0101473508, -0.0677522346, -0.0876059756, 0.0007168395, -0.009931122, 0.0149279563, 0.0151609806, -0.0292742513, 0.0077712587, -0.0271268412, 0.0006460647, -0.0049736677, -0.0037709246, -0.0219051167, 0.0151600065, 0.0056119692, -0.0304847937, -0.0090053827, 0.021563042, -0.0240593888, 0.030408686, -0.0030719347, 0.0468457714, 0.0001652476, 0.1038699001, -0.0339703374, 0.0153742824, -0.0025269045, 0.0325627774, 0.0096861692, 0.048581291, 0.0172997098, 0.0593813807, 0.0001262875, -0.0029474662, 0.0239204466, 0.0342067927, -0.0744241849, -0.0126236714, -0.0095408168, -0.0092453761, 0.0366699211, 0.0245668758, 0.0659282953, -0.0163374785, -0.0128724845, -0.0054692305, 0.0610215589, 0.0112467613, -0.0106283491, -0.0602219477, -0.0020554473, -0.0505610332, 0.0014785292, 0.0799996331, 0.0458959639, -0.0107958177, -0.0068826377, -0.0672427416, -0.0339110605, -0.0173174776, -0.0006828839, 0.0250015818, 0.0218689963, -0.0129313739, -0.0402644128, -0.0308939833, 0.0473995097, -0.0210445691, 0.0340087637, -0.0168314818, -0.0199765228, 0.070091188, -0.0057606502, -0.0179389324, -0.0737942979, 0.0256576948, -0.0105256215, -0.0053425939, -0.0297590531, -5.510058138e-33, 0.0143587328, -0.1332232952, -0.0285561755, 0.0033645136, 0.0211748611, 0.0401591621, -0.0467106067, -0.0197221115, -0.0003534724, 0.0262497608, -0.0121158753, 0.0211599395, 0.0092207929, -0.0047823093, -0.0152996732, -0.0286666974, 0.0334245674, -0.002785885, -0.0382763632, 0.0406199768, 0.0613357201, 0.008412201, 0.0105278818, -0.1336231381, 0.0822141618, -0.0270919502, -0.0651212558, -0.029172916, -0.0475225449, 0.0581179857, 0.0024853372, 0.0262087453, 0.0443685725, 0.0148294671, 0.0023055421, 0.0460068136, -0.0435642637, -0.0274158884, -0.0053874827, 0.025482703, -0.0062037511, -0.0871529654, 0.0383322872, -0.0528095812, -0.015313914, 0.023470562, 0.0881151631, -0.0190529935, 0.0000347548, -0.0635944158, -0.0352069102, 0.0242665336, -0.0107500032, -0.0096534444, 0.0469507873, 0.0436564796, 0.0258853193, 0.0107879313, -0.0857202038, 0.0131592918, 0.0134104323, -0.0270734355, 0.0334772281, -0.005581703, -0.0404062197, 0.0429223701, 0.0489174724, 0.0168061368, -0.0453354679, 0.0046841991, -0.0236749612, 0.050278727, 0.107415475, 0.0833490565, 0.034553282, -0.012030012, -0.0463937297, 0.0520804487, -0.0038055668, -0.0333119184, 0.0396388993, -0.0067905686, -0.0706325322, -0.0181906838, 0.0087123988, -0.0678968653, -0.0527313091, -0.0064002997, 0.0013719177, -0.0264949892, -0.0506158546, -0.0153022334, 0.0052515697, -0.0633776411, -0.0188082345, 0.0058333296, -0.0220020805, 0.0411332548, 0.0115597052, -0.0301818736, -0.0063391584, -0.0154650565, 0.0231778938, 0.0178834461, 0.0436266065, -0.0000336647, 0.0707356706, 0.0212779026, -0.041976463, 0.0102707893, -0.0110500241, -0.0175426546, 0.0038221623, 0.0119817005, 0.0092742126, -0.0193633828, -0.0033934154, -0.0031583786, 0.0233180784, 0.0258551128, -0.0268048216, -0.0316417255, -0.0664545894, 0.0044117854, -0.0200234298, -0.0099738641, 0.0031167788, 0.0124637615, 0.0267291702, -0.0631070137, 0.0156171853, -0.0399211235, 2.472e-7, 0.0346620865, 0.0303372648, 0.0401891954, -0.0481389426, 0.0435966402, 0.0634961724, 0.0573908128, 0.0006120087, -0.034920305, 0.0050699939, 0.0256358776, 0.0073444471, 0.0737670511, 0.0125981849, -0.0212574489, -0.014574497, -0.0483648665, 0.0062740822, -0.0741443112, 0.0467525683, 0.1197619215, 0.0040701111, 0.0331401005, -0.0380031653, 0.0091978004, -0.0699991882, 0.0200476553, -0.0116197625, 0.056831032, 0.0123184379, 0.0552088208, -0.0112781562, 0.0099658649, 0.0257500336, -0.0277944226, -0.0330835022, -0.0139423143, 0.0663847476, 0.0147109376, 0.0081693372, 0.0111441743, -0.0313477442, 0.0344434604, -0.0413747989, 0.057645984, 0.0564699546, -0.0302299447, 0.020809792, -0.0170499031, -0.0807809234, 0.0139838979, -0.0183297824, -0.0074510719, 0.0255645495, -0.0161370225, -0.0161448792, -0.0065006162, -0.0573457293, -0.0210841056, -0.03154625, -0.0215602387, 0.0006281202, -0.0573193319, 0.0171872396, 0.0732513443, -0.0646611154, 0.0240367744, 2.363365831e-34, 0.0059466399, -0.0176515896, -0.0094666686, 0.0136357127, 0.0503650829, 0.0543818958, -0.0045336708, 0.0238228794, 0.0405566469, 0.0006763691, 0.0357709341 ]
http://arxiv.org/abs/2406.18256v1
http://arxiv.org/pdf/2406.18256v1
LLaMIPa: An Incremental Discourse Parser
2024-06-26T11:08:17
2024-06-26T11:08:17
Kate Thompson, Akshay Chaturvedi, Julie Hunter, Nicholas Asher
cs.CL
cs.CL
This paper provides the first discourse parsing experiments with a large language model (LLM) finetuned on corpora annotated in the style of SDRT (Asher, 1993; Asher and Lascarides, 2003). The result is a discourse parser, LLaMIPa (LLaMA Incremental Parser), which is able to more fully exploit discourse context, leading to substantial performance gains over approaches that use encoder-only models to provide local, context-sensitive representations of discourse units. Furthermore, it is able to process discourse data incrementally, which is essential for the eventual use of discourse information in downstream tasks.
LLaMIPa: An Incremental Discourse Parser: This paper provides the first discourse parsing experiments with a large language model (LLM) finetuned on corpora annotated in the style of SDRT (Asher, 1993; Asher and Lascarides, 2003). The result is a discourse parser, LLaMIPa (LLaMA Incremental Parser), which is able to more fully exploit discourse context, leading to substantial performance gains over approaches that use encoder-only models to provide local, context-sensitive representations of discourse units. Furthermore, it is able to process discourse data incrementally, which is essential for the eventual use of discourse information in downstream tasks.
[ 0.0079549598, 0.0531854294, -0.0058052894, 0.0312751867, -0.0148743503, 0.012582927, -0.0011739391, -0.0425985977, -0.048385825, -0.0710578561, 0.0026928722, 0.0089596743, 0.0017258873, 0.0371309891, -0.0053191907, -0.0313571915, 0.0335577764, -0.0436487496, 0.0532787219, -0.0132676922, 0.0045493785, -0.0097870072, -0.0102357995, 0.0253068022, -0.000137898, -0.0485359393, -0.0171019901, -0.0256143566, 0.0079791872, -0.0407926142, 0.0428437516, 0.0066420049, 0.0389793739, 0.0106547233, 0.0000021325, -0.0275331382, -0.0139750279, -0.0167811289, -0.0386389941, 0.0251728278, 0.0355904661, -0.0198513512, -0.0187545773, -0.0000137581, -0.0343746059, -0.0046716817, -0.0135032982, 0.0507621765, 0.0611226782, 0.0475203581, -0.0220170282, -0.0739397407, 0.0376029387, 0.0012677618, -0.1042219624, -0.041715268, 0.0106928311, 0.0580374897, -0.000056436, 0.0708277896, -0.0032723465, 0.0528530292, 0.0173054002, -0.0062758769, 0.0197920129, 0.032146126, -0.0002447542, 0.0187381227, -0.0229407307, 0.0277835522, 0.0457630269, 0.0021927326, 0.0002387713, 0.0189666897, -0.0224462431, -0.0260472316, -0.0438733436, 0.0230619907, -0.0409770198, 0.0652645528, 0.0528737083, -0.0077282488, -0.0184061024, 0.0010589457, 0.0011201641, 0.0603426956, -0.0181623772, -0.0269886646, 0.0040590931, 0.0164662469, -0.0355610661, -0.0480344594, 0.0538307652, 0.0515789799, -0.0265384391, 0.002883363, -0.0111501664, -0.0075321845, 0.0352858156, -0.1141994968, 0.0378168635, 0.0339612067, -0.027014697, 0.0308592115, 0.004013923, 0.0664573386, -0.0489085615, -0.0513582565, -0.0229634494, -0.0205498356, -0.0435071699, -0.0429405086, -0.0376752988, 0.0618661717, -0.1366886199, -0.0301217996, -0.0269505884, 0.0570407547, 0.0299550388, 0.0158274584, 0.0201884061, -0.0160837937, -0.0259450376, 0.0498742685, -0.0343205482, 0.0481918342, -0.0429953784, 0.0311084986, -0.0208669752, -0.0382404439, -0.0289235413, 0.0135829588, 0.0036927764, -0.0206198692, -0.0320422836, 0.0395548791, -0.0251680035, -0.0208793227, -0.0111435205, -0.0240629911, 0.0030219632, -0.0238350928, 0.015550144, -0.0635829717, 0.0017872341, 0.0039481083, 0.0257620197, -0.0367383212, 0.0106848227, 0.0277489182, 0.0023606527, 0.0373877399, -0.0076812645, 0.0020145313, 0.073364079, -0.0221128184, 0.0435680859, 0.0865108296, 0.0437131748, -0.0123478118, -0.0023613081, 0.0213357657, -0.039254453, 0.0381166562, -0.0290712882, -0.0154263405, -0.0193891712, 0.0358395875, -0.0246281084, 0.0156284887, 0.0588003919, 0.0606306195, -0.0422441922, -0.011113557, 0.0083610145, 0.0527081415, -0.0636065751, -0.0172495767, 0.0370852388, 0.0282204971, 0.0008207776, 0.0233123172, -0.0007484197, -0.0018710663, 0.0299653076, -0.0204225555, -0.0020268322, -0.0038129436, 0.0031564119, -0.0339833163, 0.0119842775, -0.009489377, 0.0122544589, -0.0348717384, 0.0328653231, 0.0141491713, -0.0315555334, 0.0399423763, 0.0523582473, -0.0168275125, -0.0412533507, -0.0406262986, 0.008250976, 0.0306781325, 0.0413740203, -0.0329168253, -0.0365386941, -0.0372019522, 0.0539059639, 0.0656763464, -0.0236308984, -0.026301289, -0.0177924428, -0.0282305088, -0.0129631115, -0.0434460863, 0.0271697026, 0.0335912667, -0.0277316812, -0.0163357873, -0.0104915155, 0.0439110994, -0.0050591705, 0.0195132643, -0.0359567255, -0.0081536295, -0.0156759843, -0.0234432947, 0.0496420898, 0.0274665467, -0.0094935643, -0.0096530644, 0.0398346446, -0.0651518181, 0.0145235825, 0.037258029, -0.0575876348, -0.0520494469, 0.0350904576, -0.0152525557, -0.0296321362, -0.0317082405, -0.0218719821, -0.0329802297, 0.0099780569, -0.0575919412, 0.021644678, -0.0279655941, -0.0011225905, -0.005734148, 0.0019432053, -0.0214509536, 0.0660094693, -0.0214022696, -0.008770409, 0.0440206788, -0.0194841772, -0.0018290386, -0.0436026268, -0.009423269, 0.0262331814, 0.0341425315, -0.0394416265, 0.0173459966, -0.0129479375, -0.0020029095, -0.0337901041, -0.1279220879, -0.0061275065, -0.0280804578, 0.0577967949, 0.0276806112, -0.0055426629, -0.0016327965, -0.008259506, 0.0241751038, -0.057098832, -0.0062311599, 0.0257779397, 0.044481609, 0.0042092055, 0.003443535, -0.0449355058, -0.0040284409, -0.0199324042, 0.045237653, 0.0085996287, -0.0567456335, -0.0660937801, -0.0085232742, -0.0062772292, 0.0068753455, -0.0173662659, -0.0416823924, 0.0354857035, 0.0327182338, 0.0852515325, 0.0383430794, -0.0052559283, 0.0878039971, 0.0056071626, -0.0563538782, -0.0408628359, 0.0285901967, 0.0128113655, 0.0223103557, -0.0837629586, 0.0058416086, -0.0037343539, 0.0978593454, -0.0058754073, 0.0437862128, -0.0039801011, 0.0276884045, -0.0497807488, -0.0194533411, -0.0006276687, -0.0552156307, -0.0333613902, 0.011461067, -0.0164298154, 0.0261931904, 0.0323251896, -0.050691627, -0.0112142721, -0.0430687368, -0.0004057724, -0.0281907804, 0.0541581213, 0.0048791468, -0.0122363679, -0.0192151666, -0.0651143566, 0.0296768826, 0.0098498752, -0.0050794897, -0.0278335381, 0.0689181238, -0.0769194886, 0.0235202312, -0.0091392957, -0.0422925465, -0.0368394479, -0.0645492598, -0.0299104173, 0.0478624627, -0.0319091529, 0.0318277478, -0.0232756585, 0.01382815, -0.0268710181, 0.049340371, -0.0213591997, 0.01080216, -0.0674963593, 0.0566796213, 0.024725223, 0.0166316312, -0.0084221642, -0.0019637137, 0.0335437171, 0.050771974, -0.0483873673, -0.0197361596, -0.0370338559, 0.0019702024, -0.0263995733, -0.0323357843, -0.0307625998, -0.053678114, 0.0897916704, -0.0583808646, -0.0439273044, 0.1265826672, 0.0208150372, -0.0087284138, -0.0084681371, -0.0453199781, 0.0038628674, 0.0256092437, 0.007391626, -0.0598630235, -0.0117405262, 0.0252484959, 0.0017856789, -0.0171836298, -0.0488543734, 0.0281252488, -0.0541041121, 0.0188249927, -0.0100141549, 0.0379692204, 0.0279585663, -0.0573001355, 0.000498545, 0.0006609763, 0.0454979464, 0.0251044314, -0.0231386982, 0.0169522632, 0.028027771, 0.0143428855, -0.0292624868, -0.0003060774, -0.002581066, 0.045074828, 0.0302785262, 0.049818933, -0.010675245, 0.0164305456, 0.0642799735, -0.0426393338, -0.0029761859, -0.0087266406, -0.0203130525, 0.0132077402, -0.0140137784, -0.0152570577, 0.0201862641, -0.0283255987, -0.0031109715, 0.0300033446, 0.0378090739, -0.0537046939, 0.0451156758, -0.0123293027, -0.0250030383, -0.0168929324, 0.0285194274, 0.0511897318, -0.0001179724, -0.0369498283, -0.0455020145, -0.0195475593, 0.0103965625, 0.0255921055, -0.0158156902, 0.0323874019, 0.0215736292, -0.0559792854, -0.0497587733, 0.008169475, -0.0250149872, -0.0047807083, 0.0042606043, 0.0624861605, -0.0211042427, -0.0530440584, -0.0083269998, -0.0223140549, -0.0067044026, -0.0389671847, -0.0118782884, 0.0852316096, 0.0149659477, -0.0453677848, 0.000875407, -0.0462219641, -0.0467073545, 0.0067476379, 0.0387093239, 0.0160806663, 0.0325857736, -0.0304713268, -0.0159753729, 0.0373928212, 0.0240182094, -0.0383979417, 0.0611029565, 0.0052852766, 0.0444234349, 0.0046578404, 0.0176872052, 0.0214629602, 0.0429407097, -0.0526410192, 0.0437306166, -0.0395168923, -0.0181607753, 0.0199430641, -0.0078536319, 0.0569397248, 0.0248435922, 0.0160025395, -0.0187830813, 0.0314671211, -0.0554966368, 0.0029586293, -0.032641463, 0.0274668224, -0.0127905263, -0.0221030451, -0.0227259025, 0.0013650607, -0.0175746121, 0.0474953018, -0.0456535183, 0.0441698506, 0.0150233442, -0.0038385505, -0.0144875171, 0.0053525376, -0.0238243099, 0.0243744757, 0.0048531215, 0.0426641516, -0.0369209982, 0.0398047604, 0.0034689973, 0.0241936855, -0.0089688376, 0.0276278201, -0.0308615472, 0.0346819125, 0.0188056957, 0.0199795067, 0.0489695705, 0.0039755208, 0.0913854092, 0.0096903071, 0.0181552302, -0.0216879044, 0.056042064, 0.0011192139, -0.0411477201, 0.0308138784, 0.0122400653, -0.0317955166, -0.0025206574, 0.0585011989, -0.0388555974, 0.0424525626, 0.0336988717, 0.0156450048, -0.0490936451, -0.0889131203, -0.0613954812, 0.0094006499, 0.0132883834, 0.0262730438, 0.017962411, 0.001606568, 0.042195037, -0.016384542, -0.0250805374, -0.0486070924, 0.0348596908, 0.0265503302, -0.0086193699, -0.0307590552, -0.0344497934, 0.0198910162, -0.0148684978, -0.0232539847, -0.0248416848, -6.428089471e-33, -0.0921708047, -0.0138123631, 0.0434863381, 0.0299896207, -0.049802877, 0.0177557785, -0.0774706006, -0.0040855384, 0.0353589319, 0.0353767909, -0.0187993273, 0.018913703, 0.0221617334, 0.0181245189, 0.0300010312, -0.0248500519, 0.0769639611, -0.0020877097, 0.0138371782, 0.0109818494, 0.093806982, 0.0366147421, 0.0577723607, -0.0354431607, -0.0109551763, -0.008331052, 0.0077118999, 0.0297867116, 0.0261375066, 0.0548487268, -0.0257542953, 0.0059971632, 0.0347653627, 0.0071330643, -0.0160817243, 0.0262053888, -0.1007430553, -0.0054675862, 0.0286747683, 0.030102035, -0.0194167215, -0.0588634834, 0.0370186195, 0.0206184518, -0.0128878988, 0.0302069653, 0.0283336975, -0.032818079, -0.0134694828, -0.0158155933, -0.0610534474, 0.0239104256, -0.0211219788, 0.0037464211, 0.0367203951, 0.0090658031, 0.0165862087, -0.0487041399, -0.0302502718, -0.015584765, -0.0046135504, 0.0344263949, -0.0124376267, -0.0246135872, 0.0280207805, 0.0618986189, -0.030302627, -0.0507040434, -0.0599128269, 0.0279845744, -0.0002121719, -0.0022870225, -0.0402423218, -0.0436774641, -0.029493738, -0.0009441305, -0.0525136441, -0.0295040961, 0.0498504937, 0.077883929, 0.0333442837, 0.009977703, 0.0105192401, -0.0156557672, -0.0267869905, -0.0227799602, 0.0286533423, -0.0161943175, 0.0116875423, -0.0134282997, 0.0129247876, 0.0118283909, 0.0459622294, -0.0218608603, -0.0418442115, 0.0154163875, 0.0079084579, 0.0201689117, -0.0052497359, 0.0371611528, -0.0266714655, -0.0219906308, -0.014852413, -0.0093839448, 0.0208527893, -0.0053050951, -0.0076761674, 0.0027477846, -0.0397990681, 0.0167504568, -0.0443950109, 0.0357219763, -0.0497826077, 0.105786927, 0.0332643241, 0.0225301757, -0.0096014794, 0.0837438554, -0.0155381151, 0.000149213, -0.0521342568, 0.0114111975, 0.003710255, 0.0474428572, -0.0438148119, -0.0039147716, -0.0165031999, 0.0706886575, -0.019937668, -0.0457888097, 0.0106707197, 0.0226143505, 2.645e-7, 0.0198577717, 0.0715401173, 0.0492739789, -0.040081434, 0.0669929683, -0.0294606239, -0.0027452267, 0.0414138101, -0.0009876946, -0.0347427912, -0.036866121, 0.036406491, 0.0153609822, 0.0412677936, -0.1049301699, -0.0361446664, -0.0870374292, -0.0376564749, -0.0466326997, 0.0027492377, 0.0426932126, 0.0624516793, -0.0301596317, 0.0269076042, -0.0246683508, 0.0146164084, 0.0693687499, 0.0225211531, 0.002283327, 0.0404052548, -0.0231886078, 0.0006003177, -0.0097700041, -0.0148031805, 0.0347440466, 0.0466172993, -0.0179743879, 0.0960770771, 0.0044687479, -0.0028362607, -0.0014470583, -0.0016131833, -0.0004858143, 0.0114593469, 0.0354069509, -0.0532946661, -0.0429270267, -0.0101912627, -0.0420507416, -0.0036771849, -0.0080667064, 0.0209334902, 0.0100511285, 0.0310155712, 0.0069981115, -0.0322944969, -0.0222091693, -0.0714653954, 0.0054064635, -0.0089022443, -0.0728616863, -0.0143426107, -0.0149774347, -0.0045747682, 0.0028571195, -0.028533984, 0.000707293, 2.01205465e-34, 0.049724821, 0.0069714016, -0.003175949, 0.0768955871, 0.015011291, 0.0021401145, 0.0140767861, 0.0382206254, -0.0141224731, -0.1003911197, 0.0124165155 ]
http://arxiv.org/abs/2406.18245v2
http://arxiv.org/pdf/2406.18245v2
Weak Reward Model Transforms Generative Models into Robust Causal Event Extraction Systems
2024-06-26T10:48:14
2024-06-27T10:19:55
Italo Luis da Silva, Hanqi Yan, Lin Gui, Yulan He
cs.CL
cs.CL
The inherent ambiguity of cause and effect boundaries poses a challenge in evaluating causal event extraction tasks. Traditional metrics like Exact Match and BertScore poorly reflect model performance, so we trained evaluation models to approximate human evaluation, achieving high agreement. We used them to perform Reinforcement Learning with extraction models to align them with human preference, prioritising semantic understanding. We successfully explored our approach through multiple datasets, including transferring an evaluator trained on one dataset to another as a way to decrease the reliance on human-annotated data. In that vein, we also propose a weak-to-strong supervision method that uses a fraction of the annotated data to train an evaluation model while still achieving high performance in training an RL model. Our code is available at https://github.com/oyarsa/event_extraction/tree/causal-event-extraction.
Weak Reward Model Transforms Generative Models into Robust Causal Event Extraction Systems: The inherent ambiguity of cause and effect boundaries poses a challenge in evaluating causal event extraction tasks. Traditional metrics like Exact Match and BertScore poorly reflect model performance, so we trained evaluation models to approximate human evaluation, achieving high agreement. We used them to perform Reinforcement Learning with extraction models to align them with human preference, prioritising semantic understanding. We successfully explored our approach through multiple datasets, including transferring an evaluator trained on one dataset to another as a way to decrease the reliance on human-annotated data. In that vein, we also propose a weak-to-strong supervision method that uses a fraction of the annotated data to train an evaluation model while still achieving high performance in training an RL model. Our code is available at https://github.com/oyarsa/event_extraction/tree/causal-event-extraction.
[ 0.0009898614, 0.0835860595, 0.0126949959, 0.0283023678, -0.0146010164, 0.0447953604, -0.0206507333, -0.0110982694, -0.0093576014, -0.0004821142, 0.0096464157, 0.0382750221, -0.0625306219, 0.0178534649, -0.0115772728, -0.0833471045, 0.0490502529, -0.0083851302, 0.0465780906, 0.0340333357, -0.0010662002, -0.0220608767, -0.0107587948, 0.0167630464, -0.0223541744, 0.0078875748, -0.0004337361, -0.0088803638, 0.0518591814, -0.038517978, -0.0083701303, 0.0178204998, 0.0176691897, 0.0411739722, 0.0000019821, -0.0159486327, -0.034523312, 0.0179952942, 0.0142102735, -0.0418714806, -0.0269253962, -0.0077453097, 0.0057941661, -0.0247711036, -0.0209829416, -0.0425975285, 0.0277183615, 0.0804798752, 0.0262935758, 0.0900327191, -0.0163714532, -0.0430404171, 0.0009350023, -0.0418819003, 0.047888469, -0.0590116754, 0.0318630189, 0.0019275108, 0.0139428256, -0.0556133986, -0.0377638638, -0.0048211934, 0.0369406156, -0.0135425515, 0.0234199185, 0.0324931182, 0.002035764, -0.0206252001, 0.0021290742, 0.069245249, 0.0216689222, 0.009417681, 0.0567146577, -0.0210571159, 0.024646135, 0.0207842942, -0.0370136835, 0.0114761684, -0.0173229575, 0.0095605515, 0.0141635099, 0.0370778479, 0.0229303818, 0.0188331548, -0.0002943828, 0.047340326, 0.008429979, -0.0406116433, -0.0247954801, 0.0311656836, -0.0063473703, -0.0266769435, -0.0081965122, 0.0153927635, -0.0549629182, 0.0160797443, -0.0167729687, -0.0003848611, -0.0196285937, -0.0409045182, -0.0048517431, -0.0048581688, -0.039181456, 0.0771691129, -0.019386109, 0.0435352102, 0.0014390787, 0.0806254894, 0.0120432377, 0.0049730036, -0.051740434, 0.0319294631, 0.0034096895, 0.0052241087, -0.0205329843, 0.0031878566, -0.0482111797, 0.0048440844, -0.0509403162, 0.00340648, -0.0059299073, 0.0317299403, -0.0506573878, 0.037628103, -0.0318879411, 0.0032164203, -0.0865944028, 0.0357746668, 0.0281250756, -0.0205683708, -0.0200152602, 0.0127552068, 0.0381639078, -0.0358737893, 0.0007880304, 0.075455457, 0.0099384151, -0.0264603402, -0.0582683347, -0.0371411704, 0.000957294, -0.0283696726, -0.0103475275, -0.0361272059, 0.0111911567, -0.02748104, 0.0040956279, -0.0366808288, 0.0061667226, -0.0479633175, -0.0362368263, 0.0509201623, 0.0097391987, -0.011141167, -0.0183545966, -0.0076491246, -0.0266719367, 0.0338330641, 0.0390234478, 0.0710801706, 0.0217497814, 0.032227017, 0.0089366483, -0.0069137248, 0.0289714169, -0.0083567789, -0.0059037274, -0.0295773223, -0.020912515, 0.003749046, 0.0584431328, -0.0008386369, -0.0433796905, 0.029577842, 0.0372585505, -0.0162987877, 0.0003488645, -0.0286895577, 0.0640339628, -0.0319293179, -0.019699771, 0.0258798189, -0.043054834, 0.0251345988, 0.0099676894, -0.0006562117, -0.0129517429, -0.0382037982, -0.0057640108, -0.0474984832, 0.0315866992, -0.005442915, -0.017203493, 0.0282880682, 0.0935371667, 0.0164435543, -0.0220351573, -0.0212606434, 0.0357097611, -0.0251139626, 0.0255271625, -0.035785079, 0.0814236552, 0.0561237074, 0.0441939905, -0.0467112921, -0.0055259173, -0.0117630726, -0.0139825772, -0.0157045443, 0.0033418038, -0.0571821257, 0.0154095944, 0.0039479178, -0.0704782382, 0.0033701095, 0.0254895352, 0.0299736969, 0.0340760723, -0.0409307927, -0.0130088348, 0.0141241606, -0.0248704646, 0.0452711582, 0.0056862948, -0.0509124137, -0.0273234509, 0.0235859361, 0.0592334606, -0.0044341935, -0.0419990905, -0.0017307154, 0.0331544541, -0.0281832553, -0.0376884602, 0.0524571352, -0.0234584548, -0.0310738813, -0.0098467004, -0.051993113, -0.0472060814, 0.0358251669, -0.06145503, -0.0485567078, 0.0160955563, 0.006289307, 0.0033351232, -0.0336814187, -0.0139146661, 0.0140443156, -0.0531646833, -0.0509074144, 0.0895522311, 0.0154417735, 0.0091095725, -0.0137177082, 0.0139856543, 0.005152097, -0.04926477, -0.0193391703, -0.0017748374, -0.0246015266, -0.0203487277, 0.0728567615, 0.0026471475, -0.0538237281, -0.0182970334, -0.0374164172, 0.0049582208, -0.0203984883, 0.0197749436, 0.0214882866, 0.0163286831, -0.0156871025, -0.0194427278, -0.0050230455, -0.0587831289, -0.0139590809, 0.0048302785, 0.0206686854, 0.0367243998, 0.019058127, -0.0623294301, 0.0169398133, -0.0269832108, 0.0357069559, 0.0703968406, -0.0127708316, -0.0116391378, -0.007759918, 0.0121110454, 0.0172906741, -0.0128418198, -0.0189892463, 0.0126365162, -0.0381026827, 0.0246145856, 0.0848237798, -0.0097593302, 0.0785176978, 0.0050913077, -0.0429535322, -0.0620697327, 0.0138503741, -0.0599276572, 0.0083442917, -0.0144135645, 0.0020203632, 0.0056363223, 0.0737986043, -0.0028274963, 0.0165276751, -0.0582965948, -0.0115609029, -0.0189805925, -0.0531774759, -0.0207102764, 0.0031518622, -0.0360756367, 0.0152192377, -0.0698095709, -0.0436671264, 0.0794360563, 0.0380285382, 0.002885524, -0.0156901293, 0.0203484539, -0.0323677287, 0.0593107492, 0.0170610026, 0.0087508392, -0.0067893369, -0.0272563118, -0.0371023007, -0.0229006894, 0.0480031781, 0.0152752092, -0.0067668017, -0.0588457957, 0.0331610814, 0.0000660511, -0.032444708, -0.0638685748, -0.0432644561, 0.0331734419, 0.0150391115, -0.0052008531, -0.0057598082, -0.0389135033, -0.009280541, 0.031613525, 0.0445756726, -0.0327994227, -0.0063783764, -0.0214090496, 0.0093545485, 0.0684862211, -0.0689129755, -0.0284724813, -0.0428060666, -0.0031993545, 0.0318218395, 0.0603683479, 0.0051062508, -0.0432644077, 0.0397667922, -0.0231973697, -0.0400453135, 0.0398965925, 0.0256126001, 0.0315164365, -0.0528177023, -0.0250757355, 0.0805541053, 0.0120722437, -0.0122185703, -0.0016895018, 0.0122337174, 0.0118665779, 0.0248375796, 0.0113857118, -0.0553980395, -0.0233813133, -0.0129637159, 0.0559316315, -0.0013335596, -0.0517501831, 0.060042996, -0.0760755688, 0.0421912633, -0.014780771, -0.0611787662, -0.0182819981, -0.0359921306, 0.0310362335, -0.0529251285, -0.0659006685, 0.0090338327, 0.0162004121, 0.0166024417, 0.0393649228, 0.0123666823, -0.071110934, 0.0319418572, -0.0522965863, 0.0176307186, 0.1036811993, 0.0144664235, -0.0078919511, -0.0092029748, -0.0080987168, 0.0167320054, -0.0254166592, 0.0159116965, 0.0656473041, 0.0035772235, -0.017573446, -0.0095486231, 0.0729049668, 0.0270117819, -0.0264399443, -0.0018336424, 0.0800908208, -0.0204835851, -0.0040094154, -0.0031566704, -0.0366345979, 0.0185678378, 0.0619024113, 0.0278513171, -0.0229760371, -0.0634424165, 0.0261680372, -0.0163676869, -0.0849672705, -0.0208093282, -0.0188989844, -0.0411897413, -0.0823352411, -0.0780683681, -0.0737140924, 0.0134885842, 0.0345784836, -0.0218581744, -0.0289263148, 0.0133907106, 0.0049778465, -0.0264863316, 0.08422295, -0.0266314242, -0.0006022014, -0.0273456816, -0.0370562263, 0.0006739087, 0.0113723762, -0.0371770561, 0.057037808, -0.0324928313, -0.013582306, 0.0126962047, 0.0557616316, 0.0164812841, 0.0849517435, -0.0259450395, -0.0153303286, -0.0049409866, -0.0006691758, -0.0267430022, -0.0248289406, 0.0376959853, -0.0196424052, -0.0184290558, 0.0050367587, -0.0602339879, 0.05723897, 0.0160244815, 0.0393254459, -0.0303945318, 0.0389805622, 0.0411688797, -0.0272501111, 0.0933844298, -0.0369789712, -0.0362194777, -0.063186422, -0.0366723649, -0.0058972607, -0.0186311658, -0.0440037772, 0.0255449042, 0.0138523839, 0.0348727778, -0.0216560531, -0.024748005, -0.008707202, 0.0444121584, -0.0958061963, 0.0245027523, 0.0675186217, 0.0212420914, -0.0277884752, 0.0184261762, 0.0070913807, -0.0061208494, 0.0206175987, 0.0343306623, -0.0114369867, 0.0345786437, 0.052673351, 0.0438455157, 0.0262494814, -0.0333951265, -0.0911419615, 0.0475676097, -0.0153396623, 0.0226723086, 0.070305571, -0.0240229294, 0.0244900826, -0.0127375545, 0.0024801851, 0.0042083338, 0.031512145, 0.0138114439, 0.0101687219, 0.0215206631, 0.0368538871, 0.0021810723, -0.0226599537, 0.0405001491, 0.0031887214, 0.0805201679, -0.0633455664, -0.0268639214, -0.0340896398, -0.0497126468, 0.0433866866, -0.0057229358, -0.0296764374, -0.0214356072, -0.0425722972, -0.0095506599, -0.0142203085, 0.0231899209, 0.0181277338, 0.0109044714, 0.0475085713, 0.0297423303, -0.0351883024, -0.018948406, -0.0455606952, 0.0000833966, 0.0127251437, -0.033196345, -0.0211494267, -6.272263736e-33, -0.0416088179, -0.0131509127, 0.0176612828, 0.0083780643, -0.0049416604, 0.0177339502, -0.0578584149, -0.0376760066, 0.0204199255, 0.0434481502, 0.0225236155, -0.0005836642, 0.0208631046, 0.0041134339, 0.0504299439, -0.0284848344, 0.0509751141, 0.0114769908, -0.0091189733, 0.0390438773, 0.0656125173, 0.0323897824, 0.0165985264, -0.1279200912, 0.0550147928, 0.0145769371, -0.0362313688, -0.0096945632, 0.0163374767, 0.0467668734, 0.0249203555, -0.0295002405, 0.0123165157, -0.0844299942, -0.0098290155, 0.0036990626, -0.1368454546, -0.0077325492, 0.0194522012, -0.0105891582, -0.0599808171, -0.0309473183, 0.0246467814, -0.0014188399, 0.0290772542, 0.0123937558, 0.056809254, -0.0289665982, -0.0037290833, -0.0734231621, -0.0696093887, 0.0418114401, -0.0181661993, -0.0122234114, 0.0269818474, 0.008100668, 0.0137172239, 0.0359391496, 0.0237839669, -0.0081308875, 0.0118717439, 0.0094598504, 0.0333149359, 0.0618736856, -0.0159464255, 0.063447088, -0.0363729335, -0.0327782407, 0.0013232722, 0.001729586, -0.0154084945, 0.0978671238, 0.0230035726, 0.0339222066, 0.0164232701, 0.0521200374, -0.1002888903, 0.0607745983, 0.0098939352, 0.043996729, -0.0060236743, 0.0256731603, -0.0352810435, 0.0209541954, -0.0265335366, -0.0104395701, -0.0164681524, 0.0045829075, 0.0429086871, 0.0376737006, -0.010100944, 0.0322100595, 0.02949946, -0.0262202024, 0.0209910236, 0.016158171, -0.0272296593, 0.0184772722, -0.0197728071, -0.0365229435, -0.0460981689, -0.0283664726, 0.048925247, 0.0035352584, 0.0444505364, 0.0214148164, 0.0604865439, 0.0002466589, -0.0039675212, 0.0079548443, -0.0378603078, -0.0017123071, -0.0267728213, -0.0089019006, 0.0216698032, 0.016728634, -0.0282557309, 0.0367707573, 0.0132755814, 0.061650075, 0.0181032121, -0.0445939526, -0.0105401287, 0.0120425178, -0.0245024879, -0.0182077009, -0.0060758214, 0.0259871595, -0.0248981509, -0.0311904065, 0.017391881, 0.0031340825, 2.681e-7, 0.0197222363, 0.0099701378, 0.0604605004, 0.0153716225, 0.01033244, 0.0187695418, -0.0083129043, 0.024720883, -0.007157024, 0.0291814469, 0.0419766568, 0.0073886733, 0.0389304981, 0.0014490975, -0.0539968759, -0.031783767, -0.0789753571, -0.0264228508, -0.0393871367, -0.0180370584, 0.0938912705, 0.0660145208, 0.0610884055, 0.0049266191, 0.0174628515, -0.020579664, 0.0443204753, -0.012564159, -0.0110546555, 0.0041779494, 0.0240608696, -0.0103863776, -0.0217734519, 0.019867884, -0.0263932776, 0.0034294687, -0.0017921372, 0.0550484098, -0.0355684198, -0.0063933269, -0.0252408776, -0.001151, 0.051941406, -0.0195997357, 0.0487619936, -0.0557117425, -0.0152576612, 0.0409880169, -0.0512926169, 0.0193420518, -0.048370406, 0.0049005677, 0.0115590291, 0.0128174964, -0.006883976, 0.0057499064, -0.0326705463, -0.0654782057, -0.0167269669, 0.0271208491, -0.0093806703, -0.0505767912, -0.0081897173, 0.0291521288, -0.0075916033, -0.0791193098, -0.0379406959, 2.30869042e-34, 0.0141766705, 0.046271462, -0.0573712625, -0.014683513, 0.0433315709, 0.0085108951, -0.047708381, 0.0250074733, 0.0189857967, -0.0130938003, -0.0069823917 ]
http://arxiv.org/abs/2406.18239v1
http://arxiv.org/pdf/2406.18239v1
Zero-shot prompt-based classification: topic labeling in times of foundation models in German Tweets
2024-06-26T10:44:02
2024-06-26T10:44:02
Simon Münker, Kai Kugler, Achim Rettinger
cs.CL
cs.CL, cs.AI
Filtering and annotating textual data are routine tasks in many areas, like social media or news analytics. Automating these tasks allows to scale the analyses wrt. speed and breadth of content covered and decreases the manual effort required. Due to technical advancements in Natural Language Processing, specifically the success of large foundation models, a new tool for automating such annotation processes by using a text-to-text interface given written guidelines without providing training samples has become available. In this work, we assess these advancements in-the-wild by empirically testing them in an annotation task on German Twitter data about social and political European crises. We compare the prompt-based results with our human annotation and preceding classification approaches, including Naive Bayes and a BERT-based fine-tuning/domain adaptation pipeline. Our results show that the prompt-based approach - despite being limited by local computation resources during the model selection - is comparable with the fine-tuned BERT but without any annotated training data. Our findings emphasize the ongoing paradigm shift in the NLP landscape, i.e., the unification of downstream tasks and elimination of the need for pre-labeled training data.
Zero-shot prompt-based classification: topic labeling in times of foundation models in German Tweets: Filtering and annotating textual data are routine tasks in many areas, like social media or news analytics. Automating these tasks allows to scale the analyses wrt. speed and breadth of content covered and decreases the manual effort required. Due to technical advancements in Natural Language Processing, specifically the success of large foundation models, a new tool for automating such annotation processes by using a text-to-text interface given written guidelines without providing training samples has become available. In this work, we assess these advancements in-the-wild by empirically testing them in an annotation task on German Twitter data about social and political European crises. We compare the prompt-based results with our human annotation and preceding classification approaches, including Naive Bayes and a BERT-based fine-tuning/domain adaptation pipeline. Our results show that the prompt-based approach - despite being limited by local computation resources during the model selection - is comparable with the fine-tuned BERT but without any annotated training data. Our findings emphasize the ongoing paradigm shift in the NLP landscape, i.e., the unification of downstream tasks and elimination of the need for pre-labeled training data.
[ 0.0322009735, 0.0003520919, 0.0186987314, -0.0183550138, -0.0318113677, 0.0436610021, 0.0298580918, -0.0144568589, 0.0105845062, -0.0912301391, 0.0241907984, -0.0256491005, -0.0505229086, 0.0220346861, -0.003421417, -0.0616241172, 0.0431958735, 0.0042798407, 0.0425809547, 0.0103337644, 0.0175891519, 0.0149916569, -0.0193646774, 0.0248807818, -0.0112614725, 0.0401547775, -0.0116264755, -0.0136710862, 0.0016155241, -0.0184490271, 0.0147019252, 0.1005994827, 0.0310936458, 0.052363053, 0.0000016759, -0.0702910125, -0.0291432366, -0.0229851175, -0.0054503474, -0.0315599442, 0.0425143875, -0.0373608284, 0.009442932, 0.0165314041, -0.0393432155, -0.0054881372, 0.0448831059, 0.0682411939, 0.0584160015, 0.0868528187, -0.0236037858, -0.0156324077, 0.0266708247, -0.0334821716, 0.0024128614, -0.0490424745, 0.0322558917, -0.0490480848, -0.0027377002, 0.0353467278, 0.0276256166, 0.020787647, -0.0169280935, 0.0033717563, 0.0328170173, 0.0233124271, -0.0345933251, -0.0579291061, -0.0192305464, 0.048550874, 0.030152455, -0.0406228006, 0.0229552034, -0.0165309925, -0.0101503413, 0.0166660696, -0.014199933, 0.0299304593, -0.0265746526, 0.0043738713, 0.061640285, -0.0189877711, 0.0055788197, 0.0485738032, 0.0300930999, 0.0883042514, -0.019562196, -0.0567515269, 0.0489660725, 0.0670659915, -0.0237061698, -0.0814109668, -0.0082084546, 0.0285455566, -0.0172607768, 0.0058259661, -0.0515042506, -0.0438568816, 0.0051570544, -0.0413467623, -0.0143655669, 0.0332596563, -0.0155860009, 0.0708324611, 0.0383090228, -0.0379137062, 0.0103784911, -0.0221728459, 0.0086402055, 0.0036610942, -0.0831336826, 0.0344193392, -0.0590786375, -0.0011769707, -0.0223983806, 0.005539353, -0.0079257693, 0.0244107842, 0.019432256, -0.0131283989, -0.0304559451, -0.0447915085, -0.0561061502, 0.0288407765, -0.008821018, 0.0382373519, -0.0558986366, 0.0048889141, -0.0156458858, -0.0188940112, -0.0004343992, 0.010701918, 0.0220897049, -0.0271322504, -0.0037078969, -0.0091619985, -0.0273629744, -0.0499861762, -0.0603263862, -0.003022155, -0.0017009625, -0.0059891678, 0.0078785894, -0.0655469447, 0.0124137318, 0.0343493856, -0.0199145507, 0.0034161771, 0.0141185056, 0.0185191724, -0.0619050488, 0.0520479307, -0.0858003125, 0.0093762148, 0.0574947558, -0.0233455822, -0.0322277509, 0.0941870511, -0.0099961273, 0.0272401925, 0.0655336305, 0.0085796975, 0.0102422005, 0.0447485484, 0.0028904758, 0.0436596088, -0.0017688441, -0.0027526065, -0.0208841879, -0.0064542596, -0.0102128498, -0.021572927, 0.0007388997, 0.0401763581, 0.053180255, 0.044082962, 0.0081690717, 0.0071365917, 0.0185051709, 0.0373318493, 0.0373132601, -0.014041041, -0.0298722386, -0.0182159804, -0.0503005013, 0.002909973, 0.0100069968, -0.0233155899, -0.0186090525, 0.0017396931, -0.019578496, -0.0158118047, 0.0004425278, -0.0246207416, 0.0629269704, -0.0010194927, -0.0021172245, 0.039158009, 0.041632127, -0.0222987626, 0.0010602921, -0.0243393667, 0.0312917046, 0.027796546, 0.001635546, -0.0601665415, -0.0385555662, -0.0497761965, 0.1003308296, 0.0654455498, 0.0449955501, -0.0088929692, -0.0127009451, 0.0114252875, -0.0337328427, 0.0055924142, 0.0316467583, 0.0055408236, 0.0249463804, -0.0196273178, -0.0154128103, -0.0168580431, -0.0148244649, 0.0362008661, 0.0034551255, -0.0369815938, 0.0135953892, 0.0073371395, -0.0087603899, 0.0448541194, -0.0233728625, -0.0276612025, -0.0024146207, -0.0138798254, -0.0276114661, -0.0230696406, -0.0039576855, 0.012782027, 0.0000804766, 0.0217397138, -0.0522658937, -0.0131823318, -0.0633090213, -0.0225146245, -0.0145374546, -0.0039017785, -0.0123813823, -0.0442447588, -0.0179620199, -0.056856215, -0.0152147245, -0.0326903947, 0.080183506, 0.0227737669, 0.0053359605, -0.0090381308, -0.0259612687, 0.0117104286, -0.022966817, -0.0263921171, 0.0068889442, -0.0017952377, 0.0154812178, 0.0577956624, 0.0291750096, -0.054570917, 0.0135164503, -0.1301018149, 0.0417818055, -0.041839581, 0.0273280963, 0.0412221998, 0.0400189757, -0.0259456653, -0.0173974931, -0.0047308835, -0.0126414886, 0.000611253, 0.0327836387, 0.0537295863, 0.003984544, 0.0216967724, -0.0215459075, -0.001219146, -0.0591169745, 0.0138268154, 0.0646993443, -0.0682272166, -0.0248473436, -0.0028776519, 0.004965832, 0.0240254086, 0.0206136145, 0.0068915542, -0.0059176479, -0.0252426956, 0.0055442923, 0.0242495034, 0.0220578276, 0.0568066351, 0.0100890072, 0.0338933878, -0.0183568913, 0.0127792368, -0.0373009965, -0.0188212208, -0.0141755519, -0.0147371795, 0.039220687, 0.0840538666, -0.0362002887, 0.0134095661, -0.0543859303, 0.0033390324, -0.0433184355, 0.0236628894, -0.0582374409, 0.006518221, -0.0017954879, 0.0300473478, -0.0161309708, -0.0155917564, 0.049321156, 0.0168873016, 0.0300303288, -0.0667669922, -0.0703940988, -0.0296072643, 0.0125600966, 0.0003848628, -0.0043225079, 0.0413892418, -0.0018709614, 0.0033618945, -0.0141199566, 0.0304885488, 0.0135476673, 0.0251656286, -0.0280596539, 0.0362836905, -0.0132953711, -0.0276082661, -0.0098440247, -0.113352932, -0.0545036383, 0.0240542982, -0.030026501, -0.0209379476, -0.0236233361, 0.0041763787, 0.0309790876, 0.0384110622, -0.0167148896, -0.0263313204, -0.0115229916, 0.0421812534, -0.0098509211, -0.01098435, -0.0472388193, 0.0005078801, 0.0306807216, 0.0089899637, 0.0696175992, 0.0222748313, -0.0349407643, -0.0017301093, 0.0101901237, -0.0767191947, -0.0251815338, -0.0074287928, 0.0445477366, -0.0124529088, 0.0170446318, 0.1397732049, 0.0454343371, -0.0452124551, -0.0292048659, 0.0133855045, 0.0062552183, 0.0238869861, 0.0362829827, -0.0849560425, -0.0341198258, 0.0082738847, 0.0639306828, -0.0236022845, -0.0624816529, -0.0016126173, -0.0850004479, 0.0123444078, -0.0309797041, -0.0315236598, 0.0332836136, -0.1048115343, 0.0409808382, -0.0443297885, -0.0511634722, -0.0428398103, -0.0064095482, 0.0390140638, 0.0306419805, -0.033277832, -0.0768407509, 0.0126566002, -0.0394966938, 0.0306434352, 0.0195411425, 0.0131690288, -0.0087871933, 0.032641951, 0.0502797328, 0.0069328425, -0.0031357196, -0.0309600513, 0.0646462142, -0.0183166508, 0.0038787613, 0.0292907692, 0.0060742451, -0.0198400132, -0.0257105771, 0.005538193, 0.0498935282, -0.0096789598, 0.032157965, 0.0125258053, -0.0003069084, 0.0504992194, 0.0694958866, 0.0157256667, -0.0672160015, 0.0040309071, -0.0406883247, 0.0247845128, -0.0352600217, -0.0179564282, 0.0016407127, -0.0818070695, -0.0957017392, -0.0341605991, -0.0328646936, 0.0566210561, 0.0302781034, 0.0081066415, -0.0294997469, 0.0251989719, 0.0096086273, -0.0449571535, 0.0530535281, -0.0228666514, 0.0033968524, -0.0568104908, -0.0329193324, 0.0309648253, 0.0048836316, -0.0190801024, 0.0081436383, -0.0458543301, -0.0414529406, 0.0353010967, -0.0215484276, 0.0243996475, 0.0181842875, -0.0212091189, -0.0548565164, 0.0077330279, 0.0744117275, -0.0237818658, -0.0185662657, 0.0185836405, 0.018427806, -0.0062807724, 0.0162815861, -0.0500115976, 0.0515874736, -0.0034181273, 0.0116601111, -0.0325672701, 0.0063216798, 0.0035329601, 0.0217884406, 0.1149986014, 0.0015583549, -0.0658446103, -0.0566729903, -0.0536573082, 0.0213952959, -0.032848563, -0.0166572426, -0.0004933694, 0.007490966, -0.0053024236, -0.0186767504, 0.0547824539, -0.0387254432, 0.0473911501, -0.0616420992, 0.0100352829, 0.0211233366, 0.056136366, -0.0092034107, -0.0206913203, -0.0231768358, 0.011258048, 0.0548331067, 0.0093150642, -0.0246687997, 0.0498497672, 0.0108461352, -0.0030123298, 0.0241582971, 0.0083487006, -0.0797903091, -0.0120944092, -0.051560536, 0.0628853664, 0.0427652411, -0.0354249179, 0.0338161439, -0.0428128764, -0.0392236635, 0.0028811391, 0.0017791478, -0.001212543, -0.0261566639, 0.0025425812, -0.0048776614, -0.0347361974, 0.0394762121, 0.0306941345, 0.0017038595, 0.0364678539, 0.04503024, -0.0237869676, -0.0251762494, 0.0215060003, 0.0226641335, 0.0009081053, 0.0127682779, 0.0149392784, -0.0450338051, -0.0126131065, 0.0532285795, -0.0562750138, -0.0433734693, 0.0063012051, 0.0419773683, 0.0795101151, 0.0021703301, 0.0372215509, -0.0571628027, 0.0322131589, 0.0200448278, -0.054249052, -0.0374546126, -6.712332087e-33, -0.0040507643, -0.0150103709, -0.0150019825, 0.0012875322, 0.0186124034, 0.0019389328, -0.0029138816, 0.0159691535, 0.0013911946, 0.0176150016, -0.0183674097, 0.0218613241, 0.0194719695, 0.0094684409, 0.0014729649, -0.0320204943, 0.030038571, -0.0318534523, -0.0499721356, 0.0368748941, 0.0381472893, 0.0196157787, 0.0421147011, -0.0984209403, 0.0486255884, 0.001415398, 0.0055537494, 0.0315273702, 0.0986592397, 0.0179801397, 0.0098605966, 0.0706382915, 0.057750918, -0.0657491088, 0.032614395, 0.0824239701, -0.0591949821, -0.0504977554, -0.0306594726, -0.0113837635, -0.0122310491, -0.0104297427, -0.0151320165, 0.0100531895, -0.0044322819, 0.0239018165, 0.0458035991, 0.0168742463, 0.0140998876, -0.0284092035, -0.0560034811, 0.0532483831, -0.0145370057, -0.0231907498, 0.0062796427, 0.0701377019, 0.0241785105, 0.0127255023, -0.0260521192, 0.0245040022, 0.0142758032, -0.009255603, 0.0145696215, -0.0355790369, -0.0429085828, 0.0199238546, -0.0011737957, 0.0617560409, -0.0226568803, -0.0165420696, -0.0217949934, -0.0096422546, 0.0061620665, 0.0418468118, 0.0273950342, -0.0029226625, -0.0603210367, 0.0501562767, 0.0598290563, -0.0075562843, -0.0136987325, -0.0218026396, -0.0151841389, -0.016556887, -0.0239678007, -0.037471734, -0.030907141, 0.0138820745, -0.004227059, -0.009268255, -0.0187350623, 0.0317600556, 0.0312385038, 0.0136207528, 0.0134983091, -0.0160902254, -0.0074459617, 0.0584920831, -0.0006751217, 0.0222794246, -0.068760559, -0.0517391078, 0.0151222777, 0.0367247388, 0.030582279, -0.0108859111, 0.0357762761, 0.0137399854, -0.0429614112, 0.0002845889, -0.0058584958, 0.0442632101, -0.027187502, 0.0396891646, 0.0066858977, 0.0182044562, 0.0218668245, 0.0472607389, 0.0667549372, -0.0001292567, 0.0104554808, -0.0245250482, -0.0206443015, -0.0079052085, 0.0162013788, 0.0017836477, 0.0133704022, 0.0446197838, -0.0002976096, -0.0584000833, -0.0154274348, -0.0079145236, 2.646e-7, 0.0638331994, 0.1003044546, -0.0158969015, -0.024086671, 0.0225438997, 0.0638603568, -0.001456399, 0.0378182456, 0.0019823667, -0.0109454058, 0.0021338426, -0.0139876539, 0.0427363291, 0.0430530757, -0.0774786472, -0.0669244975, -0.0550261922, 0.049281694, -0.0758878812, 0.0390954837, 0.0483166762, 0.0575217418, 0.028104987, -0.0008282633, -0.0360544957, -0.0625526235, 0.0400851108, -0.0136095621, -0.0204141699, 0.0205687899, 0.0433847383, 0.0237037633, -0.0019789853, 0.0414448045, -0.0330263413, -0.0006162477, -0.0109105865, 0.0352566876, -0.0017518032, 0.0093028201, 0.0023116202, -0.0268534981, 0.0275981817, -0.0189568307, 0.0686719343, -0.096163474, -0.0301237367, -0.0247552134, -0.0145286778, 0.0023594568, 0.0126949754, -0.0082224999, -0.0461736582, 0.0143353716, 0.0102356793, 0.0111788763, -0.0326417275, -0.0323772803, -0.0038805744, 0.0231591798, -0.0102580097, -0.0459066145, -0.0137874912, 0.0005148606, 0.022330543, -0.053569112, 0.0585877337, 2.315513568e-34, 0.0182989351, -0.0150481686, -0.018478388, -0.0030222782, 0.0192744862, 0.0272539482, -0.0286124703, -0.0126454048, 0.0045932513, -0.050496567, 0.0212973077 ]
http://arxiv.org/abs/2406.18227v1
http://arxiv.org/pdf/2406.18227v1
GUIDE: A Guideline-Guided Dataset for Instructional Video Comprehension
2024-06-26T10:24:00
2024-06-26T10:24:00
Jiafeng Liang, Shixin Jiang, Zekun Wang, Haojie Pan, Zerui Chen, Zheng Chu, Ming Liu, Ruiji Fu, Zhongyuan Wang, Bing Qin
cs.CV
cs.CV, cs.CL
There are substantial instructional videos on the Internet, which provide us tutorials for completing various tasks. Existing instructional video datasets only focus on specific steps at the video level, lacking experiential guidelines at the task level, which can lead to beginners struggling to learn new tasks due to the lack of relevant experience. Moreover, the specific steps without guidelines are trivial and unsystematic, making it difficult to provide a clear tutorial. To address these problems, we present the GUIDE (Guideline-Guided) dataset, which contains 3.5K videos of 560 instructional tasks in 8 domains related to our daily life. Specifically, we annotate each instructional task with a guideline, representing a common pattern shared by all task-related videos. On this basis, we annotate systematic specific steps, including their associated guideline steps, specific step descriptions and timestamps. Our proposed benchmark consists of three sub-tasks to evaluate comprehension ability of models: (1) Step Captioning: models have to generate captions for specific steps from videos. (2) Guideline Summarization: models have to mine the common pattern in task-related videos and summarize a guideline from them. (3) Guideline-Guided Captioning: models have to generate captions for specific steps under the guide of guideline. We evaluate plenty of foundation models with GUIDE and perform in-depth analysis. Given the diversity and practicality of GUIDE, we believe that it can be used as a better benchmark for instructional video comprehension.
GUIDE: A Guideline-Guided Dataset for Instructional Video Comprehension: There are substantial instructional videos on the Internet, which provide us tutorials for completing various tasks. Existing instructional video datasets only focus on specific steps at the video level, lacking experiential guidelines at the task level, which can lead to beginners struggling to learn new tasks due to the lack of relevant experience. Moreover, the specific steps without guidelines are trivial and unsystematic, making it difficult to provide a clear tutorial. To address these problems, we present the GUIDE (Guideline-Guided) dataset, which contains 3.5K videos of 560 instructional tasks in 8 domains related to our daily life. Specifically, we annotate each instructional task with a guideline, representing a common pattern shared by all task-related videos. On this basis, we annotate systematic specific steps, including their associated guideline steps, specific step descriptions and timestamps. Our proposed benchmark consists of three sub-tasks to evaluate comprehension ability of models: (1) Step Captioning: models have to generate captions for specific steps from videos. (2) Guideline Summarization: models have to mine the common pattern in task-related videos and summarize a guideline from them. (3) Guideline-Guided Captioning: models have to generate captions for specific steps under the guide of guideline. We evaluate plenty of foundation models with GUIDE and perform in-depth analysis. Given the diversity and practicality of GUIDE, we believe that it can be used as a better benchmark for instructional video comprehension.
[ 0.0308309123, 0.0033243338, -0.0008863953, 0.0451488942, 0.0077636847, 0.0288856365, 0.0102445325, -0.033835493, -0.0400295071, -0.0485712811, -0.0081558358, 0.017725192, 0.0005193803, 0.0062454557, -0.0155426078, -0.125841558, -0.0078100995, 0.0058385869, -0.0069682389, 0.054387033, 0.02618053, 0.0030563851, -0.0448878556, 0.045038078, -0.0134316627, -0.0116554685, -0.0055776453, -0.0058578318, 0.046521388, -0.037473131, 0.0051432182, 0.0320000015, 0.0611947104, -0.0424499996, 0.0000017624, 0.0036811859, 0.005364297, -0.0274735745, -0.0295997094, 0.0290288702, 0.0016263287, -0.0079977084, 0.0267806612, 0.0163887627, -0.0315699317, 0.0206578337, 0.019642232, 0.0237333868, 0.0552799515, 0.0536618456, -0.0070316293, -0.0372489132, -0.002246659, 0.0019692795, 0.0207556728, -0.0308007319, -0.0131473206, 0.0912542269, 0.0265610646, -0.0038248166, -0.0461604632, 0.035725113, 0.0325218402, -0.0067444658, 0.0038886103, 0.0299762879, -0.0161767546, -0.0347079821, 0.0100340331, 0.1001872569, 0.0232731942, 0.0392866805, 0.009260118, 0.007791304, 0.0256134998, 0.0376942828, -0.0763767511, 0.0497776493, -0.0166792981, -0.0070875618, -0.0153680108, 0.0371687338, 0.0034879863, -0.0141371451, 0.0704857334, -0.035457626, -0.0142037366, 0.0150750326, -0.0263059754, -0.0102435732, -0.0465045422, -0.0312707797, 0.05384022, 0.0274631791, 0.0081438562, 0.0283346381, -0.0522794016, 0.0363780744, -0.0201823246, -0.0374904722, 0.0770185739, 0.0559023544, -0.0371418037, 0.071655564, -0.0249532871, 0.0542435683, -0.0047820415, 0.0337420143, -0.0628285781, 0.0383421518, -0.0965459347, 0.0287842266, -0.0536814965, 0.0321267508, -0.0275654923, -0.012497182, -0.0329857282, 0.0035504024, -0.0255569275, 0.0402126387, -0.0264416374, 0.0321983919, 0.0126531953, 0.0038775152, -0.0154625326, -0.0172743611, -0.082242988, 0.0462749563, 0.0061835833, -0.0432059541, -0.0082011661, -0.0020470193, 0.0085418848, -0.0008378077, 0.022967536, 0.015629055, -0.060051024, -0.0330095552, -0.0417146757, -0.0356533341, 0.0092180539, 0.0091998549, 0.0296380762, -0.0392553173, 0.019749783, -0.0145314131, 0.0099172201, -0.0237316191, -0.0056084674, 0.0266940221, -0.0118875299, 0.0636450574, 0.0301344767, 0.0285136923, 0.0257097315, 0.0645890385, -0.0639268085, 0.0458580069, 0.0303143244, 0.0165295601, 0.0633354932, 0.0127656329, 0.0348867998, 0.0116521856, -0.0021904835, 0.0132297203, -0.0339420848, 0.0453641675, 0.0067762542, 0.0349567942, 0.0313446037, 0.0167028699, -0.0187828373, -0.00729784, 0.0370974168, 0.0314227566, -0.0307665281, -0.004286197, 0.0871017203, 0.0525551476, 0.0335130692, 0.0492728353, -0.026821712, 0.0509018488, -0.0007953264, 0.0097860014, -0.0042975745, -0.008555416, -0.0692880005, -0.0672482327, -0.0078926682, 0.0003259655, -0.0022670827, -0.0228174031, 0.0175425243, 0.0277468357, -0.0788810998, 0.0320914611, -0.0217198059, -0.0281147733, 0.0361819565, -0.009985147, 0.0086264769, 0.0482901856, 0.0108592371, -0.021097675, -0.0395002775, 0.0084642097, 0.0990919173, -0.0327905938, 0.0197781809, -0.0612752736, 0.0368210375, -0.0810511559, -0.050879471, -0.0130125135, 0.0159910824, 0.0079597281, 0.0118871387, 0.0081343809, 0.0385596529, -0.0323942453, -0.0561431423, 0.0605000146, 0.0001241845, -0.0278476197, -0.0169812515, 0.0108261881, 0.0139783798, 0.0431331806, -0.0060765883, -0.0196188502, 0.011456619, -0.0214111656, 0.0084846988, 0.0373580903, -0.0604419485, -0.0368823744, 0.0257962551, -0.1029157564, -0.0005018361, 0.0482897833, -0.0366757102, -0.0454520583, 0.0048103132, -0.0015988562, -0.018274432, -0.0094122663, -0.007905608, 0.0476024151, -0.0322779343, 0.0153637277, 0.02388441, -0.0041600615, 0.0196096115, 0.0467392989, -0.0029392019, -0.0103346445, -0.0538297966, 0.0122949984, -0.0318882465, -0.0444889776, 0.0013454623, 0.0245021544, 0.0241807494, -0.0881763548, -0.0165913831, -0.0410159752, 0.0159005728, -0.0114109498, 0.0052056368, 0.0232506022, -0.0272409413, -0.0217996296, -0.0555776507, 0.0114667267, -0.0223150402, -0.0206556153, -0.0042250114, -0.0148167871, 0.0680609047, -0.0262945984, -0.046956528, -0.0072578089, 0.0002318361, 0.0118894475, 0.0470243096, 0.0297867134, -0.0168764703, 0.0180711951, 0.017679235, -0.0076364297, -0.0188376456, -0.0213762615, 0.0143396249, 0.0060784854, -0.0118691586, 0.0504629575, -0.0077471924, 0.0514297262, 0.0072112419, -0.0391627178, -0.0254551824, 0.0200104117, 0.0093981614, -0.0193837099, -0.0165267717, -0.045642145, -0.0194636043, 0.0107368585, 0.0063364389, 0.0358126462, -0.0674541146, 0.0056021791, -0.0260837413, 0.0183859766, -0.0473760478, 0.0003258307, 0.0154465698, 0.0089747347, -0.0233139601, -0.0288762357, 0.047424987, -0.0271552615, -0.035722848, -0.0140899876, -0.0040407758, -0.0216446817, 0.0521913134, 0.001665361, 0.0120633729, 0.0004026887, -0.0386103727, -0.0510400087, 0.0068416926, 0.0720803589, -0.0044863308, 0.0589921288, -0.0801251456, -0.0292971358, -0.0278629512, 0.0117211994, -0.0511192046, 0.0013882833, -0.0147533184, -0.0184908863, -0.0273569655, -0.0060322927, -0.0625245646, -0.0105886916, 0.0109722093, 0.0520001985, -0.011153237, -0.0049706092, -0.0042946837, 0.0319092534, 0.0408621579, 0.0309208557, -0.0608367622, -0.0513599478, 0.0979962274, 0.0217577107, 0.029876044, -0.0636620969, -0.0278217457, 0.026515156, 0.0120639298, -0.0467343815, 0.0034903681, 0.0720243305, 0.0529627427, -0.0755962431, 0.0246552378, 0.1168534309, 0.0160605758, -0.0228399802, -0.0095771775, 0.0303500425, -0.0182233397, 0.0522609688, -0.0118396319, -0.0584857017, -0.0455480739, -0.0158227682, 0.0390126593, 0.0288615599, 0.0082847355, -0.0179118589, -0.0595267378, 0.0631615445, 0.0292505622, -0.0427301154, -0.0288609099, -0.0335407443, -0.0169566423, -0.046629034, -0.0069871149, -0.0152867893, -0.047954157, 0.0443079136, 0.0467030741, -0.0241203308, -0.0741692483, 0.0319325514, -0.0349415503, 0.0426966213, 0.0259057861, 0.0075535607, -0.018591132, 0.0007943062, 0.0067020901, -0.0275875814, -0.0068210321, -0.0131972907, 0.0872605667, -0.0132962763, -0.0066260798, 0.0169893038, 0.0647282153, -0.0326552726, -0.0381018855, 0.0565872379, 0.0470122173, 0.0062630153, 0.0182496514, -0.064838618, 0.0338889249, 0.0227423366, 0.1063061431, 0.0152188791, 0.0114530679, 0.0320736691, 0.0128806578, -0.0066206777, -0.0764152408, -0.0063434509, -0.0144683141, 0.0142896511, 0.0488726161, -0.0688091516, -0.0412712619, 0.0761562958, 0.0229958557, 0.0064998115, 0.0115385856, 0.0288856607, 0.0469038263, -0.0739140734, 0.0488295108, -0.0119018154, -0.0172789451, -0.1015055478, -0.0416984521, 0.0346239656, -0.0274388473, -0.0222573392, 0.0368704125, -0.0176787693, 0.0033672254, -0.0284710303, 0.058028996, 0.0188615844, 0.0573992506, 0.0320102088, 0.0175744575, -0.0316614099, -0.0276957266, 0.0019444952, -0.0169067662, 0.0144577408, -0.0179005228, -0.0195052531, 0.0039735609, -0.0372159183, 0.0136106256, -0.0475450531, 0.0280924886, -0.0498412848, -0.0378175415, 0.0201772172, -0.0487184897, 0.0243278407, -0.0003456127, 0.0115417782, -0.0429684557, 0.0128227267, 0.0105215935, 0.0291375145, -0.0634402856, 0.0534623452, 0.045876883, -0.0012769813, 0.0005702659, 0.0439557359, -0.0294192843, -0.0025928677, -0.0612858199, 0.0028212522, -0.0310711209, 0.0125800911, -0.036679972, 0.011107604, 0.0048058825, 0.0061405613, 0.0183590595, 0.0267768465, 0.0483553559, 0.0902250931, 0.0163512528, 0.0255131312, 0.0157684553, 0.0420806184, -0.0104128718, -0.0182433203, -0.0387180671, 0.0098551596, 0.0250426903, -0.0667236447, -0.0025214825, -0.0079317763, 0.0129481535, -0.0343137346, 0.0491732135, -0.0362483598, -0.0593945161, 0.0403656214, 0.0147289233, 0.0215371232, 0.0079144305, 0.1003496051, 0.0046000252, 0.0448119789, 0.010764705, -0.023496205, -0.0358468816, -0.0256132521, 0.0217603482, 0.0201730039, -0.0399156213, 0.0191804431, -0.0331234336, -0.0462793484, -0.0527364686, -0.027910227, 0.0205319747, 0.0233319905, 0.0749000236, 0.0125644561, -0.0345450826, -0.020275278, -0.0192269292, -0.0305878595, -0.0638343692, -0.0059782723, -0.0294802189, -5.168284257e-33, -0.0463222116, -0.0069341059, 0.0172053166, 0.0265593026, 0.0287887324, 0.0221774094, -0.0398540199, 0.0412059575, -0.0223644711, 0.044129584, 0.0299107563, 0.0143664377, 0.0209751278, 0.0370377377, 0.0298446044, -0.0330766104, 0.0628532544, -0.0229124334, -0.0308495276, 0.0357856862, -0.000626117, 0.0077726734, 0.0297121704, -0.0754888654, -0.0175741222, -0.0182510763, -0.0469878428, 0.0071729901, 0.0337589383, 0.0746406913, -0.0021498851, -0.0594019666, 0.0188197214, -0.0435038581, -0.0372171104, -0.0152491638, -0.079515785, 0.0011487105, 0.0134605244, 0.0264248904, -0.0631272867, -0.0246520713, 0.0084153749, -0.0142355738, -0.0084593622, -0.0229970906, 0.0868918821, -0.0450165719, 0.0337440893, 0.0139508629, -0.0121668167, -0.0062216655, -0.0659532174, -0.1027056426, 0.0205586925, 0.0253474899, 0.0180361066, -0.0282320399, -0.0488019846, 0.0122181326, 0.0102729676, 0.0421415046, 0.0930869952, -0.0111184185, -0.023010999, 0.0402896442, -0.0118242884, -0.0542150289, 0.0050949291, -0.0147809442, -0.014879195, 0.0888844579, -0.0017309359, -0.0215130504, 0.0498519093, 0.0441414006, -0.0344914198, 0.1013067216, -0.0342647135, 0.0400486253, -0.0120254317, 0.0036706978, -0.0328168571, 0.0408035927, -0.0287304744, -0.0023792887, 0.0040511596, -0.0153170107, 0.0325375833, 0.0089222481, -0.0038669596, -0.0277764574, 0.032979086, -0.0321253911, 0.0271554161, 0.0068761171, -0.0278125275, 0.0290318932, 0.0118419835, 0.0030585476, -0.0807797909, -0.0340662524, 0.0394563638, -0.0084798187, 0.0130260903, -0.001507719, 0.0228320938, 0.0532202125, -0.0153994272, -0.0009399807, -0.0070757223, -0.0142739071, -0.0399120525, 0.0213854611, 0.015371725, 0.0404435843, -0.0855160132, 0.0579117797, -0.0364305712, -0.00208778, -0.0141063286, -0.026142925, 0.0198372602, 0.0107206497, -0.0264812652, -0.0059167328, -0.040975716, -0.009603492, 0.0387757346, -0.0301090796, -0.001171625, 0.0337649249, 2.38e-7, -0.0198214483, -0.0124517754, 0.0122350231, 0.0037161761, -0.0070546069, 0.057801228, -0.0363359153, -0.0057051806, 0.0019163286, -0.048476126, 0.0512089431, 0.0474831648, 0.0147271734, -0.0105558187, -0.0555529743, -0.0473124944, -0.0183205791, -0.017885806, -0.0286409575, -0.0185028911, 0.0281272121, 0.0597462431, 0.0101931384, 0.0066621285, 0.0008847417, -0.0534410886, 0.0480584055, -0.0132393166, -0.0095016686, 0.0064243488, 0.054744035, 0.0369654149, -0.0460992083, -0.0037028533, -0.0304284431, -0.016551381, 0.0587682761, 0.0638610497, 0.0172576886, -0.0110883666, -0.0408795774, 0.0344470665, -0.0115203811, 0.000719059, 0.0241231415, -0.0250548441, -0.0247300696, 0.021030331, -0.0294784922, 0.0275398064, -0.0256444849, 0.0298141688, -0.005818957, 0.0124237062, 0.0073155425, -0.0226553548, -0.0418487862, -0.0635369271, 0.0072208592, 0.003717314, -0.0274136644, 0.0024713089, -0.0493998341, -0.003664135, -0.0098925531, -0.0212029405, -0.0525105707, 1.662610256e-34, 0.0144377695, -0.0034605132, 0.0050324625, -0.0146374144, 0.0176505763, -0.0049670665, -0.0749069005, 0.0074946014, -0.0088662459, -0.0248900782, -0.0055910409 ]
http://arxiv.org/abs/2406.18221v1
http://arxiv.org/pdf/2406.18221v1
Enhancing Data Privacy in Large Language Models through Private Association Editing
2024-06-26T10:08:47
2024-06-26T10:08:47
Davide Venditti, Elena Sofia Ruzzetti, Giancarlo A. Xompero, Cristina Giannone, Andrea Favalli, Raniero Romagnoli, Fabio Massimo Zanzotto
cs.CL
cs.CL, cs.AI
Large Language Models (LLMs) are powerful tools with extensive applications, but their tendency to memorize private information raises significant concerns as private data leakage can easily happen. In this paper, we introduce Private Association Editing (PAE), a novel defense approach for private data leakage. PAE is designed to effectively remove Personally Identifiable Information (PII) without retraining the model. Our approach consists of a four-step procedure: detecting memorized PII, applying PAE cards to mitigate memorization of private data, verifying resilience to targeted data extraction (TDE) attacks, and ensuring consistency in the post-edit LLMs. The versatility and efficiency of PAE, which allows for batch modifications, significantly enhance data privacy in LLMs. Experimental results demonstrate the effectiveness of PAE in mitigating private data leakage. We believe PAE will serve as a critical tool in the ongoing effort to protect data privacy in LLMs, encouraging the development of safer models for real-world applications.
Enhancing Data Privacy in Large Language Models through Private Association Editing: Large Language Models (LLMs) are powerful tools with extensive applications, but their tendency to memorize private information raises significant concerns as private data leakage can easily happen. In this paper, we introduce Private Association Editing (PAE), a novel defense approach for private data leakage. PAE is designed to effectively remove Personally Identifiable Information (PII) without retraining the model. Our approach consists of a four-step procedure: detecting memorized PII, applying PAE cards to mitigate memorization of private data, verifying resilience to targeted data extraction (TDE) attacks, and ensuring consistency in the post-edit LLMs. The versatility and efficiency of PAE, which allows for batch modifications, significantly enhance data privacy in LLMs. Experimental results demonstrate the effectiveness of PAE in mitigating private data leakage. We believe PAE will serve as a critical tool in the ongoing effort to protect data privacy in LLMs, encouraging the development of safer models for real-world applications.
[ 0.0315022096, 0.1154447868, -0.0240769517, 0.0835983604, -0.034967877, -0.0194263849, 0.0623105615, -0.0059271432, 0.0102836275, -0.0162161775, -0.0271275844, -0.0235526524, 0.0218127817, 0.0016257963, 0.0491319709, 0.0287220534, 0.0573213436, -0.0436584912, 0.0005931275, -0.0286942385, -0.0170251541, -0.0235465094, 0.0297028404, -0.008224953, 0.0188422687, -0.0234375391, -0.0177247152, -0.0264435243, 0.0237118918, -0.0137506109, -0.0106648151, 0.0700728968, 0.000566324, 0.0731672272, 0.0000021392, -0.0357972421, -0.0767137781, 0.0538189262, -0.0459742583, 0.0637825951, -0.0268010367, 0.0095628873, -0.0167976469, -0.0002689244, -0.0197075363, 0.0124674551, 0.0447387211, 0.0236385558, 0.0936271846, 0.0719651505, -0.0080187051, -0.028433118, 0.0325861759, -0.0262408033, -0.0043283952, -0.0439411774, 0.0612344742, -0.042756781, -0.054310903, 0.0012992594, -0.0007764671, 0.030234864, -0.0056094895, 0.0000696754, -0.0294865295, 0.0076834601, -0.017217312, 0.0018822798, 0.011679505, 0.0233881138, 0.0213654581, -0.0249804389, 0.0312017724, 0.0191802103, 0.0220993515, 0.0194612369, 0.0113129746, 0.0403316468, -0.005871383, -0.009928368, 0.0178198293, -0.0503881387, 0.0421251878, 0.0002133126, -0.0091974726, 0.0191529021, 0.0381617509, -0.020659769, -0.0358075127, -0.0541658401, 0.0277629513, -0.0628994852, 0.0437726751, 0.0105267214, 0.0460933037, 0.0348157808, -0.075967744, -0.0526875779, 0.0086800745, -0.0475204512, 0.0330921337, 0.0191796273, 0.0207476672, 0.0253587645, -0.0115407519, 0.0467116833, -0.01238144, 0.0420935862, 0.0498426259, 0.0284686685, -0.0238378253, -0.005971822, -0.0097301761, 0.0656543374, -0.008992563, 0.0500990264, -0.0075190268, 0.0429543965, -0.0166345667, 0.0604184493, -0.0019206228, -0.0262408797, -0.0020286872, -0.024350632, -0.005685077, -0.0441563167, 0.0206953213, -0.0194706451, 0.0015757809, -0.0468507484, -0.0131840538, -0.0539312698, -0.0108403238, -0.0292394273, -0.0258109178, 0.0734068081, 0.0211066622, -0.0094992639, -0.0668199733, -0.093877852, 0.0321845599, -0.0254685599, 0.0051916363, -0.0614085272, 0.0043596905, -0.0307295639, -0.0201067831, -0.107992813, -0.0054112487, 0.0162252337, -0.0369303003, 0.0492485166, 0.0008270711, -0.0014404313, 0.0175802428, 0.0467391014, -0.0015876197, 0.006585049, 0.0264488533, 0.0310992263, 0.0690306276, -0.0209387969, 0.0491061546, 0.0482204258, 0.0225606225, 0.0014689171, -0.0408105217, 0.0370048545, 0.0318767913, -0.0741856098, -0.0200065654, 0.0413121656, -0.0144693376, -0.0646093786, 0.1067450792, 0.0377158038, -0.0189667158, 0.0594181903, -0.0609752089, 0.0120584164, 0.0119148577, -0.0131576862, -0.0574255921, -0.0053588981, -0.0789655522, -0.0247977842, 0.0022714692, 0.0114507545, 0.0055050249, 0.0157899, 0.020738529, -0.0462808385, 0.0333458744, 0.018437814, -0.0020457609, 0.0272993557, 0.0238678306, 0.0695355311, 0.0241775401, -0.0630644411, -0.0200926196, -0.0717050731, -0.0055345669, 0.0623447299, 0.0533615015, 0.0004316549, -0.0065871147, -0.0420601964, -0.0556663163, -0.0020527293, 0.0069020349, -0.0009913705, 0.0633449256, -0.0163113344, -0.0324093662, 0.0287762303, 0.0245780386, 0.0164362341, 0.0249739569, -0.0368706696, 0.0191379562, -0.0166496467, -0.0118958009, 0.0073363185, -0.0314528309, -0.051381655, 0.043590527, 0.0353371352, 0.0029441256, 0.0309671201, -0.0291483197, 0.0136462897, 0.0196463391, 0.04866568, 0.0406837873, -0.0312227793, 0.0031891228, 0.0240893047, 0.0001809895, -0.0106207523, -0.0026322501, 0.053125035, -0.0009328014, -0.0047293031, 0.0060484549, -0.0265092403, 0.0444700383, -0.011390077, 0.0391208977, -0.0180013292, 0.014712302, -0.0026405021, 0.0018453432, -0.033926148, -0.0436371192, -0.0406552218, -0.0019271898, -0.0037384774, -0.0552275889, -0.0221223813, 0.0206648801, 0.0019456472, 0.0130231455, -0.0015805386, -0.0135799525, -0.0289277341, -0.0285369065, -0.0070093484, 0.0083417138, 0.0118887993, 0.0395728871, 0.0162317269, 0.0203129798, -0.0395967662, -0.0219418164, -0.0370177627, -0.0359534547, -0.0577192418, 0.0306228958, 0.0203743596, 0.0192019399, -0.0453384258, -0.0048479978, -0.001517039, -0.0000397823, 0.0347439684, -0.053501606, -0.052817408, -0.0300328583, 0.032278195, -0.0489550233, 0.0244599357, 0.0089734225, -0.0252851397, -0.0097667733, 0.0012810504, -0.0014385515, 0.0414025187, -0.0060623847, -0.0202504043, 0.0423605852, -0.0337944068, -0.0369265191, -0.0264568236, 0.0269292258, -0.0133481463, -0.0206994303, -0.0159929618, -0.0029196902, 0.0292670224, 0.0228206888, -0.0225895848, 0.0052363868, 0.036624033, 0.0233172085, -0.0207066853, -0.0030921523, -0.0430964716, -0.0191522501, 0.0000025312, -0.0309332814, -0.0546573699, 0.0666537732, -0.0194513611, 0.0443229638, -0.0228495076, -0.0299693961, 0.0220204126, 0.0756363049, 0.0074495981, 0.0367262736, 0.0577475987, -0.0544990413, -0.032469213, -0.0349890515, 0.0251103304, -0.053389255, 0.0289769564, -0.0443504229, -0.0412095599, 0.0138793001, -0.0090820128, -0.0567200519, -0.053098321, -0.0132350624, -0.0177125819, -0.0087487856, -0.0496043898, 0.0019013248, -0.0439579152, 0.0570442043, 0.0183427893, -0.0211116355, 0.0209765993, -0.017931683, -0.020952858, 0.01552564, -0.0739058107, -0.0248851739, -0.0337734818, -0.0112301176, 0.0843409896, 0.0589917377, -0.0268463325, 0.0077426927, -0.0319952108, 0.0101638418, -0.0432658345, -0.0121209128, 0.0272640586, -0.0102626961, -0.0619946644, -0.013286721, 0.0671070814, 0.0187184252, -0.070564568, -0.0019822826, 0.0267588161, -0.0124276932, -0.006888771, 0.012790842, -0.0901030824, 0.0108933821, 0.0134160938, 0.0312107299, -0.0285729263, -0.0469697788, -0.0441990346, -0.0952237993, -0.0243891366, 0.0404007435, -0.0275052581, -0.0229533948, -0.0368213654, 0.0161089413, 0.0152451079, 0.0124643305, -0.0266665164, 0.0121844029, -0.0127523821, -0.0066702403, -0.0326180309, -0.0229640119, -0.0186313987, -0.0285702106, 0.0635775179, 0.0538164563, -0.003133473, -0.027885912, -0.0120681059, 0.0216346234, 0.0694645047, -0.0225802995, -0.0000085151, 0.0449795499, 0.0119412616, 0.0012699183, 0.0217493661, 0.0322208516, 0.0481087491, -0.0196753033, 0.046631135, 0.0404403619, -0.0268031806, 0.0351093411, 0.0232235771, -0.0512775294, 0.0131877698, 0.055322241, 0.0016126495, -0.0133977393, 0.025767684, -0.0670879185, 0.0224291086, -0.0112632969, 0.0101393303, -0.0179201998, -0.0208173506, -0.0517961048, -0.0291000269, -0.0023658937, -0.0160472151, -0.001715149, 0.0754356161, 0.007193936, 0.0487313047, -0.0011516394, -0.041087307, -0.0193247907, -0.102187857, 0.0202545188, -0.0499389134, -0.0132526541, -0.0029974906, -0.0039619538, -0.0057580676, -0.0049316185, -0.0630406439, -0.0933557525, 0.0401974916, 0.0500314645, 0.0200809259, 0.0418474004, -0.0033604871, -0.1029401645, 0.0685485378, 0.0146631878, -0.0464531183, 0.0049616676, 0.0614074469, -0.033096049, 0.0103600956, 0.018609887, -0.0231837705, -0.0409762673, 0.0705832615, -0.0438511707, -0.0466850139, -0.0394647233, -0.0345059149, 0.0723424777, 0.01400756, 0.0290370956, 0.0339004621, -0.0365952253, 0.0638967305, -0.0617672652, 0.0041700602, 0.0140453866, 0.0155631462, 0.0377787985, 0.0089064669, 0.04484367, -0.0581141859, 0.0030387717, 0.0275483839, -0.011008217, 0.0210114941, 0.0209731944, 0.0274528265, 0.0312103312, 0.0096822074, -0.0130848261, 0.0447983369, 0.0542658269, 0.0330534466, -0.029712338, 0.0527793802, -0.0137985274, -0.02439503, -0.0355080627, 0.0330059417, -0.0344788656, -0.0038166451, -0.00130067, 0.052825395, -0.0178983156, 0.0116135674, -0.0034656511, -0.0312629305, 0.00502053, -0.0404140949, -0.0227924995, -0.0137426481, -0.0329170525, -0.0235592127, -0.0184098128, -0.0074038729, 0.0571642891, 0.0565242358, -0.0196884591, 0.0530614853, -0.0321970023, -0.0205092132, -0.0736769438, -0.0165250935, 0.0338206887, 0.0113859279, -0.0669521615, -0.0532133617, -0.0332964212, -0.0634274557, -0.0159245227, -0.0130980806, 0.0282735899, 0.0041947062, -0.0196721125, 0.0099104112, -0.0088473242, -0.0016910079, -0.0903770328, 0.0247321203, 0.0304904915, -0.0366326235, 0.0052770567, -5.484526033e-33, 0.0164426342, 0.0069758925, -0.0270300694, 0.0285570454, 0.0257167425, 0.0268445779, -0.0331257544, 0.0176356789, -0.0074682306, -0.0308349561, -0.0460856408, -0.0079105161, 0.0152795361, 0.0151716201, -0.0164585933, 0.0020787821, 0.0361161493, 0.0013802856, -0.012768914, 0.0264740512, 0.0365367681, 0.0298260562, 0.0573365167, -0.0948287398, 0.0607268028, -0.0171701144, -0.0455105752, 0.0177251846, -0.0536698066, 0.0761775598, 0.005605862, -0.0357274897, -0.011373587, 0.00667318, -0.0179108884, 0.0197292361, -0.0340097174, -0.0159093421, 0.0013386981, 0.0070734904, 0.0490832031, -0.0784338191, 0.0226270165, -0.0376581587, -0.0212622974, 0.0065794005, -0.0156220105, -0.0214203689, -0.0037063819, -0.0058823545, 0.0130877346, 0.007173209, -0.0124794571, 0.0130715016, 0.0833168775, -0.0403838679, 0.024219688, -0.0451409295, -0.060072355, 0.0554402359, 0.0783440843, 0.0041716942, 0.0460776351, 0.0185347907, 0.0292798113, 0.0387211815, 0.0485275798, 0.0389601476, 0.0536285676, 0.0178331248, -0.0088356109, 0.0226666462, -0.0170011241, 0.037108209, 0.0231519118, -0.0301046167, -0.0152858896, 0.0219505019, -0.0218343455, 0.0090458626, 0.038758155, 0.0198780578, -0.063772358, 0.0241017621, -0.0431722738, 0.016890673, -0.0473201908, -0.0097582517, -0.0196243227, -0.010545671, -0.0675636008, 0.0170982648, 0.010949376, -0.0596610047, -0.0313531831, 0.036661014, 0.0455771089, -0.0303642284, -0.0108347982, -0.0025505342, -0.0193234161, 0.0029847529, 0.0060860557, -0.0427809507, 0.0156265907, 0.0137834176, 0.0357497893, 0.021840442, -0.0999265015, -0.0273053832, -0.0253117792, 0.0187575947, 0.0322766118, 0.0102597978, 0.0162489694, -0.0011557175, 0.0165426582, 0.0192297325, 0.058319848, 0.0165897366, -0.0507759079, 0.055828914, 0.0000360067, 0.0096953437, 0.0312277395, 0.0005020247, 0.0422527343, 0.0326720998, 0.1037614569, 0.0036112021, 0.0200431701, -0.0367869586, 2.865e-7, -0.0103297234, 0.0248779468, 0.0000575811, -0.016180886, 0.0237722658, 0.0755645931, -0.0006768323, 0.0504916273, 0.0238725524, 0.0164876413, -0.011121518, -0.0035418386, 0.0365363099, 0.0386476666, -0.038203802, 0.0257830061, -0.0169076566, -0.0056155492, -0.0611204244, 0.0124026658, 0.0157888848, 0.0751640871, 0.061477948, -0.0271093491, -0.0385811999, -0.0364348441, 0.0138550615, -0.0332288183, -0.0504616462, -0.0024967124, 0.0236580577, 0.0660673603, 0.035063535, 0.0087828198, -0.0591695234, 0.0181175638, -0.0658968389, 0.0792330056, 0.0099811796, 0.0004670117, 0.0039916802, -0.0642906129, -0.0172889046, 0.0005887641, 0.0416343547, 0.0227587428, -0.015144676, 0.0011144807, 0.0116863139, -0.0495372266, 0.0228407253, -0.0192105286, -0.0097407484, 0.0239222888, -0.0010695662, 0.0132266032, 0.0048537315, -0.0404815637, -0.0560954697, -0.0554503128, -0.0121643292, -0.0336743481, 0.0204696227, 0.0091596041, -0.0147594381, -0.0435009599, 0.0525304489, 2.672447606e-34, -0.0298465397, -0.0356343873, -0.0268574245, 0.0375882722, 0.0595514104, -0.0098800259, 0.0191800389, 0.0333530679, -0.006989622, -0.0428536162, -0.0333793275 ]
http://arxiv.org/abs/2406.18219v1
http://arxiv.org/pdf/2406.18219v1
A Closer Look into Mixture-of-Experts in Large Language Models
2024-06-26T10:07:57
2024-06-26T10:07:57
Ka Man Lo, Zeyu Huang, Zihan Qiu, Zili Wang, Jie Fu
cs.CL
cs.CL, cs.LG
Mixture-of-experts (MoE) is gaining increasing attention due to its unique properties and remarkable performance, especially for language tasks. By sparsely activating a subset of parameters for each token, MoE architecture could increase the model size without sacrificing computational efficiency, achieving a better trade-off between performance and training costs. However, the underlying mechanism of MoE still lacks further exploration, and its modularization degree remains questionable. In this paper, we make an initial attempt to understand the inner workings of MoE-based large language models. Concretely, we comprehensively study the parametric and behavioral features of three recent MoE-based models and reveal some intriguing observations, including (1) Neurons act like fine-grained experts. (2) The router of MoE usually selects experts with larger output norms. (3) The expert diversity increases as the layer increases, while the last layer is an outlier. Based on the observations, we also provide suggestions for a broad spectrum of MoE practitioners, such as router design and expert allocation. We hope this work could shed light on future research on the MoE framework and other modular architectures. Code is available at https://github.com/kamanphoebe/Look-into-MoEs.
A Closer Look into Mixture-of-Experts in Large Language Models: Mixture-of-experts (MoE) is gaining increasing attention due to its unique properties and remarkable performance, especially for language tasks. By sparsely activating a subset of parameters for each token, MoE architecture could increase the model size without sacrificing computational efficiency, achieving a better trade-off between performance and training costs. However, the underlying mechanism of MoE still lacks further exploration, and its modularization degree remains questionable. In this paper, we make an initial attempt to understand the inner workings of MoE-based large language models. Concretely, we comprehensively study the parametric and behavioral features of three recent MoE-based models and reveal some intriguing observations, including (1) Neurons act like fine-grained experts. (2) The router of MoE usually selects experts with larger output norms. (3) The expert diversity increases as the layer increases, while the last layer is an outlier. Based on the observations, we also provide suggestions for a broad spectrum of MoE practitioners, such as router design and expert allocation. We hope this work could shed light on future research on the MoE framework and other modular architectures. Code is available at https://github.com/kamanphoebe/Look-into-MoEs.
[ 0.0724910796, 0.0707233921, 0.001186475, 0.0371967517, -0.0287535023, -0.0045648101, 0.0428110436, -0.0128656449, -0.0382816195, -0.011674582, -0.058836475, -0.0352468938, -0.0001394138, 0.0149993198, 0.042634856, -0.034138009, 0.0143964794, -0.0553087369, 0.0059464467, -0.0276063811, 0.0025942978, 0.0013505116, -0.0082583176, 0.06474033, -0.0023010944, 0.037165083, 0.0091933981, -0.0258957837, -0.0086369384, 0.0293198023, 0.0042146966, 0.0652513131, 0.0347899608, 0.0523170978, 0.0000018764, 0.0001725784, 0.0169847328, 0.0116029698, -0.0173587855, 0.0061900271, 0.0187347457, 0.0012073237, -0.0063832388, -0.0002364398, 0.0019093005, -0.0330138803, 0.066116266, 0.0478373058, 0.0032192471, 0.1218123138, 0.0106221847, -0.0215120334, 0.0745067969, -0.0410430618, 0.0098006651, -0.0171443783, 0.0196032524, -0.0123399254, 0.0014224657, 0.0229475349, -0.0290429238, 0.0122420862, 0.024726538, 0.0391888618, 0.0418821871, 0.0530247465, 0.0495079793, -0.0301005114, -0.0128246518, 0.0388276204, -0.0620806515, -0.004172944, 0.0022592272, -0.0008354703, -0.0163502973, 0.0247431044, -0.0402392931, 0.0207652412, 0.0089453831, 0.0213018712, 0.0186035838, 0.0237035248, 0.0550905578, 0.0235413667, -0.0731767789, 0.0304449983, 0.0216994118, 0.0000352515, -0.0191665702, -0.0147869317, -0.0316704288, -0.046590887, 0.0230133086, 0.0283045135, -0.0169646256, -0.0055135284, -0.0234876722, -0.0255438648, -0.0143008595, -0.0016184911, 0.025833888, 0.0423097201, 0.0092461724, 0.0614691265, -0.0298163388, 0.0218972713, -0.0363719836, 0.0278811604, -0.0336233601, -0.0194963962, -0.0518583357, -0.0063337055, -0.0180710517, 0.0311845131, 0.0057533006, -0.0391438417, -0.0298693273, -0.0028162245, -0.0109725017, -0.0013332341, 0.0207427591, 0.0029873995, -0.0357054621, -0.0420633443, -0.0219625607, 0.03620122, -0.0488620028, 0.0287667532, 0.0371371172, -0.0383733287, 0.0114895608, 0.0422231443, -0.0179173443, -0.020545641, 0.0257316306, 0.0516900085, -0.0437022448, -0.0375417098, 0.0018290586, 0.0166952796, -0.0339639485, -0.0810819268, 0.0161600616, -0.0943888426, 0.0635831431, 0.0091700898, 0.0278533194, -0.0450906791, -0.0090043172, -0.0145953968, -0.054360453, 0.025509024, -0.0332121961, 0.0109855849, 0.0164571032, 0.0135231121, -0.0475400724, 0.0574990883, 0.0182959847, 0.0234940015, -0.001581579, -0.0032531221, 0.0189942829, -0.0030823371, -0.0129321143, -0.0112082809, -0.0195700135, -0.0236191712, -0.0220152941, 0.0199005455, 0.0436237678, 0.0768005475, -0.070723772, -0.0201426595, 0.0554333553, 0.0328904241, 0.0278342757, 0.0647105053, 0.0616196543, -0.0284110941, 0.0326719694, 0.0358339623, -0.0424523987, 0.0130981477, -0.0335933156, -0.0088953432, 0.0289445389, 0.0684608668, -0.0577378087, -0.0316011794, 0.0202828944, -0.0066470113, -0.0495672449, 0.0190322939, 0.0744363964, -0.0339422934, 0.0091268485, 0.0035962642, 0.039833691, -0.0037497703, 0.021280231, -0.0795962363, 0.0285691693, 0.059210062, 0.0237214528, -0.0319580398, -0.0088962298, -0.0021762967, -0.0348458812, -0.0100685116, 0.0316830538, 0.0124039194, 0.0117051844, -0.0784050301, -0.0102583775, -0.0281008296, 0.0352361761, 0.0397626162, 0.021535363, -0.0008387326, 0.0188775696, -0.0254959892, -0.0080929743, 0.0215484649, -0.0260765851, 0.020275984, 0.0112793455, 0.0056096627, 0.0293406844, 0.0004542638, -0.0412266031, 0.0389227755, 0.0003836788, -0.0462265611, -0.0184658393, -0.024109818, -0.0770139918, -0.0163503308, 0.0429417267, -0.0100706778, 0.0135163506, -0.0195008311, -0.0125198485, -0.0576494709, 0.0237965547, -0.0338846594, 0.0339701846, -0.0458865091, -0.0150207793, 0.0157156866, 0.0241521075, -0.0195953529, 0.0772003233, -0.0294455495, -0.008856752, 0.015909465, 0.0149640292, -0.008594485, -0.0098710069, -0.0300508291, -0.0092633488, 0.0087795639, 0.0580697283, 0.0092366626, -0.0214694291, 0.0037321339, -0.1169217899, -0.1096547544, 0.0166191161, -0.0441635028, -0.0174502153, 0.0007602156, -0.0022870174, 0.0007319707, 0.0336471759, -0.0154118622, 0.0165741798, -0.0058038863, 0.0171843823, -0.0203088764, 0.0204851534, -0.0266943537, -0.03238298, -0.0722451657, -0.0035495139, 0.0602561906, -0.0462462753, -0.0180714447, -0.0410131067, -0.0189962368, -0.0415844843, 0.020071201, -0.033734601, -0.0194599163, 0.0061281468, -0.0653625429, 0.0941839516, 0.0604634807, -0.0029528488, 0.0508425497, 0.0527949706, -0.0094332127, -0.0115225697, 0.0209559258, -0.0214404259, 0.0317161418, -0.0581217594, -0.0024484347, -0.0325507559, 0.0634319857, 0.0025866635, -0.0467069633, -0.0518357307, 0.00376103, -0.0017004956, -0.0452623256, -0.0036281545, -0.0010141219, 0.016319247, 0.0362954922, -0.0260656774, -0.0315446593, 0.0451357961, -0.0060334667, 0.0006102319, -0.0055438941, -0.0534651428, -0.0390205644, 0.014695758, 0.0310133453, 0.0330935605, -0.0356671028, -0.0061654337, -0.0465214103, 0.0082490463, -0.0059771198, -0.0076155672, 0.0276376437, -0.018592149, 0.0148351872, 0.0033370042, -0.0329266936, 0.0200199261, -0.0161875561, -0.0442538895, 0.0586738586, 0.0489877686, -0.0459882431, -0.0491650663, 0.0277492572, 0.0258288607, 0.0546032824, -0.0407075509, 0.0168616343, -0.037880525, 0.0367918722, 0.017300589, 0.014862204, 0.0323276632, -0.0666266158, 0.0153698325, 0.0145125603, 0.0503667966, -0.0321364813, -0.0538637191, 0.0051778383, 0.0060927751, 0.00006918, -0.0320202298, 0.0183418449, 0.0558816493, -0.0522280708, 0.007475026, 0.100640133, -0.0158466697, -0.0663488284, -0.0098885158, 0.0421135947, 0.0492384806, 0.0038154777, 0.0377338529, -0.0164007228, -0.0114797382, 0.0048110075, -0.0254326947, 0.0073271948, -0.0164342746, 0.0118327849, -0.0755695105, 0.0230283178, 0.0195810143, -0.047676973, -0.0437246934, -0.0145341149, -0.0147148827, -0.0263010319, 0.0251932628, 0.0064017042, 0.0052925344, -0.0156724863, 0.0331858508, 0.0178978425, -0.0597034208, 0.0210185926, -0.056551706, 0.083525151, 0.0250211637, -0.0201226175, -0.015744498, -0.0011034256, 0.0208053421, 0.0021535431, 0.0165181886, 0.0043355417, 0.0136001976, -0.0224730112, -0.0206911284, -0.0004240216, 0.0130402138, 0.0026466034, -0.0015128178, -0.0288858302, 0.0778639391, -0.0120545728, -0.0425935946, -0.0110007636, -0.0266617201, -0.0319197737, 0.0545484014, 0.036098171, -0.0464132205, -0.0251031984, -0.0180356819, -0.0348423123, 0.0061603882, 0.0050278138, 0.0161572658, 0.0246406589, -0.0850966871, -0.0156562794, -0.0116400719, 0.052894637, 0.0642219707, -0.0043435888, 0.0072907866, 0.0024607438, -0.0012026121, 0.008256319, 0.0203679744, -0.0269190967, -0.0015558252, -0.0887361467, -0.0291861724, 0.0417713113, -0.0231860597, -0.0103278989, -0.0343092829, -0.095527038, -0.0840057284, -0.0178320426, 0.0807803199, 0.0209602062, 0.0333884843, -0.0167313423, -0.1046627909, 0.0815327391, 0.0323780999, -0.0536694601, 0.0050170748, 0.0205616616, -0.0307071637, -0.001200394, 0.0305653028, -0.0084849661, 0.0090403929, 0.0497776605, 0.022244839, -0.0484929569, 0.0168005526, -0.0028242471, -0.0450331122, 0.0234909412, 0.0037824877, 0.0336293131, -0.0545120351, 0.0073723681, 0.0054242574, -0.000601758, -0.0514177792, -0.0130541576, 0.0555561036, 0.0181678329, 0.0175759681, 0.0127613693, 0.0678713918, 0.0318901837, 0.0011275082, 0.0115743596, -0.0067509282, -0.018152874, 0.0224257521, 0.005916879, -0.0228369087, 0.0271935761, -0.0129763279, -0.002997404, -0.0208605025, 0.035820391, 0.0013799081, -0.0258259811, -0.0428993367, 0.0191267151, -0.1023337916, -0.0774046183, -0.0025271417, 0.0772177055, 0.0601798631, -0.0164755322, 0.0299662743, -0.0126634762, -0.0158089902, -0.0367507301, 0.0061872602, -0.0147873303, 0.0202057362, 0.0140597885, 0.0357679501, 0.0142240217, -0.0068175234, 0.074303627, 0.0372365266, 0.0096722515, 0.0216337275, 0.0079604108, -0.031619817, -0.003011178, 0.019647738, 0.0326565653, -0.0511513688, 0.0164561607, -0.021267565, -0.0466849543, -0.0344712809, -0.0034814882, 0.0169883203, 0.0046293531, 0.0191084072, 0.0416986048, -0.0417285487, -0.0017983522, -0.0210837778, -0.0153410668, 0.0338800661, -0.0007382539, -0.0061310134, -6.566450299e-33, -0.0046085287, 0.0008195708, 0.0301571544, -0.0657246336, 0.0185774341, -0.0184152946, -0.0045204242, 0.0197437946, -0.0342639722, 0.0074071735, -0.0388231091, 0.005235007, 0.0041126329, 0.0321452096, 0.0142169911, 0.0159885027, 0.0655080974, -0.0142311836, 0.0159171931, 0.0370848961, 0.0045052152, 0.0093747908, 0.0542962514, -0.0337090269, -0.039100185, 0.031738434, -0.0230314136, 0.0333974399, -0.0378916152, 0.0745810345, -0.0235145837, 0.0244150385, 0.0320249945, 0.0086031267, -0.0308885444, -0.0373799503, -0.0522450916, -0.028407827, 0.0063389116, -0.0304942522, -0.0113176005, -0.0399781317, 0.0777825788, 0.003842236, 0.0131299496, -0.0327021517, 0.0327209495, -0.0050179497, -0.0535339378, -0.0049428828, -0.0267766174, 0.0396848135, 0.0079855006, 0.0112600308, 0.0188084859, -0.0360638872, -0.0155522786, -0.0059892694, -0.0033138967, 0.0044096303, 0.0412255116, 0.0819242895, 0.0043546101, 0.0492284633, 0.0091740219, 0.0325650908, 0.0661690086, 0.0252733082, -0.0316526107, -0.0165417716, 0.043845769, 0.0624207929, 0.0591674261, -0.0273909494, 0.0388729945, 0.002757299, -0.0437245294, -0.0263719242, 0.0477569774, 0.0406795479, 0.0671771392, 0.0209180713, -0.0204416011, -0.0155303515, -0.0283087622, -0.0554504953, -0.0531431101, -0.0300683081, 0.0001818968, 0.0480898954, -0.0207639989, -0.0353206284, 0.0407797471, -0.0505064242, -0.0104661798, 0.00563872, -0.0115023898, 0.0403700434, 0.0084417546, 0.0338977836, -0.1161245704, -0.0579544194, 0.0035646807, -0.0146989953, -0.0371914171, 0.011346316, 0.0261679683, 0.0307528079, -0.0524302274, -0.0210747365, -0.0318345763, -0.0002587953, 0.0308944415, 0.0781358406, 0.0330410153, 0.0105118016, -0.0047556129, 0.0010235051, -0.0046452088, -0.0063576791, -0.0075774421, -0.0081685642, 0.0077220416, 0.008475991, -0.0116480049, 0.02095473, -0.0762035251, -0.0510250814, 0.0342133157, -0.0179759096, -0.0101002492, 0.0166160539, 2.594e-7, 0.0003502619, 0.0398030914, -0.0122461664, -0.0265069269, 0.0274798479, 0.0064163501, -0.0270558968, 0.0082319975, 0.0034469489, -0.0370843038, 0.0388249606, 0.0676485598, 0.02259106, 0.0116182426, -0.076542221, 0.0127530228, -0.0356502905, 0.0353026092, -0.0427348614, -0.0189365689, 0.0088479426, 0.1371036321, 0.0277131386, -0.01143914, 0.0070052012, -0.0342745408, 0.0583703071, -0.0246982835, -0.0057565318, 0.0432410762, 0.0813886598, 0.0288815591, -0.0075421198, 0.0353326909, 0.0392597765, 0.0605632067, 0.0247668419, 0.0834068805, -0.0271388106, 0.0059519052, 0.0452902429, -0.0063681607, -0.0103707379, -0.0507871695, 0.0549176373, -0.0187286306, -0.0196174532, -0.0357718989, 0.0130406534, -0.0173921678, 0.0074855527, -0.0478493534, -0.0455265567, 0.014301382, 0.0116538219, -0.021354951, 0.0093869166, -0.0757474229, 0.0099748187, -0.0006192673, -0.0797348991, -0.0694376007, -0.0331698433, 0.0000428687, 0.0704825372, -0.0250928458, -0.0136629883, 2.5048065e-34, 0.0614622943, 0.0058537545, 0.0037061118, 0.0185681302, 0.0365827829, -0.0223699231, -0.0658717826, -0.00370358, 0.0059254295, -0.104552269, -0.0608532541 ]
http://arxiv.org/abs/2406.18200v1
http://arxiv.org/pdf/2406.18200v1
SEED: Accelerating Reasoning Tree Construction via Scheduled Speculative Decoding
2024-06-26T09:33:41
2024-06-26T09:33:41
Zhenglin Wang, Jialong Wu, Yilong Lai, Congzhi Zhang, Deyu Zhou
cs.CL
cs.CL
Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by surpassing the capabilities of chain-of-thought prompting, encouraging exploration of intermediate steps. However, such methods introduce significant inference latency due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SeeD, a novel and efficient inference framework to optimize runtime speed and GPU memory management concurrently. By employing a scheduled speculative execution, SeeD efficiently handles multiple iterations for the thought generation and the state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate superior speedup performance of SeeD, providing a viable path for batched inference in training-free speculative decoding.
SEED: Accelerating Reasoning Tree Construction via Scheduled Speculative Decoding: Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by surpassing the capabilities of chain-of-thought prompting, encouraging exploration of intermediate steps. However, such methods introduce significant inference latency due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SeeD, a novel and efficient inference framework to optimize runtime speed and GPU memory management concurrently. By employing a scheduled speculative execution, SeeD efficiently handles multiple iterations for the thought generation and the state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate superior speedup performance of SeeD, providing a viable path for batched inference in training-free speculative decoding.
[ 0.0354654938, 0.0510721244, -0.0068561719, 0.0225629266, -0.0503505692, 0.0237316955, -0.0222291239, -0.0351965316, -0.0230378546, -0.0199615266, 0.022843441, -0.0732704401, 0.0074073495, 0.0252464078, -0.0335558541, -0.0356248803, 0.0295971539, -0.0753278807, 0.0339827426, -0.014014639, -0.0322943702, 0.0381735452, -0.0252163727, 0.0279686879, 0.0093247602, -0.042409569, -0.0601513572, 0.0074368254, 0.0013226286, -0.0441259891, -0.0472744964, 0.0332993232, 0.0249079131, 0.0474010594, 0.0000022405, -0.0308989082, -0.0207750294, 0.0009304957, -0.0386648886, 0.0051248963, 0.010631307, 0.0238156449, -0.0171062574, 0.0276917797, -0.0274770483, -0.0232524052, 0.0296349693, -0.0207456462, 0.0003242138, 0.0187464561, -0.0078034392, -0.0330652334, 0.0036381641, 0.0126662925, 0.0078195343, -0.0496255271, -0.0000479801, -0.002532006, -0.0199477095, -0.0183278266, -0.015297248, 0.0234605316, 0.0469563082, -0.0005803216, -0.031620618, 0.0382818058, 0.0116266478, -0.0492098629, -0.0294609312, 0.0125763947, -0.0262373146, -0.0448086187, -0.0174341463, -0.0030635609, -0.0436976328, -0.0598191209, -0.0087078512, 0.0909534097, -0.0708337203, 0.0112195611, 0.110860981, 0.00549964, 0.0239949785, 0.0151780434, -0.0535040759, 0.0977391601, 0.030429557, -0.0214071181, -0.0100816078, 0.0106749581, -0.0067158793, -0.072011672, 0.0244959481, 0.0396343321, -0.056662105, 0.0385169648, -0.0430814475, -0.024428241, -0.0167084783, -0.000915183, 0.0286746472, 0.0958460793, 0.0206337906, 0.0230981018, -0.025922846, -0.0076648281, -0.0154085169, 0.0146244606, -0.043661464, 0.0138744805, -0.0306143258, -0.0076129041, -0.0530067794, 0.0305239707, -0.0593413673, -0.0167388003, -0.0084401807, 0.0397028029, 0.0132408142, -0.0036282197, 0.0418676697, 0.0133456271, -0.0085140895, 0.0005996519, -0.0257341992, -0.0101209283, -0.0358718373, 0.0035934127, 0.0117494296, -0.0426542424, -0.0149098188, 0.0031718214, -0.0117886113, -0.0021192455, 0.001883234, 0.0299605764, -0.0491030924, -0.0163509995, -0.0552745089, -0.0334919803, 0.0266374089, -0.0731214732, 0.0096339211, -0.0444496833, 0.0243118331, 0.012381237, -0.0094335023, 0.02951416, 0.0105472822, 0.0320729092, -0.0002076197, 0.050593745, -0.0392726474, 0.0028295009, 0.0102434112, 0.0147782816, 0.0182496458, 0.0917361975, 0.0415563658, 0.0450843796, -0.0163252745, 0.0228268746, -0.0144644259, 0.038110707, 0.008691798, -0.0306310132, 0.0211167149, 0.0211484395, -0.0384862274, 0.0539594479, 0.0066811917, 0.0442332886, -0.0298840143, 0.0209482908, 0.0284357872, 0.060782034, -0.0370263867, 0.0330939777, 0.0719763562, 0.0008552099, -0.0130227804, -0.019096626, -0.025188867, 0.018920077, -0.0494221374, 0.0250493195, -0.0078576915, 0.0499410294, 0.0055129803, -0.0033142304, -0.0106402747, 0.0284087695, 0.017837096, 0.026598841, 0.0620571747, 0.0193004217, -0.0596191846, 0.0324816853, 0.0043548523, -0.0078580258, -0.0695226416, -0.0531733707, 0.030025389, 0.0810852423, 0.0400401056, 0.0164783765, -0.0428964943, 0.0290659238, -0.0145522738, 0.0387309082, -0.0118863042, -0.0228706487, 0.0195871964, -0.0456856154, -0.0088632405, 0.0422064625, 0.0036362465, 0.0452753603, 0.0443166085, -0.0041675903, -0.0262129325, -0.0069143367, -0.0253725071, 0.0164855905, 0.0054438235, -0.014418887, 0.010324914, 0.011000609, 0.0114092398, -0.0163291339, 0.0064716656, 0.0160092637, 0.0286042541, -0.0388046056, -0.0021490154, 0.0174844023, -0.0653780475, -0.0510732345, 0.0204747263, 0.0265294891, -0.0203040522, -0.0163278244, -0.0341631509, -0.0351330191, 0.0224104635, -0.0271657612, 0.0558167323, -0.0389909595, 0.0411423929, -0.0000901402, -0.028773658, -0.0053216978, 0.0753427744, -0.0469374843, -0.0197561625, 0.0436378606, 0.0368126296, 0.0026087561, -0.027722111, 0.0187793951, 0.0262601171, -0.0276333019, -0.026761882, -0.0296217371, 0.0064819329, -0.0291903578, 0.0196712725, -0.0753120631, 0.0095430696, -0.0143553512, 0.0272584297, 0.0122744776, 0.0039158161, -0.013677693, 0.0022760038, 0.0171502326, 0.0483831614, 0.0028481255, 0.0197659433, -0.0107024787, -0.0048283166, -0.0506768972, 0.0022017958, -0.0090552531, -0.0333744511, 0.0403551348, -0.0048478842, -0.0470557809, -0.054108344, 0.0533621646, 0.0011473916, 0.0251400471, -0.0030260619, -0.0006191307, 0.0308128949, -0.0341855735, 0.0270547252, 0.045551993, -0.0133518307, -0.0100064632, 0.0650354773, -0.0174514167, -0.0253543016, 0.0302108023, 0.0027720437, -0.0107152835, -0.0191915985, -0.027097838, -0.0340552144, 0.0896086246, 0.0088633774, 0.0060713915, -0.0605571531, 0.0112705128, -0.0005872685, -0.0553326532, -0.0260496158, -0.0253459439, 0.0049150367, 0.0803743973, -0.031551417, -0.0655768737, 0.0798257738, -0.0494332463, -0.0064603528, -0.0461708605, 0.0067058131, -0.0114878239, 0.0997706428, -0.0146207195, 0.0118483799, 0.0041659107, -0.026752973, 0.0013301763, 0.0329558104, 0.0093139708, -0.0318578966, 0.0170372259, -0.0558098815, 0.016984025, -0.0238276199, -0.0344593301, -0.043287497, -0.0034365579, -0.0678511411, -0.0129714301, 0.0788614601, -0.0336109921, 0.0192023236, 0.0144372303, -0.0366448089, 0.0593957119, -0.0008424907, -0.0225282814, 0.0033645416, -0.0123584941, 0.0357815139, -0.0212296396, 0.0308662355, -0.0073864427, 0.0682144165, 0.0601633079, 0.0549190938, -0.0265594982, -0.0671282187, 0.0246887058, 0.0229521636, 0.0224658884, 0.0102894502, 0.0413585529, 0.0296248253, -0.0912641734, -0.0596366487, 0.0893802494, 0.0451210253, -0.0514183939, -0.0033291739, -0.0365415514, 0.0728680715, 0.0065913028, 0.0550048798, -0.1082508639, -0.0398690887, 0.0412216261, 0.0049990108, -0.0272428412, -0.0516262539, -0.0285489932, -0.109439455, 0.046633359, -0.0118600596, 0.0075572454, 0.0168147385, -0.0163130611, -0.0455981605, 0.0365656018, -0.0581038892, 0.0357448049, 0.0262741093, 0.0250419863, 0.002826425, -0.0146738542, 0.0046231928, 0.0557205677, -0.0012502015, 0.0022488297, 0.0413780734, 0.0044374876, 0.057079874, -0.0498117916, -0.0243215486, -0.0049642278, 0.0090611866, -0.0632406771, -0.0052024438, -0.0228276514, 0.0262883939, -0.0212018639, 0.0252103116, -0.0477576405, -0.039111279, 0.0133975018, 0.1065455377, -0.0500647351, -0.0443797559, -0.0103776371, -0.0112192547, 0.0327602737, 0.0637418851, 0.000517758, -0.0205694754, -0.0545555986, -0.0546660498, -0.027665399, -0.094616808, -0.0101691848, -0.0023280096, 0.0723605305, 0.0161649603, -0.0190362651, 0.0079496745, 0.0578504279, 0.0520259328, -0.0144459335, 0.0026882438, 0.0073195347, 0.0085480707, -0.093016386, -0.0112318862, -0.0022848747, 0.0137906428, -0.0115675526, -0.0148687596, 0.0132057508, -0.0280852374, -0.0174722262, 0.0520231761, -0.1173035875, -0.0277787335, -0.0116883619, 0.0355687104, 0.0252025295, 0.0470933206, -0.0267692246, 0.0009426802, 0.0234661028, 0.003973674, -0.0191273, 0.0060556377, 0.004456121, -0.0089865234, 0.0168483518, 0.0370293893, 0.0003916537, 0.019261606, 0.0095801037, 0.0410290621, 0.0139566958, 0.0122431358, 0.0183703396, 0.0255453344, -0.0006722556, -0.0472573377, -0.0259061512, -0.0094073955, 0.0497105755, -0.0080041885, -0.0243640337, -0.0309587996, 0.0427131988, 0.0242260229, -0.0208730493, -0.0095110424, -0.0439488627, 0.0470012724, -0.0247314703, 0.0281768981, -0.0319024138, 0.0504425466, 0.0147349881, 0.0438921414, -0.0248406995, -0.0121092228, 0.006374374, -0.0146776522, 0.0552944206, -0.0198715236, 0.0592969321, 0.0253549777, -0.0043987315, -0.0332452878, -0.0095930612, -0.0520333573, -0.1053349376, -0.0176499188, -0.0288847834, 0.0504564643, 0.0294409283, 0.0124825304, -0.0403488278, 0.0134284412, 0.018908279, 0.0120477621, 0.0013134169, -0.052534245, 0.0378216654, 0.0223464761, -0.0148556, -0.0049811481, 0.022977259, 0.0062275426, 0.0422346443, 0.042474743, -0.0376212038, -0.0273940563, -0.04364888, -0.0960864723, 0.0354135409, -0.0148076778, -0.0039110635, -0.0430357642, -0.039536614, -0.0016297265, 0.0407940596, 0.0041778339, 0.0046844874, 0.0384410284, 0.057135053, 0.0105545688, -0.0493414328, 0.0009656748, 0.0300683752, -0.0113806846, 0.0034869448, -0.0446745865, -6.585178128e-33, -0.0040344954, -0.0302414671, 0.082805112, -0.0438257344, -0.0067120343, -0.0130111808, -0.0554789603, 0.0001925966, 0.0105496859, 0.0166755114, -0.0298636071, 0.0314815715, 0.0238643177, 0.0385697074, 0.0308806133, 0.0179636106, 0.0330889374, -0.018633252, -0.0136478292, 0.0192237999, 0.0371473134, 0.0394236445, 0.0102135707, 0.0054522911, -0.0437954925, 0.0215102937, 0.0201267879, 0.0184218287, -0.0514162146, 0.0189596005, -0.0354453065, -0.0055880817, 0.0261261109, -0.0201585293, -0.0197029542, 0.0135057587, -0.0868539587, -0.0077781328, 0.0403860807, -0.0073346896, -0.0723080635, -0.0028915666, -0.0474341698, 0.0421574935, 0.0216828827, -0.0403715931, -0.0491259247, -0.0206869654, -0.005217934, -0.042979762, -0.0646290481, 0.0255442876, 0.0247075632, 0.0098345084, 0.0616269074, -0.0073342593, -0.025859464, 0.0064259926, -0.0228351112, 0.0090176212, -0.0172783677, 0.0234931037, 0.0890944228, 0.019177258, 0.007915535, 0.033869341, 0.0229357183, -0.0090248939, -0.0676099136, 0.0481201746, 0.0059861061, 0.0375213027, -0.0203763545, 0.0022270118, 0.0795102417, 0.0225954261, -0.0613949932, -0.0180156995, 0.0321221389, 0.0384374708, 0.0621513017, 0.0161208808, 0.000364728, -0.0457076095, -0.0380910747, -0.0249921344, -0.0350447036, -0.0034123294, 0.0445899516, 0.0252676606, -0.0341814198, -0.0006657999, 0.0379904993, -0.0025583175, 0.0378165729, 0.0215727985, -0.0007055041, 0.0200661402, -0.0086509455, 0.057158757, -0.0137130599, -0.0303428434, 0.0124270907, -0.0069233142, 0.0233732145, 0.015688803, -0.0189747233, 0.0088893073, -0.0502163023, -0.0073583056, -0.0684273019, 0.0415291227, -0.0081782751, 0.0510402992, 0.053562291, 0.0243978184, 0.0056948671, 0.0217514113, -0.004036393, 0.0474097989, -0.0326474421, 0.0126713626, -0.0052415226, 0.0215108339, 0.0017029732, -0.009495154, -0.0364712067, -0.0271274932, 0.0150185637, -0.0968181267, -0.0130194267, 0.0801337808, 2.945e-7, 0.0146338353, 0.0481906645, 0.0128548369, 0.0200205334, 0.0273854863, -0.0108774789, 0.0287253205, -0.0339842178, 0.0459815413, -0.0397871286, 0.0042775576, 0.0299819261, 0.0447445288, 0.0401126444, -0.0883639529, -0.0402558595, -0.0294155423, 0.0116255637, -0.0472562686, -0.017981831, 0.0272278301, 0.0997704118, -0.0194462165, -0.0291214101, 0.0508529991, 0.0097397799, 0.0019833865, 0.0073274686, 0.0182254426, -0.0109333275, -0.0162550956, 0.0103566516, -0.0080748796, 0.0441857241, -0.0069612483, 0.013445708, -0.076822117, 0.0344676487, -0.0108196503, 0.0124373175, -0.0050403331, -0.04181014, -0.005485429, -0.03594587, 0.0594017208, -0.0327595659, -0.0172209702, 0.0076496592, 0.0238983482, -0.0016151076, -0.044869259, 0.0102809928, 0.0426220261, 0.0159646887, 0.0128362123, -0.0262213405, -0.0061019291, -0.0783160031, -0.0139021706, -0.0037185473, -0.0493115187, -0.0398908444, -0.0587601513, 0.0107925227, 0.0297335777, -0.0589091256, -0.0008661965, 3.197913785e-34, 0.0385197252, 0.0241025835, 0.0045110597, 0.0660147071, 0.0241723582, -0.0009011799, 0.0481540114, 0.0028632646, 0.001340832, -0.1067744046, -0.0392374955 ]
http://arxiv.org/abs/2406.18192v2
http://arxiv.org/pdf/2406.18192v2
Methodology of Adapting Large English Language Models for Specific Cultural Contexts
2024-06-26T09:16:08
2024-06-27T02:17:19
Wenjing Zhang, Siqi Xiao, Xuejiao Lei, Ning Wang, Huazheng Zhang, Meijuan An, Bikun Yang, Zhaoxiang Liu, Kai Wang, Shiguo Lian
cs.CL
cs.CL, cs.AI
The rapid growth of large language models(LLMs) has emerged as a prominent trend in the field of artificial intelligence. However, current state-of-the-art LLMs are predominantly based on English. They encounter limitations when directly applied to tasks in specific cultural domains, due to deficiencies in domain-specific knowledge and misunderstandings caused by differences in cultural values. To address this challenge, our paper proposes a rapid adaptation method for large models in specific cultural contexts, which leverages instruction-tuning based on specific cultural knowledge and safety values data. Taking Chinese as the specific cultural context and utilizing the LLaMA3-8B as the experimental English LLM, the evaluation results demonstrate that the adapted LLM significantly enhances its capabilities in domain-specific knowledge and adaptability to safety values, while maintaining its original expertise advantages.
Methodology of Adapting Large English Language Models for Specific Cultural Contexts: The rapid growth of large language models(LLMs) has emerged as a prominent trend in the field of artificial intelligence. However, current state-of-the-art LLMs are predominantly based on English. They encounter limitations when directly applied to tasks in specific cultural domains, due to deficiencies in domain-specific knowledge and misunderstandings caused by differences in cultural values. To address this challenge, our paper proposes a rapid adaptation method for large models in specific cultural contexts, which leverages instruction-tuning based on specific cultural knowledge and safety values data. Taking Chinese as the specific cultural context and utilizing the LLaMA3-8B as the experimental English LLM, the evaluation results demonstrate that the adapted LLM significantly enhances its capabilities in domain-specific knowledge and adaptability to safety values, while maintaining its original expertise advantages.
[ 0.0487638041, 0.0087474547, -0.0345802978, 0.0282128938, -0.0124724004, 0.0337785892, 0.0202198103, -0.0312877223, -0.003225876, -0.0309516396, -0.0678492561, -0.0503319614, -0.0193364136, 0.0127135087, 0.0156283434, -0.0427160747, 0.0509420298, -0.031619519, 0.0098941531, -0.0128660435, 0.0186944269, 0.0082448311, -0.0137929143, 0.0329530053, -0.0002811034, 0.0028183432, 0.0165024288, 0.016283825, 0.0433889143, -0.0372249223, 0.0344975889, 0.0168953631, 0.0346769579, -0.0060303151, 0.0000021322, -0.0609787479, 0.0091973217, -0.0431702696, 0.0089099156, 0.0286390185, 0.0781809613, -0.0123516135, -0.052639097, 0.031543538, -0.0454740338, 0.0114461193, 0.034950234, 0.0416009091, 0.0333890431, 0.0790389851, 0.004593567, -0.0221739989, -0.0046797963, 0.0083180936, -0.0352037251, -0.0540587232, 0.0343836471, 0.0661805943, 0.0448131561, -0.0034323784, -0.0549866818, 0.0267533548, 0.0286667496, -0.0008695302, -0.044341363, -0.002199227, -0.0269653294, 0.0133614466, -0.0280701444, 0.0514919199, -0.0141401561, -0.006948472, -0.0152603723, -0.0143973511, -0.0175625402, -0.0296152439, -0.034656167, 0.0295449141, 0.0087857787, 0.0080309594, 0.0559135489, -0.0202075504, 0.0264368784, 0.0395957641, 0.0009416438, 0.0810397938, -0.0085476525, -0.0185244195, -0.0010181367, -0.0243777782, 0.0019653912, -0.014359518, 0.0432407334, 0.0602970459, -0.0118093854, 0.0429399051, -0.0151819922, -0.0076061343, 0.0132991197, -0.0910767466, 0.0455760434, 0.0529214703, -0.014925668, 0.0124472557, -0.0091837477, 0.0518774018, 0.0029050962, 0.0296839327, -0.0745236054, -0.0000644847, -0.0577004403, 0.0037012433, -0.0190687291, -0.0035186002, -0.0105459457, -0.0032507142, -0.0768575817, 0.0349416621, -0.0041078422, -0.0130852945, 0.0058219763, -0.0216419324, -0.0365859233, -0.0078554861, -0.0124538131, 0.0379704088, -0.0373089649, 0.0331460014, -0.0329463929, -0.0402690247, 0.0410259888, 0.0184523929, -0.0309193507, -0.0164412856, 0.0162522923, 0.0650546923, -0.0282071158, 0.0033224893, -0.0861039385, 0.0455435216, -0.014333792, -0.0308556426, 0.0333670676, -0.0438063592, 0.0104431361, -0.0207705405, 0.015987305, -0.0569925383, 0.0103416182, 0.0075630588, -0.0079086451, 0.0221251547, 0.00443738, 0.0251221433, 0.0468783863, 0.0137918349, -0.0052380334, 0.1035771891, 0.0613263585, 0.019212354, 0.0411142223, -0.0271538608, -0.0404208302, 0.0595891364, -0.0229348801, 0.0128753334, -0.0512504652, -0.0063605625, -0.0152473236, -0.0033086862, 0.0205029361, 0.0340211391, 0.0246271752, 0.0307466164, 0.0412264168, 0.027037384, -0.0038991245, 0.0300440788, 0.0622730516, 0.0031716966, 0.0157425776, -0.014225808, -0.0561782308, -0.0024339776, -0.0610647798, -0.0063690511, -0.0577053763, 0.0633400157, -0.0306732375, -0.0492521822, -0.0268236417, -0.0059381509, 0.0418767147, 0.0169831477, 0.0302826818, -0.0182723477, 0.017140381, 0.0239119455, 0.0356340222, -0.0138929198, -0.0147911841, -0.0491078384, -0.007382819, 0.1109391376, 0.0050495523, -0.0131710554, -0.0602946915, 0.0250787772, 0.0201354623, 0.0304683913, 0.0570656359, -0.0260517839, -0.0120552527, -0.0576261654, -0.0208290089, 0.022657223, -0.0032596961, 0.0178455859, 0.0388595089, 0.032304883, 0.0052954238, -0.0009565879, -0.0767398104, 0.0125735644, -0.0777225271, -0.0267919637, -0.0204992965, 0.0109266071, 0.0281926431, 0.0276580621, -0.0024690835, -0.0161018763, -0.0040893294, -0.0380028524, -0.0079012383, 0.0259995591, -0.0567830727, 0.0315971896, 0.0477368571, -0.0237539187, -0.0076652449, 0.0244383495, -0.0433044359, -0.0279194582, -0.0109773241, -0.0299508441, -0.001528025, -0.0352509692, 0.0313100964, -0.0493255332, 0.0135229453, -0.0300258733, 0.0569858886, -0.0255944934, -0.0442756154, -0.0228738319, 0.0272636805, 0.0012434336, -0.0246515516, -0.0546169542, 0.0057818014, 0.00375978, 0.0543392822, 0.0147850048, 0.0695604458, -0.0431937315, -0.0155624459, -0.0639499649, -0.0109553616, -0.0181019846, 0.0236512646, 0.0171548985, -0.0234623272, -0.0455578342, -0.0085522216, -0.0563830696, 0.032370314, -0.0208649784, 0.0251712371, 0.0038202198, 0.00052062, -0.0439556353, 0.0066062114, -0.0090074074, -0.0231021848, -0.0076596765, -0.0151863648, -0.051868286, -0.0723800212, 0.0459602773, -0.0339309312, 0.015918931, -0.0002433568, -0.0674757287, 0.0004683269, 0.0288231634, 0.0844090432, 0.0631441176, -0.0093738567, 0.0079704085, 0.0762941316, -0.0019295488, -0.0188425742, 0.0120435031, 0.018354468, -0.0296311118, -0.032862056, -0.0651755705, -0.0227508415, 0.0581576005, -0.0319261365, -0.0064017614, -0.0504191741, -0.0068828692, -0.0632595569, -0.0492153354, -0.0034639298, -0.0190110933, -0.012364191, 0.033892747, -0.0181016345, -0.0241140816, 0.061023891, 0.0108571863, 0.0342975855, -0.0649540275, -0.0278323945, -0.0228970516, 0.0328441896, -0.0338361859, 0.0492187925, -0.0047916248, -0.0219256468, -0.041336149, -0.0031264294, -0.0072524641, -0.0525772721, 0.0493879095, -0.0703990906, -0.021202758, -0.0119926063, -0.0259003155, -0.0344000645, 0.0239338223, -0.0280693006, 0.043884825, 0.0048898882, 0.0041604633, -0.0275759865, 0.0194030404, 0.0372314006, 0.003650472, -0.013237345, 0.0400327146, 0.0027092621, -0.0586349629, 0.0518338196, 0.0207278132, 0.016531935, -0.032385271, 0.062439695, 0.0274013095, 0.0565815568, -0.0444189645, -0.0243367925, 0.031912636, 0.0252537448, -0.026320355, -0.013514067, 0.0562880002, 0.0587222464, -0.0720513985, 0.0064208158, 0.0747082829, 0.0078613376, -0.0201366823, 0.026932558, -0.0500389561, -0.0175531674, 0.0217642263, 0.0565866828, -0.0622702427, 0.0029797058, -0.0084910644, 0.0186736546, -0.0050214208, 0.0031243868, 0.0131736137, -0.1045574099, 0.0246228538, 0.0146045899, -0.0618755408, -0.0407117382, -0.0019920326, -0.0084319022, 0.0010303597, 0.0741092563, -0.0141348131, -0.0372320972, -0.0549631268, 0.0211615749, -0.0259274058, -0.0966583043, 0.0386305228, -0.0279727578, 0.075013414, 0.0481382199, -0.0001298261, -0.0251073223, -0.0187988263, 0.0031345144, 0.0187266786, 0.0203922782, -0.0214594062, -0.0172129516, -0.0003469748, 0.000159039, 0.0070807631, -0.0107830772, 0.0058250753, -0.0105057275, 0.0437816046, 0.0665932149, -0.0177125186, -0.0068546631, -0.0320287719, 0.0126926759, 0.004264602, 0.0517006442, 0.0166680608, -0.0786892995, 0.0100317607, -0.0266698431, -0.0695347711, -0.0358348116, -0.0553495735, 0.0642425567, -0.083173275, -0.0406853855, -0.0578732006, -0.035714101, 0.0113347033, -0.0050284979, 0.0710735321, -0.0186557304, 0.0369333476, 0.0080758324, -0.0431448407, 0.0099849608, -0.0287234224, 0.0310661774, -0.0500836, 0.0078976676, 0.0512215681, -0.0394024588, 0.0208498389, 0.0096998662, -0.0038733438, -0.0439675413, 0.0411341712, 0.0584176816, 0.0068575479, 0.0032111497, -0.0066818749, -0.0603301935, 0.0528406948, -0.0202168338, 0.0046319063, 0.0093300119, 0.0157484543, 0.0062599871, -0.0021767637, 0.0174902994, -0.0586686581, 0.059056405, 0.0415395498, 0.0553784482, -0.049902942, -0.013046233, -0.0029519626, 0.0487825945, -0.0317664221, -0.0270064939, -0.0132347178, -0.0351211764, 0.0389044918, -0.0216458291, -0.0230610538, -0.0012321114, -0.0199350286, 0.0219529141, -0.0201366302, 0.012718237, 0.005723068, 0.0104208654, -0.0133932959, -0.0256792251, 0.0767017081, 0.0032754764, 0.0894435123, -0.003703211, 0.0349517576, -0.0383348987, 0.0140871499, 0.0262266994, 0.031068692, 0.0103082489, 0.0946612954, -0.0165064, -0.0079647787, -0.0207259227, 0.0571096428, -0.0139219873, -0.0234078728, -0.0225644838, 0.0207598358, 0.0606912076, -0.0092642186, 0.050260298, -0.0234945994, 0.0005798966, -0.0325057395, -0.0029436392, 0.0138927773, -0.0429599024, 0.0342142917, 0.0135930981, 0.0096748257, -0.0326529741, 0.0429910421, 0.000650794, 0.0443672948, 0.0196228269, -0.0119463364, -0.024465343, -0.0303708538, 0.0718319416, -0.0224544331, -0.062914975, -0.0083129704, -0.0640112236, -0.0608413629, -0.0163717642, 0.0093599577, -0.0077510416, -0.0149569558, 0.0092504676, 0.0496008061, -0.0021295666, -0.0560503006, -0.0270734485, -0.0274016578, -0.0151280295, -0.000154355, 0.0377516635, -6.414639612e-33, -0.0284872856, -0.006168223, 0.0429061428, -0.0234261528, -0.0112560168, -0.0128507018, -0.002580499, 0.00551986, -0.0392948911, -0.0009517104, -0.0207312517, 0.0103406832, 0.0139303887, -0.003596816, 0.0183356386, 0.0238835644, 0.0079894876, -0.0004134249, 0.0065630581, 0.0070646838, 0.0269416012, 0.0321347266, 0.0807713345, -0.0795512944, 0.0081695141, 0.0236781538, -0.0031960765, 0.0162178893, 0.0004766023, 0.0841975808, -0.0045720115, 0.0485846065, 0.0346243978, -0.0380303152, -0.0219493788, -0.0035978835, -0.0609819479, -0.0302074943, 0.0384737179, 0.028893942, 0.0289346501, -0.0908687338, 0.0463808589, -0.0244954675, -0.0316511169, 0.0091175614, 0.0298009347, -0.0658900514, 0.018114537, -0.011832267, -0.0844318867, 0.0182058029, 0.0112034166, 0.0182156842, 0.0013665447, 0.0321502388, -0.0288726818, -0.0458606519, -0.0179111492, 0.0191883557, 0.047454007, 0.0193897523, 0.0935082585, 0.0292791408, -0.0232335348, 0.0222309958, -0.0074261082, 0.0224952232, 0.002612181, -0.0550089404, 0.0416679084, 0.0535464212, 0.0096658878, 0.0622687228, -0.0036391094, -0.0209447779, -0.1057281122, 0.0383655876, 0.0330999941, 0.0677592605, 0.0316358209, -0.0013552904, -0.0663186759, -0.0181820188, 0.0030304089, 0.064907752, -0.0155279981, -0.0610999092, 0.0430338047, -0.0171390548, -0.0516111255, -0.0161229596, 0.0089359377, -0.0527172387, 0.02240007, -0.0002816204, 0.0155666489, -0.0085501475, -0.0059902738, 0.0377584025, -0.066296801, -0.0651021898, -0.0197187997, -0.0173266195, 0.00898109, 0.0031378618, 0.0235028211, 0.0513819121, -0.0477464832, 0.009346419, -0.0093044285, -0.0180272721, 0.0437303632, 0.0034903367, 0.0254659113, -0.0001370406, -0.0056950073, -0.0025554553, 0.0638495684, -0.0384979323, -0.0192713886, -0.0592028983, -0.026556829, 0.0413481295, -0.0427856259, 0.0255056359, -0.0342699327, 0.0433257334, 0.0035787013, -0.0568870679, -0.0231776237, 0.0155721493, 2.829e-7, 0.0381940417, 0.0703518316, -0.0189448688, 0.0469423085, -0.0021074843, -0.0066562556, -0.0380369499, 0.0250387266, 0.0094071394, 0.0352203958, 0.0457617491, 0.0553602688, 0.0526965037, 0.0465505868, -0.0640946254, -0.0260886643, -0.0178987943, 0.0411097333, -0.0571559705, 0.0344644263, 0.0259469431, 0.0764719248, -0.0232370365, -0.0211126227, -0.0397134721, -0.0364174917, 0.0144396704, -0.0513212681, -0.0001815424, 0.0438352525, 0.0116404081, 0.0542072095, -0.0136164781, 0.0274689384, 0.0135881575, 0.0120457504, 0.0137170032, 0.1045833677, -0.0156542938, 0.0080911219, 0.0336046703, 0.0013162086, -0.0067879618, -0.0647002235, 0.0401264951, -0.0160223953, -0.020920163, 0.0134227788, 0.0026966401, -0.0046159001, -0.0024119129, -0.0410446376, 0.0060704681, 0.0013482054, 0.0261869002, -0.0062749269, -0.0396943465, -0.02031748, -0.0059783435, 0.0114874523, 0.0241224039, -0.0720782727, -0.0233936626, -0.0064000832, -0.0459658876, -0.0453471988, 0.0292004067, 2.096678305e-34, 0.0067808586, -0.0275115464, -0.0094441753, 0.0194263421, 0.0335492864, 0.0182070564, -0.0037183645, 0.0255726185, -0.0029979688, -0.0652800128, -0.0221727248 ]
http://arxiv.org/abs/2406.18187v1
http://arxiv.org/pdf/2406.18187v1
Selective Prompting Tuning for Personalized Conversations with LLMs
2024-06-26T09:03:52
2024-06-26T09:03:52
Qiushi Huang, Xubo Liu, Tom Ko, Bo Wu, Wenwu Wang, Yu Zhang, Lilian Tang
cs.CL
cs.CL, cs.AI, cs.LG
In conversational AI, personalizing dialogues with persona profiles and contextual understanding is essential. Despite large language models' (LLMs) improved response coherence, effective persona integration remains a challenge. In this work, we first study two common approaches for personalizing LLMs: textual prompting and direct fine-tuning. We observed that textual prompting often struggles to yield responses that are similar to the ground truths in datasets, while direct fine-tuning tends to produce repetitive or overly generic replies. To alleviate those issues, we propose \textbf{S}elective \textbf{P}rompt \textbf{T}uning (SPT), which softly prompts LLMs for personalized conversations in a selective way. Concretely, SPT initializes a set of soft prompts and uses a trainable dense retriever to adaptively select suitable soft prompts for LLMs according to different input contexts, where the prompt retriever is dynamically updated through feedback from the LLMs. Additionally, we propose context-prompt contrastive learning and prompt fusion learning to encourage the SPT to enhance the diversity of personalized conversations. Experiments on the CONVAI2 dataset demonstrate that SPT significantly enhances response diversity by up to 90\%, along with improvements in other critical performance indicators. Those results highlight the efficacy of SPT in fostering engaging and personalized dialogue generation. The SPT model code (https://github.com/hqsiswiliam/SPT) is publicly available for further exploration.
Selective Prompting Tuning for Personalized Conversations with LLMs: In conversational AI, personalizing dialogues with persona profiles and contextual understanding is essential. Despite large language models' (LLMs) improved response coherence, effective persona integration remains a challenge. In this work, we first study two common approaches for personalizing LLMs: textual prompting and direct fine-tuning. We observed that textual prompting often struggles to yield responses that are similar to the ground truths in datasets, while direct fine-tuning tends to produce repetitive or overly generic replies. To alleviate those issues, we propose \textbf{S}elective \textbf{P}rompt \textbf{T}uning (SPT), which softly prompts LLMs for personalized conversations in a selective way. Concretely, SPT initializes a set of soft prompts and uses a trainable dense retriever to adaptively select suitable soft prompts for LLMs according to different input contexts, where the prompt retriever is dynamically updated through feedback from the LLMs. Additionally, we propose context-prompt contrastive learning and prompt fusion learning to encourage the SPT to enhance the diversity of personalized conversations. Experiments on the CONVAI2 dataset demonstrate that SPT significantly enhances response diversity by up to 90\%, along with improvements in other critical performance indicators. Those results highlight the efficacy of SPT in fostering engaging and personalized dialogue generation. The SPT model code (https://github.com/hqsiswiliam/SPT) is publicly available for further exploration.
[ 0.0500583053, 0.0212575216, -0.0197476782, 0.0256850552, -0.0070617311, -0.0149587849, 0.0064090597, -0.0366082974, 0.007016093, -0.0210404005, -0.0409590825, -0.0258729644, 0.0129937753, 0.0873595774, 0.0118570244, -0.0756979138, 0.04371012, -0.030163398, 0.0405878611, -0.0094325459, 0.0051304526, 0.0219853725, -0.0130510796, 0.0523315147, -0.0100382697, -0.0338422433, -0.0446277298, 0.0323464051, 0.0313658901, -0.0252902824, 0.0628496781, 0.052706603, 0.0271462034, 0.013792349, 0.0000022065, -0.0532291122, -0.0101815695, -0.0321105048, -0.0095276106, -0.0231525581, 0.0236588959, 0.0438501872, 0.0110209137, 0.0367833562, -0.0521212555, 0.0407593735, 0.0289544631, 0.0539678633, 0.0585264266, 0.0658017918, -0.0085043255, -0.0072444989, 0.0010318647, -0.0282937363, 0.0018943422, -0.0198936705, -0.0309604313, -0.0558481254, 0.0130900899, 0.0421003811, 0.0049436362, 0.0136304526, 0.0237842873, -0.0165231843, 0.0357750654, 0.0532936454, -0.0241592843, -0.0287652072, -0.0323701985, 0.1107803658, 0.06659922, 0.009977608, -0.0059418799, -0.0082165524, 0.0319963954, -0.0457816683, -0.0391597897, 0.0177369192, -0.0071054422, 0.0128441509, -0.0066265385, 0.0434962325, -0.0134335971, 0.0229400862, -0.0292637926, 0.0179354902, 0.0080020223, -0.0236178972, -0.0083726197, 0.0005029961, 0.0105806608, -0.0730892196, -0.0077362163, 0.0384636894, -0.0970679, 0.0113626691, -0.0607115552, 0.0110460147, 0.0344091132, 0.0337642394, 0.0183429848, 0.0465203039, -0.0469133705, 0.0250630639, 0.0130815459, 0.02079257, 0.0179853998, 0.0034015733, -0.0182960611, -0.0158916265, -0.0342378765, -0.0147574581, -0.0235759635, 0.0360946842, -0.0567332879, -0.025747912, -0.0209047087, 0.0591608137, -0.0075931875, 0.0251239538, -0.0080783395, 0.0282184724, -0.0498014279, 0.0198452752, -0.0504636057, 0.0165973343, -0.0613574386, -0.0097436821, 0.0515337549, -0.0731366053, -0.0256236717, 0.0150834341, 0.0225033369, -0.0293519385, -0.0076620714, 0.0123837031, -0.0451511964, -0.0662580058, -0.0263495687, -0.0264565293, -0.0482954346, -0.0647304356, -0.037561588, -0.0340869129, 0.0365883708, 0.0071463711, -0.0112366285, -0.0274900366, 0.0063412311, 0.0039883573, -0.0112712262, 0.039162375, -0.026352169, 0.0046745879, 0.0317539535, 0.0170403235, -0.0306389406, 0.0411204249, 0.0119506093, 0.0255143214, 0.0187126137, 0.0122685824, 0.0055581755, -0.0045539234, 0.0141285788, 0.0113692731, 0.0100700203, 0.0187771637, 0.0000517041, 0.0057746326, 0.0458600596, 0.0617157742, -0.0303562265, -0.0004793288, 0.0766466483, 0.0250251461, -0.0189002156, 0.035728097, 0.0491745397, 0.0499352142, -0.01261207, 0.0305860732, -0.0377117544, 0.0443282127, -0.0272572059, -0.0022658312, 0.0686781108, 0.0422374569, -0.0417566523, -0.0514472239, 0.04864401, -0.0125608779, 0.0243561547, 0.0299761929, -0.005818862, 0.0156630892, -0.0608827509, 0.0350901745, 0.0298433173, -0.0688881651, -0.0281620044, -0.0207837597, 0.0635356605, 0.0922965556, 0.033645086, -0.0111608161, -0.0440065525, -0.0218222551, -0.0076848678, 0.0576416105, 0.05033838, -0.0085545089, 0.057454925, -0.0540785603, -0.0149724102, -0.0547255278, -0.0162811819, 0.0519088991, 0.0118821738, 0.0043664202, 0.0115428632, 0.0383250751, 0.0141926324, 0.0029528276, -0.0466968007, -0.0136679476, 0.0402872935, 0.0185292047, -0.0007316939, 0.0477703512, 0.0067164944, -0.0046670316, 0.026615655, -0.0151276952, -0.0009247129, 0.0122920731, -0.0905090049, -0.0376503877, 0.0200033095, 0.0080265161, -0.0332505107, -0.0153322006, -0.0124961948, -0.0558766648, 0.0229741037, -0.0191031937, 0.0156317856, -0.0489793606, 0.0103365425, 0.0278109089, 0.0338664353, -0.0313537531, 0.0600391887, 0.0106158461, -0.0189703945, 0.065878123, 0.0593684241, -0.0081430916, -0.0804612264, -0.0083691906, -0.0081757708, 0.0049345461, -0.0163548738, 0.0243438352, 0.0224538278, -0.0638365, -0.0652182326, -0.1021725833, 0.0123640895, -0.0170706082, -0.0339328051, 0.0093451459, 0.0521500185, -0.0197650529, -0.0177529734, -0.0739861503, -0.0163712595, 0.0396099575, 0.0060334429, 0.0500917286, 0.0186502915, -0.0024957671, -0.0498139821, 0.0327345654, 0.0226027649, 0.0135741513, -0.0305419266, 0.0088386126, -0.0378922075, 0.024929041, -0.0386997312, 0.0230397005, -0.0352472179, -0.019959854, 0.0390972756, -0.005676385, 0.076609917, 0.0624817237, 0.0473429151, 0.0291082822, -0.0023691184, -0.0134846922, -0.0644976571, 0.0530871339, -0.0169225205, 0.0020714437, -0.0117024602, -0.0223026369, 0.0316434763, 0.0202882644, 0.0088333022, 0.0110727735, -0.0869500935, 0.0219460148, 0.0463733859, -0.0474449061, 0.0080328444, -0.0015811982, -0.02369681, 0.0366048217, -0.0250125062, -0.0395679586, 0.071566999, -0.0204084329, -0.0322715603, -0.0395332389, -0.0272199251, -0.0504547469, 0.0190425366, -0.0301941764, 0.0128811356, 0.0009955246, -0.0111212377, -0.0490704775, -0.0379461236, -0.0191461984, -0.0220277365, 0.0635807142, -0.0258016624, -0.0136383632, 0.0092428215, -0.0699359402, -0.0794145465, -0.0456900783, -0.0044659232, 0.0059674238, 0.0218622331, -0.0272479076, -0.0024489795, -0.0367942899, 0.0248775762, 0.0298744049, -0.0465125144, -0.0090081692, -0.0393127352, 0.0260851383, 0.066756174, 0.0529071167, 0.0171753448, -0.0364244953, 0.0618714765, 0.0402470045, -0.0051135868, -0.0237813164, -0.0265218485, 0.0059071672, 0.0267640594, -0.0030839781, -0.0206391495, 0.0251925681, 0.0850377232, -0.0088101355, -0.0452991799, 0.1321919858, 0.0185516961, -0.0627821237, -0.0002844902, -0.0234763548, 0.0662384331, 0.0143876718, 0.0305117313, -0.0488325506, -0.0326850079, -0.0098976344, 0.0312665291, -0.0125268633, -0.0194152985, -0.0326892883, -0.1260169744, 0.0288827587, 0.0045424146, -0.0229863357, -0.0102193337, 0.0005609798, 0.0146646546, 0.0138519667, 0.0352212861, -0.0094026718, 0.0106590996, 0.0142620113, 0.0008929512, 0.0251083225, -0.0409686938, -0.0186053123, -0.0269637089, 0.0725773424, 0.0696356744, 0.0379082076, 0.0160252098, -0.0310566779, 0.0019350632, -0.0118329311, 0.0122010456, -0.0249929242, 0.0422822051, -0.0150987795, 0.0127743417, -0.0186154675, 0.0536441319, -0.0112794042, -0.0352090932, 0.029832501, 0.0048944312, -0.0454616025, 0.0192734096, -0.0159312822, 0.0319140665, 0.0457454398, 0.0832269415, 0.0431948639, -0.0550202392, -0.0064531597, -0.0274092238, 0.0195046999, -0.0507918559, 0.0240167268, -0.0015488139, 0.0059729205, 0.0072355168, -0.0627736449, -0.0649094805, 0.0205858201, 0.0052840714, -0.0062138727, 0.0028584495, -0.0518256649, 0.0150418431, 0.001844813, 0.0324627534, -0.0302474871, 0.0223542638, -0.0777561441, -0.038087476, 0.0233061742, -0.098236993, 0.0018005988, 0.0087027894, -0.0457471758, -0.0744787604, -0.0300597399, 0.0563106164, 0.0189220179, 0.080978252, 0.0048497142, -0.0482300147, 0.0108197713, 0.0333220437, -0.0174278561, 0.0137282768, 0.0164407417, 0.0732119232, -0.0080859279, 0.031057125, -0.031261351, -0.0040283613, 0.0021264525, 0.0334312916, -0.0148949809, -0.1140116826, 0.0054340051, -0.0050997837, 0.0257018711, 0.0051991898, 0.0104611292, -0.0175364241, 0.0027526999, -0.0043093087, -0.0232351832, -0.0523472913, -0.0164111461, -0.011706871, 0.0006299538, 0.0004132171, -0.0318611301, -0.0441411808, 0.0648790598, -0.0548024252, -0.016132243, 0.0162117016, 0.0012405777, 0.0067187948, -0.0148279546, -0.0172082465, 0.0427991115, 0.0085482914, 0.0302774496, 0.0054723793, 0.0940431654, 0.0173842087, -0.0269357674, -0.0057524261, 0.0520829335, -0.0547463782, -0.0129825603, 0.0215655323, 0.0469018929, 0.0104316762, 0.0235319864, 0.0455044694, -0.0130971111, 0.0259081852, -0.0092713414, 0.0421478562, 0.0325834677, -0.0267024618, 0.0328197032, 0.0140980249, -0.03343229, 0.0092331339, 0.0544440895, -0.0623037964, 0.0052710064, 0.0123750726, 0.0379911102, -0.0508787706, -0.0048216041, 0.0265309308, 0.0349849351, -0.003119101, 0.0181582235, -0.0363013707, -0.0163747799, 0.0484138057, 0.0012391109, -0.0249810331, -0.020411158, 0.0298788734, 0.0401833542, -0.0182152484, -0.0200916957, -0.0433983542, 0.011700388, -0.0011975217, -0.04072069, -0.038121134, -6.808745399e-33, -0.0377963595, -0.0175918862, 0.0110385669, -0.0128686614, -0.0514065847, -0.0208675805, -0.064810887, 0.0016460365, -0.0160650574, -0.0243012961, -0.0503801443, 0.0087796757, 0.0060340306, 0.0476805642, -0.0293796789, -0.0242816675, 0.0668216497, 0.0375631079, -0.0252116416, 0.0682731122, 0.0367793739, 0.0317889005, -0.0095267482, -0.0646489263, 0.0296968874, -0.0158154536, -0.0108725764, 0.0322619453, 0.0063782753, 0.0546110421, -0.0230959672, 0.0225801151, 0.0059534251, 0.0130769154, -0.0184736997, 0.0651552379, -0.0666064546, -0.0453606844, 0.0195558742, -0.0168189276, -0.0525414459, -0.07301341, -0.0043459749, -0.0250335261, 0.0039509661, 0.0081951423, 0.021647647, -0.011056928, -0.0168479979, -0.0044197896, -0.0379121676, 0.0047106426, -0.03314263, -0.0419090278, -0.0263392124, 0.0522079542, 0.0169461612, 0.049487412, -0.0349814706, 0.0411389321, 0.0078060841, -0.0119494023, 0.0130343055, 0.0330922082, -0.014423768, 0.0430072732, 0.034418121, 0.0338959992, -0.0119203366, -0.0336555466, -0.0069317049, 0.0391134061, -0.0011268733, 0.0808379874, -0.0102686426, -0.0065932022, -0.0065750368, 0.0596915856, 0.0343309231, -0.0194625277, 0.0378691517, 0.0057217255, -0.0327898562, -0.0124142375, -0.0227262694, -0.0198086742, -0.0306340829, -0.0181535464, 0.0281460229, 0.0260561276, -0.0413177684, -0.0522498451, 0.0231094118, -0.0191813223, 0.0024372304, 0.0336344503, -0.0467473567, -0.0111434683, 0.0025365707, -0.0042825476, -0.0467009656, -0.0319040604, 0.0664693266, -0.0108544342, -0.0100496532, 0.0451061055, 0.0100108953, 0.0132275987, -0.0397389755, 0.0056927563, -0.0287674405, 0.0828379318, -0.0423050448, 0.031119192, -0.0003974744, 0.0118748415, -0.0346007757, 0.0636715963, -0.0009663171, 0.0089799045, -0.0604893304, -0.0580577031, 0.0047501852, -0.0015806251, -0.0368181765, 0.011400261, -0.0101245502, 0.0085146623, 0.0188071281, -0.031781666, -0.0209026448, -0.002106871, 2.834e-7, -0.0405111164, 0.0084054507, 0.0131136496, 0.0294291656, 0.0166865755, 0.0587928779, -0.0036119234, 0.0323891491, 0.0224750005, -0.0249835402, 0.0108956825, 0.0468584113, 0.0606759638, -0.0016919656, -0.0969343483, -0.0145924268, -0.0631455779, -0.007556173, -0.0526645184, 0.0120944325, 0.090160951, 0.0523529276, 0.0187681876, -0.0095500527, 0.0185505487, 0.022593474, 0.0423945002, -0.0326923616, -0.0184791107, 0.0171892904, 0.0264772009, -0.0251647998, 0.0076975366, -0.0099359006, 0.0048029814, 0.0453249738, -0.0266246237, 0.0624660477, 0.0230229963, -0.0727293715, 0.020126082, 0.031452205, 0.0135462293, -0.0399270393, 0.0629778355, -0.0174879562, -0.0231803488, -0.0144820306, -0.017209379, -0.0220165569, 0.0188586861, -0.0297382828, -0.0436571836, 0.0307767298, 0.0025039273, -0.0050873547, -0.0077853939, -0.0827826709, -0.0022119544, -0.0165468547, -0.0609086193, -0.0518188663, -0.0464620702, 0.017195167, 0.0080833407, -0.0369611755, 0.0303868931, 2.766070451e-34, 0.0193891712, -0.0062171216, 0.0285308305, 0.0291716214, 0.0203739461, -0.0043938798, -0.015324356, 0.0024360137, 0.0584366582, -0.0754018277, -0.0323576704 ]
http://arxiv.org/abs/2406.18173v1
http://arxiv.org/pdf/2406.18173v1
UIO-LLMs: Unbiased Incremental Optimization for Long-Context LLMs
2024-06-26T08:44:36
2024-06-26T08:44:36
Wenhao Li, Mingbao Lin, Yunshan Zhong, Shuicheng Yan, Rongrong Ji
cs.CL
cs.CL
Managing long texts is challenging for large language models (LLMs) due to limited context window sizes. This study introduces UIO-LLMs, an unbiased incremental optimization approach for memory-enhanced transformers under long-context settings. We initially conceptualize the process as a streamlined encoder-decoder framework where the weights-shared encoder and decoder respectively encapsulate a context segment into memories and leverage these memories to predict outputs of the subsequent segment. Subsequently, by treating our memory-enhanced transformers as fully-connected recurrent neural networks (RNNs), we refine the training process using the Truncated Backpropagation Through Time (TBPTT) algorithm, which incorporates innovative incremental optimization techniques. These techniques not only diminish time complexity but also address the bias in gradient computation through an unbiased optimization process. UIO-LLMs successfully handle long context, such as extending the context window of Llama2-7b-chat from 4K to 100K tokens with minimal 2% additional parameters, while keeping the inference cost nearly linear as context length increases.
UIO-LLMs: Unbiased Incremental Optimization for Long-Context LLMs: Managing long texts is challenging for large language models (LLMs) due to limited context window sizes. This study introduces UIO-LLMs, an unbiased incremental optimization approach for memory-enhanced transformers under long-context settings. We initially conceptualize the process as a streamlined encoder-decoder framework where the weights-shared encoder and decoder respectively encapsulate a context segment into memories and leverage these memories to predict outputs of the subsequent segment. Subsequently, by treating our memory-enhanced transformers as fully-connected recurrent neural networks (RNNs), we refine the training process using the Truncated Backpropagation Through Time (TBPTT) algorithm, which incorporates innovative incremental optimization techniques. These techniques not only diminish time complexity but also address the bias in gradient computation through an unbiased optimization process. UIO-LLMs successfully handle long context, such as extending the context window of Llama2-7b-chat from 4K to 100K tokens with minimal 2% additional parameters, while keeping the inference cost nearly linear as context length increases.
[ 0.0280601308, 0.0749914795, 0.0205439851, 0.054225862, 0.0015904781, -0.0155159719, 0.002322193, 0.0034803101, -0.0405224226, -0.022989776, -0.0580163077, -0.0527066663, 0.0167626832, -0.0147719001, 0.0002527425, -0.0213235654, 0.0188757516, -0.0203882176, -0.0087832287, -0.0110421823, -0.0044388105, -0.0016040795, -0.0021093199, 0.0291456617, 0.0051781447, -0.0247077942, -0.0068145203, 0.0052919174, -0.0094455639, -0.0501038954, -0.0027868897, -0.0012244387, 0.0057817223, 0.0346052609, 0.0000021849, -0.0412902609, -0.0105900886, -0.0411196947, -0.0362410285, 0.0339358747, 0.0560850427, 0.0097251106, -0.0113727143, 0.0291725397, -0.008363585, 0.0132523514, 0.0400408953, 0.0379945636, 0.0439443439, 0.0724180788, 0.0067745368, -0.0332215056, 0.0363008194, -0.0609954707, 0.0099716112, -0.029962305, 0.0050140228, -0.0388131663, -0.0240674689, -0.0146326181, -0.0384721085, 0.0508292727, 0.0550210066, -0.0009911988, -0.0070280479, 0.0247415155, -0.0221316777, 0.0163935982, -0.0052294726, 0.0640810132, -0.0260266811, -0.0120964255, -0.0176829025, -0.0081956983, -0.0310660657, -0.0762604177, -0.0369640142, 0.0153080262, 0.0243118778, 0.04226055, -0.0028713206, 0.0202474017, 0.0543615893, 0.0144341039, -0.0508401431, 0.050608065, 0.0115731927, 0.0009463216, -0.0151173864, -0.0107094571, 0.0017717283, -0.0246730577, 0.0316150226, 0.0825055838, -0.0104565416, 0.0352807119, -0.0073862453, -0.02882291, 0.0016780739, -0.0504580662, -0.0163707975, 0.0663985834, 0.0034982236, 0.0057269051, -0.0076117045, 0.0217584707, -0.0493582338, -0.0133053996, -0.0380361862, -0.0210586675, -0.0356127173, -0.016486926, -0.0261367876, 0.0220919922, -0.036118079, 0.0043390142, -0.0636032745, 0.0397935845, 0.0384929702, -0.0416304581, 0.0506002381, 0.0087854797, -0.0548685826, -0.0056960266, -0.0179500114, 0.0397257172, 0.0056948918, -0.010550024, -0.0071365205, -0.0419052914, -0.0403227769, 0.0283131767, 0.0009843328, -0.0306249019, -0.0083138049, 0.092158176, -0.0172433201, -0.056373775, -0.0476643331, -0.0079313535, -0.0265982803, -0.0468407534, 0.0376405045, -0.0555233769, 0.0389192253, -0.0336407013, 0.0020961752, -0.0296851303, -0.002746552, 0.0270783957, -0.0051032053, 0.0439404137, 0.0141423522, -0.0443626605, 0.0543963872, 0.0077354047, 0.0563322082, 0.0643539429, 0.0679046959, 0.0382659249, -0.0078607025, 0.0235651564, -0.0152804814, 0.0557196103, 0.0525693521, 0.0009388179, -0.0599561036, -0.0182673465, 0.0024959545, 0.0440925434, 0.037989337, 0.0496855155, -0.0195730235, -0.0106680449, 0.0428744964, 0.0137216169, 0.0304053817, 0.018674871, 0.0300031696, -0.0190987382, -0.0096592726, -0.0168730356, -0.0263428427, 0.0457226373, 0.0278710146, 0.0222281758, 0.0380241387, 0.0378687344, -0.0482902341, -0.0430318266, -0.0038797352, -0.0391185023, 0.0114950268, 0.0036921392, 0.0416408516, 0.0195382796, -0.0663555712, 0.1119872481, 0.0094628492, -0.0144983362, 0.0050007189, -0.0426554419, -0.0138633279, 0.0537774339, 0.0422042161, -0.0189080518, -0.0327582657, -0.0224090014, -0.0050526708, 0.0524248108, -0.0476955585, 0.0076605389, 0.0407069512, -0.0682916939, -0.000656967, 0.0207579732, -0.0064991834, 0.0127068926, 0.0206483975, 0.0483275503, 0.021357175, -0.0084031299, 0.0325374454, 0.0341835618, -0.0414541513, -0.0333585739, 0.0307238493, 0.0062484024, 0.0534466207, 0.0211349037, 0.0336527005, 0.0000675356, 0.0287199728, -0.064007543, -0.0281811394, 0.0193073489, -0.132514149, -0.074737072, 0.0507188179, 0.0182467811, 0.0177382305, 0.0229415949, -0.0547191128, -0.0532808229, 0.0402306691, -0.0495455116, 0.0828348994, -0.0345681906, 0.017117206, 0.0064892615, 0.0345018692, -0.0080557829, 0.0195754822, -0.0518828928, -0.0196376238, 0.0211790595, 0.0085388394, -0.0143933902, -0.02479163, -0.0158156678, 0.0573389381, -0.0119662061, -0.0114923697, -0.0025933848, 0.0246143714, -0.0088440087, -0.0342069454, -0.1006442979, 0.0097334217, -0.0341025107, 0.040416602, 0.0115495753, 0.0351678915, 0.0258483849, -0.0093791159, -0.0058671231, 0.001079516, -0.0152521078, 0.0065583149, 0.0316941775, 0.0114117749, -0.03486811, -0.0357423313, -0.0363957211, 0.0249153115, 0.072160095, -0.0278904792, -0.0464623421, -0.0430680998, 0.0144223124, 0.0013783585, 0.0171790533, -0.0283927545, -0.0489260666, -0.0245963577, 0.0121644642, 0.0577353574, 0.0336653404, 0.0263107866, 0.0150338076, 0.0348326862, -0.0079633286, -0.0349650234, 0.0189247727, -0.0234441515, -0.0441543236, -0.0236668922, 0.0036005285, 0.0044499873, 0.0860493481, -0.0434860066, 0.0104021495, -0.0383184925, -0.0427306183, -0.0159022976, -0.0174201727, -0.0089358613, -0.0290462766, -0.0439621322, -0.0074100336, 0.0091825807, -0.0379261039, 0.0446117111, -0.0423423909, 0.0114907976, -0.0439131074, 0.0054276311, -0.0602804497, 0.0496767424, 0.0056849043, -0.0053215362, -0.0341612212, -0.0478726402, -0.0314735174, 0.0149678299, -0.0334976576, -0.0094757872, 0.0421848446, 0.0004852359, 0.0122247869, -0.0187217817, -0.0244729053, -0.0453190729, 0.0069006891, -0.0128559824, -0.0050697527, 0.0283623338, 0.0022029139, -0.0244807191, 0.0147776911, -0.0126412455, 0.0747984126, -0.0077269063, 0.0033171792, -0.0047559082, 0.0366991274, 0.0153545719, -0.0000038873, 0.0188188814, -0.0310025141, 0.0107481163, 0.0314691216, 0.0445155837, -0.004606443, -0.0274757091, -0.0172190927, 0.0023696036, -0.0104861641, 0.0078220041, 0.0129952747, 0.0332010202, -0.0711226836, -0.0126454849, 0.1057019159, 0.0250794087, -0.0997547507, -0.006448383, -0.0066885939, 0.0091864159, 0.0097981635, 0.009494368, -0.0594717003, -0.0415026248, 0.0246257037, 0.0164084379, 0.0208968166, -0.0202004332, -0.0408840477, -0.0933788344, -0.0246932227, 0.0159511715, -0.0486389771, -0.0013381991, -0.0553049445, -0.0541832894, -0.0053044101, 0.0510310866, 0.0456468649, -0.0182283707, -0.0202462431, 0.0182990059, -0.0508487411, -0.0294466615, 0.0130884359, 0.0018730013, 0.029248219, 0.045357503, 0.0239739828, 0.0227971748, -0.0153334644, 0.0777697861, -0.0410441905, 0.0035156226, -0.008926047, 0.0258648768, -0.0291668829, 0.0049389228, 0.0089792768, 0.0398429036, -0.0329847373, -0.0052700909, -0.0159039497, 0.0622683875, -0.0028063313, 0.0013514747, 0.000736563, -0.0433931574, 0.0003645717, 0.0450943895, 0.0267140213, -0.0456223711, -0.0356915146, 0.0011598467, -0.000960596, -0.0312814564, 0.0229318943, -0.00365512, -0.0033924906, -0.0365411378, -0.0712510794, -0.0296548884, -0.0018547827, -0.0001454076, -0.0087451953, 0.0048228125, 0.0155645236, 0.0225104075, -0.0819952264, -0.0053438479, -0.0483206213, 0.0031616976, -0.0667852089, 0.0482196808, 0.039251741, -0.0208499078, 0.0009574378, 0.050088685, -0.1214654669, -0.0513956025, 0.0173644796, 0.0153236249, 0.0114652375, 0.0699830055, 0.021597527, -0.0296560898, 0.092231907, 0.029285498, -0.0142847616, 0.0395918414, 0.0312633291, 0.0451043658, 0.0177666806, 0.0238173585, -0.0516572781, 0.0313829184, 0.0156726222, 0.0102493186, 0.0189894177, -0.0043217232, -0.0323723406, 0.0067201867, 0.027271349, -0.0190182179, 0.0163115729, -0.0050234911, 0.0757301673, -0.0113450922, 0.0473002121, -0.0365798511, 0.0348709188, 0.0114483265, -0.0173725877, -0.0047967923, -0.0078082918, 0.0499586612, -0.0281992648, 0.0269997586, 0.0427385941, -0.0473879911, -0.01401834, 0.0236005187, 0.0387635715, -0.0237853955, -0.0045589157, 0.0075248056, 0.0131586548, -0.048048228, 0.0237316154, 0.0259917174, 0.0283533223, -0.0687383637, 0.0309951734, -0.0348665155, 0.0118728885, -0.0377942882, 0.0129573904, 0.0439044908, 0.0182207245, 0.0306060389, -0.0045244624, 0.0216025449, -0.0244668126, 0.0352967866, -0.0155291408, -0.0217005853, 0.0375344455, -0.0031195029, -0.0277031865, 0.0337950289, 0.0818527266, 0.0047775963, 0.0540594608, -0.0059045646, 0.0462624952, -0.0236991066, -0.0135312257, -0.0220546834, 0.0218625814, 0.0055194693, -0.0279798731, 0.0186753739, -0.053539224, 0.0417992957, -0.0417505763, 0.0006791864, -0.0231139716, 0.045840662, 0.0273709651, 0.0072780754, 0.0092933923, 0.0207479112, -0.0109773427, 0.0094562247, -0.0232424997, -0.0135470694, -6.833942855e-33, -0.0215422269, -0.0211488456, 0.029010389, 0.0059967153, -0.0723928884, 0.004728415, -0.0491134934, -0.0390102826, 0.0043280548, -0.0319987684, -0.0521427877, 0.0190198384, 0.0119272266, 0.0750788823, -0.0081731007, -0.0353191495, 0.0716857538, -0.0102109453, 0.0028428244, 0.0435937308, 0.0600435212, 0.0521788038, 0.0637551844, -0.007183691, -0.0173193924, 0.0063246214, 0.0166585334, 0.0356938094, 0.0282152444, 0.0508696102, -0.0635078698, 0.0480303317, 0.0343121849, -0.0748113692, -0.0235124864, 0.0157267097, -0.1042399183, -0.0190763287, -0.0140058091, 0.0319421142, -0.0344082601, -0.0615659803, 0.0480738431, 0.0305771939, -0.0165409036, -0.0218040459, 0.0098481178, -0.0223919582, -0.0337020271, -0.0502972119, -0.0366136953, 0.0068704272, 0.0022482099, -0.0022773738, 0.0388114415, -0.0011042443, -0.0143777179, 0.0441300422, -0.0963580757, -0.023262877, -0.008661624, 0.0469252951, 0.0584180914, 0.0235859174, 0.0259436592, 0.0240054224, 0.0218105055, -0.013236952, -0.0465369672, -0.0235019531, 0.0348434933, -0.0197101068, 0.0187118631, -0.0077647488, 0.0691433847, -0.0679899678, -0.0868255422, -0.0378055684, 0.0323962905, 0.0593813062, 0.0428962484, 0.0294819959, -0.0002025013, -0.0263558105, -0.0181284659, 0.0225012042, -0.0127603142, -0.0241880212, 0.0209765844, 0.051447995, -0.0004696774, -0.0187803339, 0.0509025492, -0.0179864336, 0.0106660314, 0.0335749201, 0.0000612316, 0.0170939248, 0.0036896823, 0.0212681405, -0.0458541699, 0.0159149393, -0.0094949277, -0.0040200455, -0.0042734924, -0.0310661644, 0.0277104322, 0.0468514599, -0.0754106417, -0.0035124482, -0.0449951999, 0.042264495, -0.0038645288, 0.0508631319, 0.0482521579, 0.0023333351, -0.0167277846, 0.0817029104, -0.0016816206, -0.0009602156, -0.0046760566, -0.007732443, 0.0207776558, 0.0286640804, -0.0417257659, -0.0077570276, -0.002984138, 0.0223526936, 0.0231236834, -0.0827605724, -0.0091458634, 0.0317069776, 2.915e-7, 0.0169196464, 0.0502121113, 0.0591586754, -0.0169630405, 0.0557610504, -0.0128291445, 0.0223253779, 0.0119237872, -0.0125648333, 0.002267347, 0.0119344052, 0.0041282861, 0.0407220833, -0.0042786822, -0.1158615872, -0.0277470797, -0.0859728679, -0.0026464872, -0.0143908234, -0.0464100242, 0.0147252008, 0.1179645434, -0.0003435905, -0.0187835488, -0.0158107933, -0.0015199929, 0.0486925729, -0.0193624999, 0.0102018341, 0.0293467529, 0.0191060714, 0.0147894146, -0.0194540769, -0.0508392304, -0.0026832875, 0.0328957178, -0.0544339418, 0.0570924282, 0.0073690135, -0.0789256468, 0.0374829918, 0.0314112641, -0.0348267406, 0.0030172584, 0.0536116958, -0.0760347247, -0.0317183807, -0.0256298073, 0.0118404292, -0.0070610368, 0.0123339696, 0.0163386203, 0.0090026008, 0.0296682939, -0.0043629156, 0.0020037415, -0.0115712686, -0.1241470128, -0.0137801701, -0.0252567269, -0.0672275499, -0.0295107625, -0.0733738244, -0.0007250933, 0.034181565, -0.0518792719, 0.0037921688, 2.484701873e-34, 0.0457410999, -0.0410543233, -0.0060923439, 0.0552205704, -0.0097638043, -0.0259898417, 0.018123053, 0.014945766, 0.0164038744, -0.0690131932, -0.019673761 ]
http://arxiv.org/abs/2406.18164v1
http://arxiv.org/pdf/2406.18164v1
NeBuLa: A discourse aware Minecraft Builder
2024-06-26T08:24:44
2024-06-26T08:24:44
Akshay Chaturvedi, Kate Thompson, Nicholas Asher
cs.CL
cs.CL, cs.LG
When engaging in collaborative tasks, humans efficiently exploit the semantic structure of a conversation to optimize verbal and nonverbal interactions. But in recent "language to code" or "language to action" models, this information is lacking. We show how incorporating the prior discourse and nonlinguistic context of a conversation situated in a nonlinguistic environment can improve the "language to action" component of such interactions. We fine tune an LLM to predict actions based on prior context; our model, NeBuLa, doubles the net-action F1 score over the baseline on this task of Jayannavar et al.(2020). We also investigate our model's ability to construct shapes and understand location descriptions using a synthetic dataset.
NeBuLa: A discourse aware Minecraft Builder: When engaging in collaborative tasks, humans efficiently exploit the semantic structure of a conversation to optimize verbal and nonverbal interactions. But in recent "language to code" or "language to action" models, this information is lacking. We show how incorporating the prior discourse and nonlinguistic context of a conversation situated in a nonlinguistic environment can improve the "language to action" component of such interactions. We fine tune an LLM to predict actions based on prior context; our model, NeBuLa, doubles the net-action F1 score over the baseline on this task of Jayannavar et al.(2020). We also investigate our model's ability to construct shapes and understand location descriptions using a synthetic dataset.
[ 0.0259588063, 0.0150331371, -0.0227043126, 0.0340230167, -0.0188784413, 0.0121327015, -0.0263096411, 0.0077232416, -0.0104590412, -0.046545364, -0.0513316356, -0.0216241106, 0.0041288105, 0.0504480638, 0.0240012426, -0.0521865599, 0.0186010301, -0.0547681451, -0.0077897091, 0.0201267041, -0.0144666741, 0.0003629318, -0.0274609774, 0.0552715808, -0.0143377036, 0.0038611232, -0.0727086216, 0.0116737839, 0.0033240705, -0.0330320857, -0.0082837744, 0.0433370918, 0.0000533291, -0.0151535962, 0.0000019334, -0.0170111824, 0.0866148993, -0.0436124615, 0.0133151421, -0.0644909963, -0.0197449531, 0.0069590965, -0.0017451617, 0.019171454, -0.0665074214, 0.017339563, 0.049430564, 0.0327208228, 0.0241597332, 0.0435475968, -0.0160900261, 0.0157142896, -0.0324141458, -0.020224968, 0.0201855432, -0.027737774, -0.0236148499, 0.0546612255, 0.0327598788, 0.0619267598, -0.034684021, 0.0376159549, 0.0103980806, -0.0374509245, 0.0353664197, 0.0492265224, -0.0057705445, -0.0251773223, 0.0156173389, 0.0923326835, 0.0367592126, -0.0160908066, 0.0480116755, 0.0048812851, 0.0385922231, -0.0005208641, -0.0346556939, 0.0093773687, -0.0130481739, -0.0102844955, -0.0100749228, -0.0035073219, 0.0093379598, 0.0006347477, 0.080176577, 0.0614098534, -0.028490914, 0.0259139519, -0.0555605516, 0.0062777307, 0.0111325718, -0.0798280835, -0.0208367631, 0.0370585173, -0.0727993697, 0.0193223022, -0.0267847851, 0.062353123, 0.0047765933, -0.0532790944, 0.0464234054, 0.0302326418, 0.0315959044, 0.0560717024, -0.0093099307, 0.0035062975, 0.0152349835, 0.0155612146, -0.0481830277, 0.0086547872, -0.0461395644, -0.0120546063, -0.0096613774, 0.0255782492, -0.094097659, -0.0476941243, -0.002674612, 0.0587033927, -0.0131861931, 0.0096454015, 0.0462095141, 0.0309609976, -0.0187890008, -0.0007809994, -0.0282330867, -0.0354283862, -0.009990124, 0.0309444275, 0.0112610469, -0.0634799898, -0.0114451377, 0.0156825669, 0.0115216738, 0.0034639658, -0.0112330941, 0.0862566158, -0.0262229014, -0.0161515027, 0.0051846998, -0.0252840035, -0.0107000638, -0.0627895668, 0.0314172879, -0.0652638748, 0.0071224226, -0.0004336429, -0.0118625974, -0.0119027505, 0.0179421026, -0.0086865313, 0.0084797507, 0.0704393834, -0.00807516, 0.0312436838, 0.0540240854, 0.0392556265, 0.0230191238, 0.063743189, 0.0537944064, 0.0213642046, -0.0081800865, 0.0230515655, 0.0085702315, 0.0466937274, -0.0416680872, -0.0093264654, 0.0327372402, 0.0039463523, -0.0246056933, 0.0162918679, 0.0250888616, -0.0101925097, -0.0268653557, 0.0051758382, 0.0717036948, 0.0850325003, -0.0073596146, 0.0488206558, 0.0180596281, 0.0318222269, -0.0460427999, -0.0508333929, -0.0228112787, 0.0762284622, -0.0037705826, -0.0228115208, 0.0071328618, 0.0063789254, -0.0472449921, -0.0572923236, 0.0116883758, -0.031540934, 0.0616559908, -0.0150432289, 0.0049193399, -0.0023902415, -0.0534666777, 0.0219468903, -0.0066304109, -0.0623730086, 0.0061193909, -0.0201013926, 0.0269858036, 0.056733612, -0.0059045046, -0.0125357891, -0.0291688591, -0.0007780195, 0.0101799006, 0.051950369, -0.0023532638, -0.0299239419, 0.0491971113, -0.0070037008, -0.0270932559, -0.017198449, 0.01322289, 0.0584243648, 0.0248468742, -0.0282095596, -0.0034000925, 0.0102593442, 0.0064012902, 0.0121349795, -0.0126153249, -0.0123657212, -0.0231545642, -0.0065503796, 0.046442721, -0.0077255717, 0.0180649236, -0.0151114482, 0.0341393016, -0.0296006706, -0.0176670458, 0.0880244449, -0.0615674891, -0.0419535562, 0.017300833, -0.0246151965, -0.0153524028, 0.0107247448, 0.0112476973, -0.0324668847, 0.072301425, -0.0515796281, 0.0536085106, -0.0105177732, -0.0004067586, -0.0126653416, -0.0159978867, -0.0175463054, 0.0209877715, 0.0033110406, -0.0191645157, 0.0153553905, 0.009495168, 0.0219140239, -0.0407525115, 0.0372152403, 0.0553705581, -0.0088354107, 0.0206778403, 0.0305423941, 0.0202170722, -0.0908188596, -0.0517486148, -0.1188716516, -0.0682189092, -0.0081561683, -0.0012464867, 0.0281941555, -0.002610872, -0.0179829113, -0.0125669511, -0.0367999114, -0.0233181491, 0.0261268411, -0.0073959352, 0.0418589041, -0.0001136382, 0.0112078441, -0.0434223898, 0.016005585, 0.0019983493, 0.0190869346, 0.0324609093, -0.0379710793, -0.0537421219, 0.0721288249, -0.028092565, -0.0053918818, -0.0443353802, -0.0363465883, 0.0367716961, -0.0311932899, 0.0822129548, 0.0587734953, -0.0089183282, -0.0008423704, 0.0384610891, -0.0452970862, -0.08353962, 0.0150382258, -0.00980586, 0.0187190659, -0.0659104884, -0.0476666689, 0.0378629006, 0.0713028684, -0.0179148428, 0.0387118757, -0.0704848096, -0.0086191231, 0.0362477787, -0.0567453615, -0.0054204348, -0.0357205234, -0.0303813946, 0.0579398684, -0.027582569, -0.0338290408, 0.0213393029, -0.0295756962, -0.0512782633, 0.0137575418, 0.0002323229, -0.0181315374, 0.0715802535, 0.0155528961, -0.0040441472, 0.0215798039, -0.0335474797, -0.002743372, -0.0211988296, 0.0229348373, -0.0155472541, -0.0072915591, -0.0631863549, 0.0020020581, 0.0066907392, -0.0163863078, -0.0126785701, 0.0199728925, -0.0199689772, -0.0017957184, 0.0053854343, -0.0231105406, -0.0208545327, 0.0218033679, 0.0183530338, 0.0122567201, 0.0084232884, -0.0171347819, -0.0299363732, 0.0163033511, 0.031428922, 0.0715451464, -0.047662098, -0.0159218498, 0.0040873387, 0.0536723696, -0.0051292852, -0.0221201573, -0.0222153626, 0.017137453, 0.050794851, 0.0517010875, -0.0417625271, 0.0593030229, 0.0966956913, -0.0009295203, -0.0481537059, 0.0803551376, 0.0206806492, 0.0068742689, -0.007837995, -0.0318737514, 0.043572966, 0.0230714995, 0.0136959562, -0.0055576675, -0.0334118381, 0.0101074111, -0.0343941264, 0.0040431111, -0.0458697528, 0.0130742183, -0.1044024974, 0.0381452069, -0.000226602, -0.0085939169, -0.0078759482, -0.0320886858, 0.0257315822, 0.0208694935, -0.0480801836, -0.0189046524, 0.0089788111, -0.002297723, 0.0139563484, -0.046003785, -0.0111701665, 0.0243105888, -0.0113832699, 0.0832205489, 0.037832424, 0.0703856945, 0.0208150316, 0.0153283775, 0.0090128938, 0.0487445667, 0.0148370406, -0.0487708859, 0.0336026512, 0.0346041061, 0.0008946266, 0.0137550924, 0.0226771552, -0.0511474162, -0.0548985824, -0.0014386969, 0.0310520791, -0.0514876172, -0.0073201461, -0.0090633053, 0.0662631467, 0.041962605, 0.0041005923, 0.0185106229, -0.0198748875, -0.025104085, -0.0468027033, -0.0093358709, -0.0504428707, -0.0441817418, -0.0659492686, 0.0470726825, 0.0302196648, -0.0689495876, -0.0454836041, 0.0250200927, 0.0172688048, 0.0032085667, 0.0113367131, 0.0048284833, 0.0244020317, -0.030095201, 0.0199338198, 0.0039896341, 0.0253748894, -0.0307256524, -0.050177291, 0.0362311602, -0.0735198706, 0.0248049572, 0.0529674962, -0.052253969, -0.0666843131, -0.040396858, 0.0628422648, 0.0196156707, 0.0414965935, 0.0111188414, -0.0192767214, -0.0022856093, 0.012589342, -0.042627994, 0.0346102603, 0.021326486, 0.0593658015, 0.0021935056, 0.008833182, -0.003127038, 0.0138136577, 0.0194475595, 0.0452925228, -0.0867963284, -0.0691590309, -0.0111636929, -0.0284843091, 0.0195183195, -0.016072087, 0.005183355, -0.0090599684, -0.0171514805, -0.0177021977, 0.0206125658, -0.063097693, 0.0147065679, 0.002213683, -0.0032523246, 0.030190561, 0.0416631252, -0.0557699092, 0.0363584384, 0.0286834165, 0.0484906808, 0.0033709733, 0.011726588, 0.0056997808, 0.0468505882, 0.0095631685, -0.0010725434, -0.0014145295, 0.0499448068, 0.0122020049, 0.0397511981, 0.0115355337, -0.0270287376, 0.0051234621, 0.0488155074, -0.0537010618, -0.0210905131, -0.0141636766, 0.0451872535, 0.0498383082, -0.0130234025, 0.028269235, -0.0638965368, 0.0339550301, -0.0074903877, 0.0762799159, -0.0139869265, -0.0691176206, 0.050307706, 0.0418752357, -0.0139782829, -0.039358642, 0.0210861024, 0.0067849723, 0.0017109845, 0.0215516668, 0.0053289915, -0.0573600754, -0.0487175696, 0.0252987668, 0.0422879532, -0.0299519096, 0.0258804709, -0.0189999826, -0.0433847047, 0.0087564643, 0.0348299444, 0.0207997151, -0.0675514042, 0.0513571464, 0.0437257737, 0.0146075431, -0.0210958291, -0.0647637993, -0.0182446372, 0.0185903963, 0.0052356417, -0.0741414279, -5.407195403e-33, -0.0454487987, -0.0797492713, 0.0465302281, 0.0585192293, 0.0060784658, -0.0209127311, -0.0846685767, 0.0507871509, -0.0085046645, -0.0163512249, -0.0213876348, 0.0277716946, 0.0039535118, 0.0110250358, 0.0305455383, -0.0025831158, 0.0270815231, 0.0053959643, -0.0188019946, 0.023764953, -0.0003653252, -0.0178098772, 0.0277842749, -0.0726497546, 0.0112997638, 0.0134239066, -0.0310134869, 0.0078498628, 0.0152892508, 0.0052812602, -0.0098089539, 0.0012041157, 0.0210776106, 0.0236605424, -0.0101525616, 0.025700815, -0.0476903915, -0.0021115104, 0.027008554, -0.0053620399, -0.0490242653, -0.0302153733, -0.0126569644, 0.0103956433, -0.0493526123, -0.0273091178, 0.053943716, -0.0085505247, -0.0434193574, 0.012042488, -0.074404642, 0.0008452576, -0.0593650304, -0.0586174093, 0.0453546979, 0.0544902161, -0.0000432313, -0.0164679773, -0.0482631065, 0.1156168506, -0.0435643159, 0.032054171, 0.0566533841, 0.0592931211, 0.0361655913, 0.0785474032, 0.0340727977, 0.0384136327, -0.0091032833, 0.0110682417, 0.0194241554, 0.0704569295, 0.002414559, 0.0437293351, -0.0210762583, 0.0364274643, -0.0217992235, 0.0465634763, -0.0419788025, 0.0256907903, -0.0126577914, 0.033983089, 0.0031530857, -0.0298557561, -0.026351219, -0.0120451292, 0.0121910209, -0.0197676681, 0.0075054793, 0.0071920604, 0.023059763, -0.0269321091, 0.0393246897, -0.0259078238, -0.0055656768, -0.0105868122, -0.0173873063, 0.029829612, -0.0041741757, -0.0112022096, -0.0508239716, -0.0277180299, 0.0177315753, 0.047950536, 0.0242537186, 0.0196540691, -0.0286363866, 0.0210938398, -0.0336533189, -0.0275644101, -0.0471822917, 0.0134564014, -0.0243854765, 0.0099580018, -0.002771606, -0.0418754406, -0.0410858169, 0.0140557382, -0.0278346948, -0.019909339, -0.0521009192, -0.0302114822, -0.0398752391, 0.0052177408, -0.030315889, -0.0222618524, -0.0805057138, 0.0329360999, 0.0893747211, -0.0291206706, -0.0133240055, 0.0017040233, 2.743e-7, 0.0037601043, 0.0537012219, -0.0058697923, 0.0043826406, -0.0037826663, 0.0587310158, -0.0424325019, 0.0263822991, 0.0490759276, -0.0033717614, 0.0290016886, 0.0453805663, 0.039798975, 0.0000688719, -0.099302873, 0.0411374271, -0.0316381566, -0.0386790968, -0.0018495998, 0.0031358807, 0.0056847762, 0.0116096893, -0.0218076855, 0.0032668768, 0.0040078377, -0.0258173402, 0.0494803302, -0.0927618295, 0.005434901, 0.034358412, -0.0391715877, -0.0502115116, -0.0365785137, -0.010409629, -0.0287493952, -0.03538936, -0.0097294226, 0.0625589788, -0.0216340721, -0.0318627842, -0.0149580166, 0.0190066621, 0.0035507334, 0.0058937054, -0.0136735989, -0.0617585219, -0.0333511867, 0.045269344, 0.0040392093, 0.0187684521, -0.0172618814, 0.0075141992, -0.022229882, -0.022706721, -0.0003189448, -0.0287017636, 0.0223089401, -0.0441736877, -0.0511057191, -0.034016028, -0.0843211934, -0.0626287684, -0.0339092314, -0.0049875905, -0.0026983074, -0.0018352456, -0.0015571609, 1.880146506e-34, -0.0194314513, 0.0355227217, 0.0091595575, 0.0547467098, 0.0287367608, -0.0028990181, 0.0238878913, 0.0506781638, 0.0529468805, -0.0437218063, -0.0159883574 ]
http://arxiv.org/abs/2406.18139v1
http://arxiv.org/pdf/2406.18139v1
LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference
2024-06-26T07:44:24
2024-06-26T07:44:24
Zhongwei Wan, Ziang Wu, Che Liu, Jinfa Huang, Zhihong Zhu, Peng Jin, Longyue Wang, Li Yuan
cs.CL
cs.CL, cs.CV
Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs' KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge. In this work, we introduce LOOK-M, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. LOOK-M demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by 80% in some cases, it not only achieves up to 1.5x faster decoding but also maintains or even enhances performance across a variety of long context multimodal tasks.
LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference: Long-context Multimodal Large Language Models (MLLMs) demand substantial computational resources for inference as the growth of their multimodal Key-Value (KV) cache, in response to increasing input lengths, challenges memory and time efficiency. Unlike single-modality LLMs that manage only textual contexts, the KV cache of long-context MLLMs includes representations from multiple images with temporal and spatial relationships and related textual contexts. The predominance of image tokens means traditional optimizations for LLMs' KV caches are unsuitable for multimodal long-context settings, and no prior works have addressed this challenge. In this work, we introduce LOOK-M, a pioneering, fine-tuning-free approach that efficiently reduces the multimodal KV cache size while maintaining performance comparable to a full cache. We observe that during prompt prefill, the model prioritizes more textual attention over image features, and based on the multimodal interaction observation, a new proposed text-prior method is explored to compress the KV cache. Furthermore, to mitigate the degradation of image contextual information, we propose several compensatory strategies using KV pairs merging. LOOK-M demonstrates that with a significant reduction in KV Cache memory usage, such as reducing it by 80% in some cases, it not only achieves up to 1.5x faster decoding but also maintains or even enhances performance across a variety of long context multimodal tasks.
[ 0.009991671, 0.0543595962, 0.0114146871, 0.0829629824, -0.0264241043, 0.0240195449, 0.0264376495, -0.0032655182, 0.0000866324, -0.0640965253, -0.0341650173, -0.0585643202, 0.0338846222, -0.0099136587, 0.0348003656, -0.0277623888, 0.0494920611, -0.0713990256, -0.0205329116, -0.0023678041, 0.0007587709, 0.0171853509, 0.0027000762, 0.0219162181, -0.0342016295, -0.0284444168, -0.0481558479, 0.0304739308, 0.0028919289, -0.017171273, -0.0183640514, 0.0434551127, 0.0178792551, 0.0040499517, 0.0000020552, -0.0373674259, 0.0174767282, -0.01276429, -0.0159087386, 0.0338577405, 0.0674123392, -0.0251052566, -0.0212054737, -0.0234233476, 0.0150673343, -0.0358868167, 0.0791493207, 0.0126019055, 0.040133208, 0.0232983846, -0.015139983, -0.0412862152, -0.0156623498, -0.0232048966, 0.0488248765, -0.0775575861, 0.0379871204, -0.0262381025, -0.0283740032, -0.0119775971, -0.0312023442, 0.0697548985, 0.0293925703, 0.0044641709, -0.0028973096, 0.0612007938, 0.0109513896, -0.0295063071, -0.0225768872, 0.0613309182, -0.0061992309, -0.0042434493, -0.0000238474, 0.0105038946, -0.0044556796, 0.0046934094, -0.0304632206, 0.0308727175, -0.0288968906, -0.0154807186, -0.0040440657, 0.0030072788, 0.0185457543, 0.0361850746, -0.0464488231, 0.0016319121, 0.0469112992, 0.0117139099, -0.032124579, -0.0053660823, -0.0198945049, -0.0701267719, -0.0112497937, 0.0331298895, 0.0068565411, -0.0022015814, -0.0459666774, -0.0182896797, 0.0129929716, -0.0518518612, 0.0157446917, 0.0316101089, -0.065782927, 0.0684563592, -0.0232264716, 0.0219912902, -0.0519186556, -0.0272200014, -0.0456836298, -0.0270622689, 0.0124915056, -0.0124661662, -0.0042923484, -0.0214395802, -0.0190780312, -0.0335427523, 0.0072575402, 0.0343771987, 0.0172493979, -0.0163651388, -0.0213901624, 0.0203466862, -0.0135408677, -0.0016520341, 0.0040142667, 0.0684223995, -0.0091836825, -0.0167704858, -0.0401057117, -0.0480558686, -0.0125933634, 0.047607094, 0.0237185694, -0.0666304231, -0.0075034234, 0.0854361579, -0.0284350011, -0.0250281636, -0.0245213602, -0.0307978559, -0.0481191985, -0.0509391837, 0.0435101986, -0.0431113243, 0.0468150564, -0.0031967235, 0.0017309495, 0.0162799079, -0.0235018283, 0.0181132574, -0.0371139236, 0.0856496841, 0.0413317941, 0.0021617208, 0.0451150984, 0.0118583376, 0.0143933203, 0.0740369782, 0.0004494359, 0.0493462794, 0.0137005411, -0.0170429945, -0.0220355261, 0.070183441, 0.0158493184, 0.0209082384, -0.0146304108, 0.0012263656, 0.0654190481, 0.0068971859, 0.065435715, 0.0488078631, -0.0467589088, -0.0289804209, 0.0539051481, 0.0302204005, 0.0217373185, -0.0445727482, 0.0191920921, -0.0085253101, -0.018795481, -0.0019153183, -0.0734714121, 0.0808223784, -0.0286539309, 0.0252939258, 0.0560414158, 0.0644318014, -0.027418036, 0.0293882117, -0.0176004879, -0.0315338857, -0.0027610634, 0.0269033276, 0.0346768126, -0.0131280851, -0.0439825915, 0.0392544977, 0.0037712373, -0.0224953741, -0.0224735755, -0.0306389481, 0.0145528549, 0.0711392388, 0.0272524916, 0.0446176305, -0.0125160581, 0.0319690071, 0.028769983, 0.0194497965, 0.0011971178, -0.000979633, 0.0445674844, -0.0617400296, -0.0367339849, 0.0068638129, -0.024148602, 0.0296233743, 0.0164123103, 0.0189039558, 0.0258562453, 0.0239609387, 0.0690430552, 0.0016734638, 0.0121756978, 0.003607237, -0.0255320147, 0.0256231315, -0.0014297715, -0.0088054817, -0.0090318313, 0.0047233147, 0.0335087702, -0.0280009173, 0.0424046405, 0.0111680673, -0.0892866924, -0.0739750043, 0.0204486232, 0.0163958538, -0.0145057151, -0.0240947027, -0.0244638976, -0.0067797019, -0.0172952302, -0.0092949774, 0.0546368398, -0.0301794391, 0.0016599682, 0.0322906971, 0.0290820375, -0.0166185964, 0.0535967089, -0.0441013761, -0.0117746284, 0.0449500568, 0.049787011, -0.0144621311, -0.0177688245, -0.0232591759, 0.0399928167, 0.0012484831, -0.0087725194, -0.0078052166, 0.01583042, -0.0299715251, -0.0065957601, -0.1221252307, -0.0072683659, -0.0169304218, -0.0026733968, -0.012756858, 0.0210190229, -0.000168205, -0.026425356, 0.0060880575, 0.0054912427, -0.0045168577, 0.0595128573, -0.0210700706, 0.0088031748, 0.0130320638, -0.0015495187, -0.0411000773, 0.0162928328, 0.0872251913, -0.0400959998, -0.0527698137, -0.0141927386, 0.0425913185, -0.0241246503, 0.0069410042, -0.0404847227, -0.0256537776, -0.0010067902, -0.0640559047, 0.0725715384, 0.062763758, 0.0278825928, -0.0319373272, 0.0147850439, -0.0692002922, -0.042225197, 0.0043466045, 0.0248997379, -0.0458943024, -0.001661755, -0.0132660978, -0.0556366295, 0.0586767793, -0.0125708953, 0.0048216544, -0.0228009447, -0.0330871306, 0.0145295961, -0.0447986126, -0.05981021, 0.0024014506, -0.0024618795, 0.003137239, -0.0436746962, -0.0449629314, 0.0479670838, -0.0661746189, -0.0205731969, 0.0212408211, 0.0081413649, -0.0187219679, 0.064217262, 0.008476329, 0.0149975233, -0.024360925, -0.0052631274, -0.0153297279, 0.0070855008, -0.0028664817, -0.0247536115, 0.0627896339, -0.0323160365, 0.029726943, -0.0359340124, -0.0176089518, -0.0166110061, -0.016826259, -0.0310994722, -0.063320443, -0.0201427806, -0.0064878319, 0.0179728866, 0.0473514572, 0.0140107879, 0.0397229902, -0.0144049767, 0.0171228275, -0.0228876919, 0.0431824997, 0.0566308461, 0.0271984767, -0.0300368369, -0.041896604, 0.0170354098, 0.0844563991, 0.0448561646, -0.0046551386, -0.035669338, 0.0464691259, 0.0401265286, 0.0724897236, -0.0200888701, 0.0365630351, 0.0359673947, -0.0634068176, 0.0335691869, 0.0607602708, 0.0019133461, -0.0863569304, -0.0458045006, 0.004390717, 0.033305537, 0.000077159, -0.0001456412, -0.0919923708, -0.077772975, 0.0383489691, -0.0335955396, 0.0259667523, -0.0005488762, -0.0271620434, -0.1019583866, 0.017712051, 0.0115683526, -0.0183838606, 0.0016848767, -0.0249363463, -0.0180510301, 0.0182621572, 0.0153414998, 0.030647194, -0.0064349524, -0.0235828087, -0.0122166993, 0.036229115, -0.0236101355, 0.0271844137, -0.0330745168, 0.0103809722, 0.0311612971, 0.0016379758, 0.0373225249, -0.0264861379, 0.0432139523, -0.0024938218, 0.0105920266, -0.0340061039, 0.0050015147, -0.0578397736, -0.0104819257, -0.0188776571, 0.0148680182, 0.0013540279, -0.0356914103, -0.0128082903, 0.0451093167, -0.0014746103, 0.0031574185, -0.0307322182, -0.0175053254, 0.0013099546, 0.0612984151, 0.016130615, -0.017459901, -0.0174581185, -0.0055884575, -0.0163590387, -0.0278330557, -0.0230160933, 0.0073227314, -0.0007675138, -0.026779674, -0.0209652092, -0.0120281298, 0.0052202437, 0.0406666361, -0.0417482555, 0.0144743407, -0.0288032778, 0.0046063405, -0.0855506286, 0.0649782047, -0.0447058119, 0.0412206315, -0.0386949889, 0.0331982188, 0.0373049602, -0.0306678042, 0.0054894239, 0.0625941902, -0.1261573285, -0.0558745041, -0.0280264691, 0.0455028899, 0.0097189033, 0.0687373281, 0.0197043661, -0.055014804, 0.0256894007, 0.0254241657, -0.0262043197, 0.0602656156, -0.0105101727, 0.0217678454, 0.0074466863, 0.0114766601, -0.0479014255, -0.0089743994, -0.0127359536, 0.0091140112, 0.0066959658, -0.0316531211, -0.0328325815, -0.008843773, 0.0230434798, -0.0215314794, 0.0045750085, 0.0100830533, 0.037168812, -0.005143831, -0.017022673, -0.029069541, 0.0222123116, -0.0143992184, 0.0086951014, 0.0253060199, 0.0035823791, 0.0639217794, 0.0107112247, -0.0293303356, -0.0073584653, 0.0138224913, -0.0216370374, 0.0313016921, 0.016042674, -0.0114116762, 0.0213049874, -0.0074711582, 0.0115447966, -0.0502832979, 0.047556594, 0.0204511527, -0.0288781654, -0.0190686956, 0.0300827194, -0.0278409068, -0.007031003, -0.0136573203, -0.0080351736, 0.0746703371, -0.0160015989, 0.0439441614, -0.0087411143, -0.0102198096, -0.0409363322, 0.0324576497, 0.0201302934, -0.0460503995, 0.0486149006, 0.029263543, 0.0173828136, 0.0306820758, 0.0385730341, 0.0241862275, 0.0369145051, -0.0145541588, 0.0094967959, -0.0057878164, -0.0250383168, 0.0031407252, 0.0139927063, -0.0058921711, -0.0182803068, -0.024451565, -0.0773871243, 0.0131349815, -0.0063145687, -0.0455940887, 0.0148402406, 0.0432853736, 0.0345126726, 0.0313513204, -0.0420510955, 0.0064905174, 0.018727269, -0.0061836392, -0.0020327994, 0.0191892274, -6.564569508e-33, -0.0288771018, -0.0437014215, 0.0351694413, -0.0487262756, -0.0377495475, 0.0382043608, -0.0216263104, -0.0274713244, -0.012485506, -0.0323005579, 0.0078228917, 0.0044406629, 0.0212769061, 0.0230295453, -0.0210687146, -0.0116204508, 0.0181599353, 0.0136177205, 0.0128934393, 0.0183448121, 0.0248626601, 0.0346750878, 0.017552061, -0.0198163874, -0.0078328615, 0.0026110304, 0.0181746557, 0.0081283981, -0.0014083205, 0.0577848814, -0.0668658838, 0.0131523544, 0.0169463214, -0.0181804895, -0.0285887253, 0.047328338, -0.0830586031, 0.0185960606, 0.0378933102, 0.0277481824, -0.0357697681, -0.0448823124, 0.010422335, 0.0220995694, 0.0097251497, -0.0254590698, 0.0051239412, -0.0008728747, -0.0559699908, -0.0075875754, -0.020567907, 0.0202552732, 0.0141363721, -0.0058856751, 0.0662538931, 0.0213481169, 0.0390423499, 0.0673798695, -0.0814382136, 0.0197527539, -0.0343775749, 0.0586425737, 0.1052762121, 0.0594800524, 0.0181213971, 0.0358864255, 0.0369262509, -0.0424520671, -0.0588605106, 0.0049107452, 0.0274733379, 0.0451004468, -0.0137564614, 0.0684688017, 0.0250992682, -0.001239993, -0.0218822844, -0.0136439307, -0.011539463, 0.0453615524, 0.0393922254, 0.0205223057, 0.0082274228, -0.0275639948, -0.0528990477, 0.0251314826, -0.06358435, 0.009808749, 0.0456318744, 0.0360003933, 0.001167801, 0.0110792648, 0.1240593344, -0.0150186345, -0.0078597609, 0.0329864211, -0.0009295609, 0.054799147, -0.0142680956, 0.0509021245, -0.0136133591, -0.0130850188, 0.0525926948, -0.0198642276, 0.019756332, 0.0200392157, -0.0168555249, 0.0264085103, -0.049496036, -0.0276200715, -0.041131217, -0.0349231027, -0.0126230372, -0.0021300756, 0.0525466539, -0.0039449511, -0.0068426779, 0.0861599892, 0.0034355731, 0.0153377866, -0.0177383758, -0.0764731541, 0.008223691, 0.0145102087, -0.0102991434, 0.0257539935, -0.0311083216, -0.043181058, 0.0408004522, -0.0538072325, -0.0174833219, 0.0694090128, 2.702e-7, 0.0131613435, 0.019681273, 0.0384374931, -0.0003647044, 0.0227920748, 0.0445394814, 0.0120216962, 0.0204468388, 0.0049975142, -0.0574319437, -0.0232871287, 0.052178327, 0.048323825, 0.0051146355, -0.045750834, 0.0095156701, -0.0430548862, -0.0053040618, -0.0664426535, -0.0103031676, 0.0434849896, 0.0844801143, -0.0357780159, -0.0343243256, 0.053061761, -0.0141470423, 0.0641111732, -0.0190753099, -0.0334278941, -0.0218172073, 0.0013806311, -0.0240943078, -0.0106450999, 0.0025180937, 0.0117518483, -0.0502399132, -0.0033453929, 0.0552559383, -0.0042531318, -0.0788015798, -0.0044237738, 0.0000695453, -0.0640463755, -0.0414240882, 0.1083256528, -0.018706996, -0.0082510794, -0.050121773, 0.0016605246, -0.0304891206, -0.0166984294, 0.0030013025, -0.0141858244, 0.046433147, 0.0328554288, -0.0257302169, -0.0515730195, -0.1172764525, -0.0551299378, -0.012068266, -0.0680258721, -0.0314675979, -0.0679822415, -0.0428799503, 0.0752249807, -0.0701003671, -0.0416060016, 2.466318243e-34, 0.017860068, -0.0305894706, 0.0212215912, 0.0315590352, 0.0253459383, -0.0191685539, -0.0398039781, 0.0325336419, -0.014450714, -0.0895194262, -0.0454614088 ]
http://arxiv.org/abs/2406.18135v1
http://arxiv.org/pdf/2406.18135v1
Automatic Speech Recognition for Hindi
2024-06-26T07:39:20
2024-06-26T07:39:20
Anish Saha, A. G. Ramakrishnan
cs.CL
cs.CL, cs.SD, eess.AS
Automatic speech recognition (ASR) is a key area in computational linguistics, focusing on developing technologies that enable computers to convert spoken language into text. This field combines linguistics and machine learning. ASR models, which map speech audio to transcripts through supervised learning, require handling real and unrestricted text. Text-to-speech systems directly work with real text, while ASR systems rely on language models trained on large text corpora. High-quality transcribed data is essential for training predictive models. The research involved two main components: developing a web application and designing a web interface for speech recognition. The web application, created with JavaScript and Node.js, manages large volumes of audio files and their transcriptions, facilitating collaborative human correction of ASR transcripts. It operates in real-time using a client-server architecture. The web interface for speech recognition records 16 kHz mono audio from any device running the web app, performs voice activity detection (VAD), and sends the audio to the recognition engine. VAD detects human speech presence, aiding efficient speech processing and reducing unnecessary processing during non-speech intervals, thus saving computation and network bandwidth in VoIP applications. The final phase of the research tested a neural network for accurately aligning the speech signal to hidden Markov model (HMM) states. This included implementing a novel backpropagation method that utilizes prior statistics of node co-activations.
Automatic Speech Recognition for Hindi: Automatic speech recognition (ASR) is a key area in computational linguistics, focusing on developing technologies that enable computers to convert spoken language into text. This field combines linguistics and machine learning. ASR models, which map speech audio to transcripts through supervised learning, require handling real and unrestricted text. Text-to-speech systems directly work with real text, while ASR systems rely on language models trained on large text corpora. High-quality transcribed data is essential for training predictive models. The research involved two main components: developing a web application and designing a web interface for speech recognition. The web application, created with JavaScript and Node.js, manages large volumes of audio files and their transcriptions, facilitating collaborative human correction of ASR transcripts. It operates in real-time using a client-server architecture. The web interface for speech recognition records 16 kHz mono audio from any device running the web app, performs voice activity detection (VAD), and sends the audio to the recognition engine. VAD detects human speech presence, aiding efficient speech processing and reducing unnecessary processing during non-speech intervals, thus saving computation and network bandwidth in VoIP applications. The final phase of the research tested a neural network for accurately aligning the speech signal to hidden Markov model (HMM) states. This included implementing a novel backpropagation method that utilizes prior statistics of node co-activations.
[ 0.01841061, -0.0049691792, -0.010211518, -0.0063577355, -0.0535048768, -0.0270162728, 0.0107636154, 0.004554159, 0.0120672435, -0.0417970978, -0.0333051533, -0.0120590869, 0.0758139864, 0.0435211547, 0.0061588027, -0.0251142532, 0.0094955619, -0.0298232399, 0.0271859486, 0.0006937227, 0.0123297581, 0.0263257492, -0.0824543238, 0.0295224618, -0.0464689061, -0.0139185702, -0.0160565488, -0.007940338, -0.0146029051, -0.015578934, -0.062896654, -0.0002707585, -0.0008355366, 0.014846717, 0.0000017416, -0.0361974165, -0.0062147044, 0.0253442843, -0.0155922677, -0.0138511453, 0.0206868295, 0.0764221773, -0.048444014, 0.0682120845, 0.0168979876, 0.0392856188, 0.0771374255, -0.0062862528, 0.0057402919, 0.0585031398, -0.0246293191, -0.0148319062, 0.056789767, 0.0351739153, -0.0389767587, 0.0015401621, 0.0161138289, 0.00191539, 0.0240546465, -0.0536074825, 0.0057527148, 0.0090017989, -0.0347374044, -0.016516123, 0.017412534, 0.0036252013, -0.0460759327, -0.0494312048, 0.0029170548, 0.0274343472, 0.0135882171, -0.0127588408, 0.0570292994, 0.0140594412, -0.0489871129, 0.0185481329, -0.0086435387, 0.0174711756, -0.0349079743, 0.0097799832, 0.0194057301, 0.0086670704, 0.0060580778, -0.0061666765, 0.0943764299, -0.029704744, 0.0069404133, -0.036851421, -0.0181863327, -0.0471261553, 0.0487175882, -0.0018454157, 0.1052108929, 0.0208051503, 0.0023048944, 0.0227186531, 0.0165772326, -0.0576103888, 0.00873629, -0.0872981921, 0.0303902272, 0.0290314313, -0.0335812643, -0.0022215752, 0.0058777877, 0.0387651213, 0.0044902898, -0.0457377583, -0.0108711645, 0.0209729113, -0.0071171727, -0.034700837, -0.0070331455, 0.0216449555, -0.0012671944, -0.0174329448, -0.0509889573, 0.0168032441, -0.0155660436, -0.0080990521, 0.0675074384, -0.0327846371, -0.029156832, -0.0185945109, 0.0421991721, 0.0265158266, -0.0016404106, 0.0408766083, -0.0020276031, -0.1044212133, 0.0352666713, -0.0026525704, 0.0407685004, 0.0222470406, -0.0080314679, 0.0460563265, 0.008027602, 0.0021339171, -0.0614305399, 0.012843404, -0.0222090017, -0.0251043215, -0.0132498499, -0.0535794757, 0.0464204736, 0.0128600337, -0.0302073825, 0.0035870264, 0.0082639139, 0.0065570693, 0.0280593708, 0.0687502697, -0.0082166288, 0.0261965152, 0.0981728509, -0.0079529937, 0.0356655344, 0.0177381411, 0.0433632806, -0.0174840577, 0.0215593185, -0.05049688, -0.0008522971, 0.0515635945, -0.0518254749, -0.0290537588, -0.0341346152, -0.0052187047, -0.0329147652, 0.0391975641, -0.0160021279, 0.0390841812, 0.0283525474, 0.0259002335, 0.0664866418, 0.0881286785, -0.00384164, 0.0523382053, -0.0079481965, 0.0098087583, 0.0132186487, -0.0175989717, 0.0203560479, 0.0211809315, -0.0259931404, -0.0005407472, 0.015033232, 0.0085235015, -0.0046977545, -0.0010392817, -0.0203759652, -0.0479520485, -0.0183473025, 0.0195088573, 0.0485766828, -0.0230614562, -0.0419058278, -0.0162734482, 0.0258797053, -0.0568961576, 0.0099987946, -0.0558224022, -0.0077060512, 0.0314952694, 0.02463823, -0.0407127552, -0.0059388843, -0.0240003169, -0.0724359602, 0.010804317, -0.0045209033, 0.0059815864, -0.0081122611, -0.0599353984, 0.0027969328, 0.0817581117, -0.0130910119, 0.042851083, -0.07299833, 0.040530324, 0.017095428, 0.0139425052, 0.0031435271, -0.0061171143, -0.0176716708, -0.0004995537, -0.0360464826, 0.0200151615, 0.102153182, 0.0115923705, -0.0210906323, -0.0012074404, 0.030609915, 0.0060206177, 0.0262270048, 0.0363661647, -0.0152300205, -0.0042927465, 0.0639616475, -0.0452429429, 0.0306633562, 0.0741476044, 0.0481833369, -0.0914069638, 0.0457022004, -0.0520084463, 0.0041582086, -0.040441215, 0.0461233407, 0.0120941941, 0.0047983471, 0.0344520919, -0.0332161672, 0.0365355574, -0.0095375441, 0.0124128098, -0.027365841, 0.0340902098, -0.050566528, 0.004343858, 0.0415274426, -0.0534716658, 0.0348579474, 0.0103772851, -0.0266254973, -0.0384336933, 0.0083304588, -0.1268511415, -0.0048334966, -0.0002006434, -0.0062633669, -0.0111940168, 0.0733236074, 0.0079122279, -0.0582367815, -0.0262050573, 0.0097873108, 0.0162245519, -0.0087115988, -0.0083650909, -0.0370445661, -0.0461819097, -0.0353991464, -0.033611346, -0.0084856497, 0.0168118905, 0.0094461441, -0.0749888048, -0.0158446021, -0.0271346197, 0.028668439, 0.0482552722, -0.0082175033, -0.0402461849, -0.0198107585, 0.0037367342, 0.0372292213, 0.0360691622, 0.0114442362, -0.0274007637, 0.0009322069, -0.0045711622, -0.0355521366, -0.003366305, 0.0390689783, 0.0503473803, -0.013907575, -0.0139759248, -0.0160601735, 0.0212666877, 0.0499466732, 0.0253938995, 0.0115882196, -0.0844642743, -0.0473121107, 0.0165763497, 0.0060700858, -0.0287286509, -0.0946198702, 0.0246033128, 0.0248120315, -0.0344814174, 0.0207165238, -0.0560026579, -0.0162897166, -0.0719845444, -0.0469201691, 0.0103016952, -0.0082165739, -0.0171988327, 0.0022396652, -0.0348358639, -0.0741289631, -0.0274397638, 0.0154108359, 0.0110015078, -0.0789072439, -0.0651549995, -0.01785836, 0.0247350093, -0.0203501079, -0.0032831079, -0.0204644632, -0.0260691568, -0.0198362432, 0.0343924239, 0.0086453538, 0.0195053238, 0.0367709361, 0.0740344524, -0.0152974846, 0.050799273, 0.0037699365, 0.0284154322, -0.006244882, 0.0237198044, 0.0415510647, 0.0037608778, -0.047256317, 0.034057837, -0.009331842, -0.0358508937, 0.025590308, -0.0000972832, -0.0423085839, -0.0156591684, 0.0246641636, -0.0454992391, -0.019654613, 0.0423349999, 0.0193975475, -0.017516084, 0.0148637528, 0.0549377576, 0.0177360494, -0.0436934419, -0.0635146499, -0.0299886372, -0.0419893041, -0.0124514122, 0.0216746964, 0.0049452288, 0.0053667682, 0.0735578239, 0.0316706747, 0.0251573417, 0.0503080226, 0.0114215892, -0.0570478551, 0.0126242666, -0.0019571104, 0.0066887867, -0.0077955746, -0.0506599322, -0.0160789397, 0.0103365211, 0.0360543132, 0.0142973419, -0.0157959834, -0.0491068512, 0.0615628995, -0.0905316994, -0.026575515, 0.0076727532, -0.0333120748, 0.0815055966, 0.0781645849, 0.0381432399, -0.02423211, -0.0268424861, 0.0159521773, -0.0073085562, 0.0139657026, 0.0244053639, -0.0019712162, 0.0234668404, -0.0016675341, -0.0042767236, 0.0233457759, -0.0140160965, 0.0168187413, 0.0099687846, 0.0593288988, -0.0243959669, 0.013945898, -0.0117943101, 0.0632125884, 0.0216124337, 0.0120855467, 0.0107386876, -0.0325716771, -0.029870009, 0.030290857, 0.0473491251, 0.011065824, 0.0103624472, 0.0078581544, -0.0173535012, -0.0644790679, -0.0154549684, 0.0010232117, -0.0017187109, 0.0050233034, 0.0749513283, 0.0491036288, 0.073404707, 0.0395278931, -0.0476201996, -0.0375774242, -0.0145599367, 0.0500933789, -0.0535076894, 0.0015152429, 0.0361824818, -0.0811980963, 0.000696374, -0.0276321024, -0.0146475974, -0.0172340721, 0.0551080778, -0.0174618978, 0.0208590552, 0.0516367778, -0.0689632595, -0.0433734022, 0.0315425619, -0.0124639627, -0.0744422153, -0.0317200013, 0.0366083607, -0.0022477016, -0.0167772919, -0.0198804978, -0.05034687, 0.0169395339, -0.0398947969, -0.0248623416, -0.0108070103, -0.053527005, 0.0085731512, 0.0610139817, 0.0509898961, -0.0394613147, 0.0305765402, -0.0278152917, -0.0041479012, 0.0377793498, 0.0500531159, -0.0114841834, -0.0080492208, 0.0419901498, 0.0450498424, 0.0022314442, -0.0059517841, -0.0385673828, -0.0805005282, 0.0100136874, 0.1102742031, 0.0408188552, 0.0365329608, 0.0141102159, 0.0021967585, -0.019778613, -0.0026083549, 0.0495861769, 0.0101843588, -0.0144621572, 0.0269572865, -0.0210148934, -0.0111131221, 0.0182906445, 0.0148123698, 0.0118618794, 0.0452472009, -0.0099663045, 0.0603548735, 0.0541577004, 0.0148380091, 0.0229794495, -0.0128069045, 0.0394090749, -0.0409865156, 0.0519956909, -0.0417381302, 0.047733292, 0.0188379064, -0.0034189876, -0.0379289761, 0.0291330572, -0.0070338999, 0.0108721461, -0.0259086266, 0.0059643923, -0.0297305044, -0.0277189389, -0.0063166702, -0.0405105278, 0.0411176793, -0.0303989965, -0.0288411062, 0.0179958772, -0.0640580803, -0.0014426623, -0.014229957, 0.0549786203, -0.0380586274, 0.0132537605, -0.0182815306, -0.0150128379, 0.0322756134, -0.0117211407, 0.005106722, -0.0177417062, -0.0036643727, 0.0112977149, -5.483399028e-33, -0.0136538073, 0.0192621462, -0.0115018599, 0.0272111706, -0.0361656807, -0.0093866121, -0.0235581659, 0.0011007824, 0.0148361623, -0.0102948071, -0.0479929894, 0.010372756, 0.0333080441, -0.0104249921, 0.0343153663, -0.0017591905, 0.0535744838, -0.0200838577, 0.0574735515, -0.0222057309, 0.0331703499, 0.0521192737, 0.0454110876, -0.0402276888, -0.0460695401, -0.0308196638, -0.0212976728, -0.0437662564, 0.015225145, 0.0068000914, 0.0028118838, 0.0309006739, -0.0104460027, -0.0085680205, 0.0109178973, 0.0374722518, -0.1001096666, 0.0167978555, 0.0040255152, 0.1233802214, -0.0250407327, 0.0172685236, 0.0365785696, 0.0173412301, -0.0571512692, 0.0322893374, 0.0356574878, -0.0286996588, -0.0549300388, -0.0569954962, -0.0113440966, 0.0414754748, -0.0365438685, -0.0184928272, 0.0581150465, 0.0472368188, -0.0237541758, -0.0719448552, -0.0527142696, 0.0160789099, -0.0156535245, 0.0139108924, 0.0251428261, 0.0753511488, -0.0156430639, 0.0000906058, 0.0482448079, -0.1109298542, -0.0042611039, 0.0136886807, 0.0214374736, 0.0144646084, -0.007323883, -0.043862544, 0.0572740398, -0.0766318887, -0.0310154781, 0.0166883767, 0.0503977314, -0.0040891967, 0.0195275843, 0.0291705523, -0.0649405047, -0.0207805466, 0.0065418226, -0.0296265297, -0.0473322384, -0.0218823962, 0.0552728809, -0.0178462025, -0.0068771592, 0.0161366221, 0.0146444552, -0.0310300924, 0.0383893549, 0.0054536401, -0.0380480066, 0.0062911883, -0.0121773491, 0.007369637, -0.0348863564, -0.0356574133, -0.0093524186, 0.0445943847, 0.0395302251, -0.0014771682, -0.0085068289, 0.035559468, -0.0581234731, -0.0029698673, 0.017285157, 0.0300737508, 0.0228745844, 0.0401437506, 0.0244627539, -0.0493922234, 0.0051031136, 0.0358718187, 0.003758274, 0.0281555783, 0.0153860524, -0.0003663625, -0.0167191289, -0.0123908166, -0.0552313291, 0.0274781846, -0.0131610176, 0.0573896654, 0.0197020043, 0.0033988825, -0.0023173604, 0.0211628601, 2.74e-7, -0.0550161563, 0.0455392748, 0.0378658101, -0.062216673, 0.0182476006, 0.0062579201, -0.0446821563, -0.0034167857, 0.0224036518, 0.0538459942, 0.0099176141, -0.0573501438, 0.0994106457, 0.0364648551, -0.0837595239, -0.0622378625, -0.1012005806, 0.0414033979, -0.0273838341, 0.06669835, -0.0276448205, 0.0457389839, 0.0125287697, -0.0472197942, 0.0127396453, -0.0249130372, 0.0117612565, -0.0086277211, 0.0173637103, -0.0109688621, -0.0014934832, -0.0191553086, -0.0231480896, -0.021870207, -0.0540622845, 0.0311760809, 0.0253889132, 0.0460514612, -0.0233515706, 0.0155249089, 0.0365819223, 0.0359876864, -0.0072745774, 0.0369291417, 0.03874157, 0.0356461816, -0.0672289059, -0.0230834913, -0.0483269989, -0.0300042313, 0.0019594454, 0.0063266996, 0.0166587103, -0.0327064134, -0.0319761038, -0.0564516261, -0.0384541191, 0.0083960388, 0.0030945197, -0.010217228, -0.041383557, -0.0187213682, -0.0000924857, 0.0114262775, 0.0424414724, 0.0310633667, 0.002328336, 1.95864634e-34, 0.0127465436, -0.0338737555, 0.0092942026, -0.0107443938, 0.0059161512, -0.0125574619, 0.0278594345, 0.0405551903, 0.0202278644, -0.0243084673, -0.0545905381 ]
http://arxiv.org/abs/2406.18134v1
http://arxiv.org/pdf/2406.18134v1
Assessing "Implicit" Retrieval Robustness of Large Language Models
2024-06-26T07:38:24
2024-06-26T07:38:24
Xiaoyu Shen, Rexhina Blloshmi, Dawei Zhu, Jiahuan Pei, Wei Zhang
cs.CL
cs.CL
Retrieval-augmented generation has gained popularity as a framework to enhance large language models with external knowledge. However, its effectiveness hinges on the retrieval robustness of the model. If the model lacks retrieval robustness, its performance is constrained by the accuracy of the retriever, resulting in significant compromises when the retrieved context is irrelevant. In this paper, we evaluate the "implicit" retrieval robustness of various large language models, instructing them to directly output the final answer without explicitly judging the relevance of the retrieved context. Our findings reveal that fine-tuning on a mix of gold and distracting context significantly enhances the model's robustness to retrieval inaccuracies, while still maintaining its ability to extract correct answers when retrieval is accurate. This suggests that large language models can implicitly handle relevant or irrelevant retrieved context by learning solely from the supervision of the final answer in an end-to-end manner. Introducing an additional process for explicit relevance judgment can be unnecessary and disrupts the end-to-end approach.
Assessing "Implicit" Retrieval Robustness of Large Language Models: Retrieval-augmented generation has gained popularity as a framework to enhance large language models with external knowledge. However, its effectiveness hinges on the retrieval robustness of the model. If the model lacks retrieval robustness, its performance is constrained by the accuracy of the retriever, resulting in significant compromises when the retrieved context is irrelevant. In this paper, we evaluate the "implicit" retrieval robustness of various large language models, instructing them to directly output the final answer without explicitly judging the relevance of the retrieved context. Our findings reveal that fine-tuning on a mix of gold and distracting context significantly enhances the model's robustness to retrieval inaccuracies, while still maintaining its ability to extract correct answers when retrieval is accurate. This suggests that large language models can implicitly handle relevant or irrelevant retrieved context by learning solely from the supervision of the final answer in an end-to-end manner. Introducing an additional process for explicit relevance judgment can be unnecessary and disrupts the end-to-end approach.
[ 0.0637866259, 0.0484694578, -0.0221835077, 0.031572286, -0.0278004128, -0.0235915836, -0.0218353104, 0.0177982561, -0.0030334624, -0.0705785602, -0.0297070239, -0.0247442238, -0.0333563648, 0.0031294352, 0.0320084579, -0.012664835, 0.0732079446, -0.0374061055, 0.0147640398, -0.0034465452, -0.0162488073, 0.0028367452, 0.0319472812, 0.046780996, -0.0071181534, 0.0084466487, 0.008540825, 0.0111300135, -0.0133581208, 0.0004544743, 0.0269014761, 0.0333568268, -0.0193660576, 0.0613168329, 0.000001852, -0.0337830074, -0.0293215718, 0.0084724221, -0.0225508306, 0.0136207761, 0.0163601004, 0.0230040234, 0.0048133023, 0.0226830412, -0.0518873967, 0.0266978554, 0.0500048883, 0.0388094932, 0.0399376228, 0.0468927324, 0.012213368, 0.0099941809, 0.0098612243, -0.0310638417, 0.0128656765, -0.0671819746, 0.0211828779, 0.010309522, 0.0126513643, 0.0460237339, 0.0016038112, 0.0659688637, 0.0144703761, -0.029037945, 0.003452047, 0.0199547578, -0.065689072, -0.0012458026, 0.0031395531, 0.0374421887, 0.0343700871, -0.0162678342, -0.0194312707, -0.0049425773, 0.0122522665, 0.0273598917, -0.0463857502, 0.0169317927, 0.0075120428, -0.0108552389, 0.0009461119, 0.0023012431, 0.0280259307, 0.0337567739, -0.0079507995, 0.0435515121, 0.0392101035, -0.0238770135, -0.0334934965, 0.0195509177, -0.0525229238, -0.0462807044, -0.0062476541, 0.0563246943, -0.0343179107, 0.0447102338, -0.0213993546, -0.015398128, 0.047425732, -0.0468996949, 0.0383773409, 0.054136131, -0.0460540019, 0.0411587134, -0.0510328338, 0.0307454709, -0.0384136587, -0.0182680469, -0.0424198136, -0.0295762271, -0.0048577567, -0.0428511649, -0.0420164168, 0.0593937896, -0.0348156132, -0.0200401209, -0.0267787799, 0.0235278159, -0.0118462387, -0.0055846386, -0.0187464934, 0.0416368246, -0.0756100267, -0.0183726698, -0.0164167248, 0.0194518305, -0.0390949734, 0.0116643906, 0.0389736556, -0.0772902817, -0.0058861217, -0.0150227258, -0.0002342513, -0.0399667174, -0.0260788891, 0.0866928399, -0.0145578934, -0.0360746346, -0.0461217277, -0.0579869449, -0.0487093143, -0.0465103984, -0.0010695169, -0.0384623706, -0.0033893713, 0.0192215592, -0.0221835375, -0.033310838, 0.0259388536, 0.0387957022, -0.0181149207, 0.0288628675, 0.0129589327, -0.0017838844, 0.0359348319, 0.020492807, 0.0030155454, 0.0793279856, 0.0008181338, 0.025339365, 0.0560755059, -0.0025335148, -0.0079717953, 0.0178892184, 0.0380537063, 0.0145855639, -0.0459352545, 0.0227894336, 0.0248764157, -0.0086697154, -0.0150277643, 0.0810348094, -0.0340015329, -0.009671188, 0.0593563542, 0.0131295323, 0.009020119, 0.010746656, 0.0615023039, -0.0228295717, 0.0038666956, 0.0453093313, -0.0551599041, 0.0314245932, -0.1004171968, 0.0126839643, 0.0485705249, 0.0101913083, -0.0171019919, -0.0464133658, -0.0088276984, 0.016120635, 0.006063411, 0.026693495, 0.0703864396, 0.0270009637, -0.0377346985, 0.0455443822, 0.0234094094, -0.0459510796, 0.013897174, -0.0803044215, 0.0476725772, 0.0342007577, 0.0773501098, -0.0175213292, -0.0031988472, 0.0346976444, 0.0029845345, 0.0327860974, 0.0600635484, -0.0133610852, 0.0059245587, -0.0356826708, -0.0240047257, -0.0121909818, 0.0302297473, 0.0132890139, 0.0612132959, -0.0051236232, 0.0499681719, -0.0150925266, 0.0144016631, 0.0273149293, -0.0454881564, -0.0580426455, -0.0085648624, -0.0178832803, -0.0088741817, 0.0045217122, -0.0102955084, 0.0430005752, 0.0265901089, -0.0053425077, -0.0296092536, -0.0362260565, -0.0708643496, 0.0207423214, 0.0210972391, 0.0373699479, 0.0007328123, 0.0044064173, -0.006356502, -0.03059089, 0.0367557853, -0.0512628667, 0.0793031752, -0.0546292365, 0.0163620207, 0.0083203455, 0.0095837861, -0.0495189466, 0.06243724, -0.001741722, 0.0202103406, 0.0316109769, 0.0427656174, 0.0066016992, -0.0762603581, -0.0215384029, 0.0175257958, -0.0210029837, 0.0318238959, 0.064566426, -0.0264303461, -0.0438628793, -0.0287354365, -0.0716189146, -0.014348763, -0.0258177798, -0.0295956526, 0.0182373971, 0.0267547201, 0.0061773648, -0.0172464699, -0.001906881, 0.0047557838, 0.0342846066, 0.0433350876, -0.0088923359, 0.0344700404, 0.007596245, -0.0642661154, -0.0352831595, 0.0155693972, 0.053957846, 0.0010099675, 0.005279304, -0.0326082073, 0.0308814477, -0.0709075555, 0.0417359322, -0.0024802506, -0.0290050264, 0.014742245, -0.0532619208, 0.0228896309, 0.0509607904, 0.0338491946, 0.006247228, 0.0525855981, -0.0478017889, -0.0797030702, 0.012930573, -0.0294628795, -0.0929230079, 0.0070435344, -0.0053025214, -0.0134349186, 0.0439772494, 0.0086788237, -0.012408969, -0.0590272658, 0.0346246585, -0.0375780277, -0.0547422022, -0.0394141227, -0.0263565257, -0.0243969988, 0.0567980856, -0.0283016767, -0.0607383065, 0.0748794228, -0.0193755403, 0.0167788044, 0.002406999, 0.0034509716, -0.0143826464, 0.0538097844, -0.0333738141, 0.0346981511, -0.0143600376, -0.0173513833, -0.0534218289, -0.045472458, 0.0297990069, -0.0678262413, 0.0720042214, -0.0468899459, 0.0307063982, 0.003105964, -0.0407201275, -0.0785257071, -0.0316896476, -0.0011817085, 0.0362587832, -0.013396061, -0.0037616491, -0.0361181386, 0.0023660953, 0.0488110818, 0.0153006036, -0.03071866, -0.0212903284, -0.0087981867, 0.0519231893, 0.0800660104, -0.0043235812, 0.0198540296, -0.040429458, 0.0015420779, 0.0626782775, 0.0075155832, -0.0209465623, -0.0068525826, -0.0026191974, 0.0437828079, -0.0130906906, -0.0182486344, 0.0380862206, 0.03156377, -0.0661758259, -0.0516900867, 0.0618551113, 0.0397208966, -0.0352448635, -0.0095919119, -0.0494429171, 0.0139422631, 0.0249807872, 0.0504340827, -0.0411029346, -0.0166577213, 0.034261886, 0.0372127667, -0.0146132531, -0.021678865, -0.0183945671, -0.1632450372, 0.0315887183, 0.0062434967, -0.0295609608, -0.0341145806, -0.0179344807, 0.0229315627, 0.0264996421, 0.0351854227, 0.0497527085, -0.0027956055, -0.0124983983, 0.0065035177, 0.0394680724, -0.0497367494, 0.0153575595, 0.0040808651, 0.0280713737, 0.0877607539, 0.0020415005, 0.0192414429, -0.0019533578, 0.0608247519, -0.0420985483, 0.0204997603, -0.0039153034, 0.0334363766, 0.008649555, 0.004298964, -0.013680283, -0.0142801646, 0.0274900217, -0.0208807345, -0.006395, 0.0428153425, -0.0068302276, 0.0112432195, -0.0457805283, 0.0033890624, 0.0286405552, 0.0629645884, 0.0217254087, -0.0093186572, -0.0177061763, -0.0277451985, 0.0043325112, -0.0218934435, -0.0363963321, 0.0109386379, 0.013106944, 0.0285120029, -0.0696611032, -0.0445299335, -0.0200249907, 0.0207815878, -0.0033895567, 0.0065673864, -0.0532680899, 0.004123345, -0.0215028077, 0.0304839555, -0.0600427538, 0.0028807572, -0.0356108993, 0.0106289051, 0.0186953843, -0.0601258427, -0.0129246321, 0.0623410493, -0.017306963, -0.0657094792, 0.0097095482, 0.0522569343, 0.0323476084, 0.0403490439, -0.0180132054, -0.0601035133, 0.0318130814, 0.0167627465, -0.0605822466, 0.0201851279, 0.0084419409, 0.0050978907, -0.0137414224, 0.0479712375, -0.0968196988, 0.0249623749, 0.0178422667, 0.0756326094, 0.0178326517, -0.0384190381, 0.0410280898, 0.0213873107, 0.012630498, 0.0024101238, 0.0127554778, -0.0090071214, 0.0136106452, -0.0442854948, -0.0263215005, -0.0283905584, 0.0314517803, -0.000015827, -0.0173639823, 0.0057782847, -0.0177222695, 0.0100813741, 0.0787117407, -0.0371886045, -0.0227984749, 0.0022948636, -0.0004458348, 0.0280023105, 0.016953934, -0.0311106388, 0.0162273087, 0.0007363259, 0.0584380813, -0.0353046805, 0.0588567555, 0.0194096249, -0.0158798955, -0.0322937854, 0.0034793192, -0.0935699716, -0.0325319804, -0.0473451726, 0.0116051584, 0.0464373045, -0.027778551, 0.0457294621, -0.0152798966, -0.0403905585, -0.0308321435, -0.0033331816, 0.0508920699, -0.0599234253, 0.014404282, 0.0313011222, 0.0070708659, 0.0388578251, 0.0464960746, -0.0209427215, 0.0366114937, -0.0518477, -0.0046488503, -0.0238838363, -0.0352502391, 0.0211096872, 0.028531149, -0.020069804, 0.0133565944, -0.0624487735, -0.0969160348, -0.0231530294, 0.0149575835, -0.0123130735, -0.0177724306, 0.0304016154, 0.0893177316, 0.0226827133, -0.0204963312, -0.0489382595, 0.0250025727, -0.0517205931, 0.0231125113, -0.0472052172, -6.567278288e-33, -0.0599010661, -0.0370650664, 0.0113456827, -0.0636719316, -0.0136728594, 0.0304945447, -0.0782734379, -0.0017483353, -0.0336745791, -0.0157616138, -0.0305838566, 0.0227268673, 0.0054168291, 0.0089636371, 0.0117105013, -0.0051673548, 0.0387218595, 0.0141694816, -0.0380723886, 0.0326613523, 0.054593727, 0.0151518667, 0.0170908831, -0.060192436, 0.0318854824, 0.0010302782, 0.0037503135, 0.017815154, -0.0244000796, 0.0887215808, -0.0237766597, 0.0570210218, -0.0028976745, -0.003224208, -0.0246739499, 0.0023520929, -0.0646867007, -0.0248516072, 0.0139076905, -0.0353290103, -0.0360427089, -0.0248101018, 0.0240121633, -0.0179079566, -0.0156352445, -0.0114734489, 0.0221879706, -0.0180647187, -0.0314372331, -0.0246143006, -0.0659405366, 0.0208886638, 0.0198943, 0.0015219693, -0.0021465905, 0.05538892, 0.0287552532, 0.0367168225, -0.0027294676, 0.0240746345, 0.0147851435, 0.0549770407, 0.0463769659, 0.0363558382, -0.0183272231, 0.0410020649, 0.0072656395, 0.0109988526, -0.0063014068, -0.0891805962, 0.0151387062, 0.0231570527, 0.0473016128, 0.0630016848, 0.015163634, -0.0052883271, -0.0308563057, 0.0025005557, -0.0078779077, 0.0397667699, 0.0633148551, 0.006213563, -0.0337904952, -0.0084218243, -0.0501307547, 0.0067290822, -0.0619506538, -0.02817812, 0.0633532703, 0.0397484601, -0.0225724932, -0.0126216579, 0.0408552326, -0.0445489362, 0.0087237535, 0.0541004986, -0.018901227, -0.0073894602, -0.0158510376, 0.0362844579, -0.0102014365, -0.0146970572, 0.0163557287, -0.0180948488, -0.0443121977, 0.0111951167, -0.0083431611, 0.0225583464, -0.0558456592, 0.0171402395, -0.0426887125, 0.0389091745, -0.0079097198, -0.0384489857, 0.0415055864, 0.0138811413, -0.0051830062, 0.0525129735, 0.0344937891, 0.0296637975, -0.0117020998, -0.0283609461, 0.0205692723, 0.0069248751, -0.0255162548, 0.0189960487, -0.0190131906, 0.0623949356, 0.020384144, -0.0089002214, -0.0183422789, 0.0172689836, 2.634e-7, 0.0576716587, 0.0137982508, 0.0269836746, 0.0170572996, -0.0076608532, 0.0140982978, -0.0126480572, -0.0019275605, 0.05455992, 0.0141200218, -0.0194426812, 0.0353262164, 0.0457272381, 0.0241442602, -0.0375442989, 0.0115255769, -0.0770479888, -0.0016540758, -0.0155974953, -0.0130671645, 0.050236918, 0.0759602785, -0.0095584048, 0.0032526043, -0.0375857092, -0.0838722363, 0.0856659114, -0.01251119, -0.004157647, 0.0062015667, 0.0299022533, 0.0470403582, 0.022706233, -0.0235962756, -0.0086550834, 0.016877614, 0.0080650691, 0.0486004688, 0.003711794, -0.047926303, -0.0228536129, -0.0114411954, 0.0206718538, -0.0606751144, 0.0581561886, -0.0525242165, -0.0102619128, -0.0245598573, -0.0316158645, -0.0086097922, -0.0053279204, -0.0182410888, -0.0399862826, 0.007224285, -0.0028515025, 0.0515294373, -0.0089420974, -0.0996159762, -0.0479449704, 0.0129642794, -0.0442187563, -0.0595819391, -0.0257905927, 0.0169010572, -0.0275827684, -0.0601570308, 0.0155995358, 2.469057237e-34, 0.0431611277, 0.0028628069, 0.0151047977, 0.0103612086, 0.03899597, 0.0009704345, 0.004885097, 0.0256470796, 0.0391152427, -0.0843251646, -0.0427396931 ]
http://arxiv.org/abs/2406.18133v1
http://arxiv.org/pdf/2406.18133v1
ConvoCache: Smart Re-Use of Chatbot Responses
2024-06-26T07:35:10
2024-06-26T07:35:10
Conor Atkins, Ian Wood, Mohamed Ali Kaafar, Hassan Asghar, Nardine Basta, Michal Kepkowski
cs.CL
cs.CL
We present ConvoCache, a conversational caching system that solves the problem of slow and expensive generative AI models in spoken chatbots. ConvoCache finds a semantically similar prompt in the past and reuses the response. In this paper we evaluate ConvoCache on the DailyDialog dataset. We find that ConvoCache can apply a UniEval coherence threshold of 90% and respond to 89% of prompts using the cache with an average latency of 214ms, replacing LLM and voice synthesis that can take over 1s. To further reduce latency we test prefetching and find limited usefulness. Prefetching with 80% of a request leads to a 63% hit rate, and a drop in overall coherence. ConvoCache can be used with any chatbot to reduce costs by reducing usage of generative AI by up to 89%.
ConvoCache: Smart Re-Use of Chatbot Responses: We present ConvoCache, a conversational caching system that solves the problem of slow and expensive generative AI models in spoken chatbots. ConvoCache finds a semantically similar prompt in the past and reuses the response. In this paper we evaluate ConvoCache on the DailyDialog dataset. We find that ConvoCache can apply a UniEval coherence threshold of 90% and respond to 89% of prompts using the cache with an average latency of 214ms, replacing LLM and voice synthesis that can take over 1s. To further reduce latency we test prefetching and find limited usefulness. Prefetching with 80% of a request leads to a 63% hit rate, and a drop in overall coherence. ConvoCache can be used with any chatbot to reduce costs by reducing usage of generative AI by up to 89%.
[ 0.042297367, 0.052087009, -0.00990922, -0.0195653047, -0.0210927557, -0.0234667417, 0.0384964906, -0.0074972976, 0.0587018579, -0.0198166184, -0.0051700594, -0.0372155868, -0.028010536, 0.0667610243, -0.015606028, -0.0534453616, 0.0445494913, -0.0200503059, 0.0219876468, -0.0195069648, -0.0081437165, -0.0134353871, -0.0141545627, 0.0781571195, -0.0665963665, -0.0494297221, -0.0413839035, 0.0274001658, -0.0000696364, -0.0138037922, -0.0128750373, 0.0178621039, -0.019910166, 0.022992922, 0.0000022169, -0.044013001, -0.0215429496, -0.0122219464, -0.0765406415, -0.0018083997, 0.0150916176, 0.0439207479, -0.0589573421, 0.0454899631, -0.0307893269, 0.0421984866, 0.0453416184, -0.008295333, 0.0621110126, 0.0843140632, -0.0227994528, 0.0028065457, 0.0218585059, 0.0125753433, 0.006741839, -0.0640356019, -0.0346821621, 0.0029273271, 0.0025423821, -0.0391032062, -0.0221008789, 0.0467694253, 0.037296664, -0.0328216441, 0.0029798048, 0.0170362927, -0.0503895208, -0.0432456471, -0.017624924, 0.0529091172, -0.0015202941, 0.0058640093, 0.0192491617, 0.0058077089, 0.0142156351, -0.032434009, -0.0205541924, -0.0014095168, -0.036644578, -0.0181666669, -0.0462378487, -0.0048487489, -0.0219660718, -0.018154392, -0.0155775808, -0.0352819934, 0.0209797025, -0.0143736117, -0.044972524, -0.0025589759, 0.0288986303, -0.0682414323, 0.0064906767, 0.0582821593, -0.0809125677, -0.0008493111, -0.0608138405, 0.0059816749, 0.006210696, -0.0194964055, 0.0035160591, 0.0561930239, -0.0425453037, 0.0049426816, -0.008175238, 0.0477129482, 0.0024221258, -0.0369763933, -0.0332239717, -0.0267178789, -0.0400328562, -0.0149044842, -0.0629696622, 0.0428138897, 0.019835664, -0.0461978763, -0.055761639, 0.0605593994, -0.0177312773, -0.0013346878, -0.0087094223, 0.0274301954, -0.0369175449, 0.0150431069, -0.0235352702, -0.0029036738, -0.0410731472, -0.0041466039, 0.0181162395, -0.1416526437, -0.006305811, 0.002509447, 0.0201858934, 0.0215233322, 0.0072208513, 0.0694490075, -0.0219428726, -0.0260056313, -0.0134938341, -0.061115101, 0.0292200707, -0.0630600974, -0.0244527739, -0.0394753292, 0.0008355248, -0.031136889, -0.0002892095, 0.0170460753, -0.0230509043, 0.0046594432, 0.0026934212, 0.0051700561, 0.0088342577, 0.0366753042, 0.028923599, 0.0086163851, 0.0132381096, 0.0270077735, 0.031579148, 0.0175235253, 0.030659413, -0.027313726, 0.0327157453, 0.0077453093, 0.0141497515, -0.0156477373, 0.0184483081, -0.0017106921, -0.0185763892, 0.0397790894, 0.0211383402, 0.0191963799, -0.0445319563, -0.0111366715, 0.0616949201, 0.0596705042, -0.0001299437, -0.0025228949, 0.0353291444, 0.0344557837, 0.0264243763, 0.004359873, -0.0246839151, 0.033897236, -0.022346966, 0.0107033709, 0.0262872651, 0.0228674449, -0.0646316037, -0.0337670408, 0.0425655395, -0.0372440442, -0.0136644188, 0.0298105385, 0.0061568972, 0.0137752388, -0.0505646616, -0.0253076125, -0.0290891994, -0.0749787837, -0.0696584284, -0.0513449684, 0.0156677328, 0.117616564, 0.0473900549, 0.0227784105, -0.038268134, 0.0274866503, -0.0209376346, 0.0300537795, 0.0359961428, -0.0083136298, 0.0343409628, -0.0239187554, 0.0005106286, 0.0173821282, -0.0169603787, 0.0516331755, 0.0080230404, 0.0218621995, -0.0135945342, 0.0900944546, 0.0581182986, -0.0080668377, -0.0086237192, -0.0057070092, 0.001467062, -0.0033047164, 0.0730031729, 0.027546512, -0.0205294844, -0.0066440748, 0.0045021367, -0.0106226485, -0.0191711392, 0.0545217767, -0.0696885884, -0.0220783837, 0.0299060009, -0.0146865742, -0.0077393027, 0.0296335984, -0.0147332903, -0.0499655716, 0.0428155959, -0.0377530567, 0.0683347434, -0.0233442429, 0.0377846099, -0.0041512814, 0.0263537113, 0.0254099835, 0.0154592646, 0.0151904188, -0.0201517679, 0.0379578769, 0.0794920623, 0.0010325037, -0.0682948455, 0.0309538543, 0.0403684974, -0.026675947, 0.0327764824, 0.0085703833, 0.01320644, -0.0549376421, -0.0197916701, -0.1135748178, -0.0031125932, 0.0229328424, -0.0443885401, -0.0082928324, 0.003264958, -0.0122631369, -0.0348626003, -0.0334124565, 0.0304356404, 0.0353219919, 0.0320265703, -0.0183575284, -0.0386469699, -0.0537179075, -0.0481211357, 0.0226224922, 0.0496734828, 0.0050489986, -0.0027430046, -0.0080080507, -0.0278040711, 0.0068788049, -0.0371424146, 0.0296866, 0.0036558891, -0.0625898466, 0.0384859107, 0.0088401018, 0.0410970263, 0.0590730272, 0.0506649539, -0.0098120803, 0.0023703824, -0.0142309209, -0.0801803097, 0.0181658138, 0.0086185979, -0.0351481847, -0.0117310798, -0.0121139446, 0.0077406932, 0.0362359248, 0.0212864839, 0.0111388629, -0.0402695313, 0.0251334198, 0.0278269891, -0.0561508462, -0.013109819, 0.001981911, -0.053300567, 0.023821814, 0.0150195062, -0.0588805936, 0.0318976864, -0.0111374706, -0.0092391828, -0.0003012336, -0.0422100984, 0.0045728218, 0.0524896309, -0.017762756, 0.0286468267, 0.0186404642, -0.0191769414, -0.0315635838, 0.0023404749, 0.0136574833, -0.0771695152, -0.0083900224, 0.0052176928, 0.003511868, 0.0099060843, -0.0477266014, -0.0879783556, -0.0090692881, -0.0378861092, -0.0080114091, 0.0405591168, -0.0426780879, -0.0297355931, 0.0061060051, -0.0074467734, 0.042137064, -0.0021004134, -0.0216312483, -0.0101291696, -0.0067909313, 0.038128227, 0.0720116571, -0.0063447799, 0.0385687165, 0.0684500784, 0.0622357279, 0.0608534366, -0.0140665863, -0.0371942669, -0.0036265024, 0.0213903915, 0.0226990934, -0.0038419864, 0.020366827, 0.0319055319, -0.0408307761, -0.0122290952, 0.0782185495, -0.0072072563, -0.0063592247, 0.0117979906, -0.0128589282, 0.042029608, 0.0150527488, 0.0332320817, -0.0701638982, -0.043725457, 0.0644469261, 0.0418296009, -0.0519700013, -0.0102505684, -0.0207346082, -0.1297663301, 0.0238435064, -0.0142282983, 0.0040805666, 0.0012759614, -0.0254562572, -0.028215792, 0.0127675114, 0.0009835272, 0.0112689091, 0.026412122, -0.0108001847, 0.0030159319, -0.0456966422, 0.0150509933, 0.0177866314, -0.0157534629, 0.1002192497, 0.0747507289, 0.0591260456, 0.05983942, 0.0027754877, 0.0155677386, 0.0231966693, 0.0120454943, 0.0042831413, 0.0372664817, 0.0259691272, -0.0168308485, -0.0125720818, 0.0286806319, -0.00361257, 0.0062088878, 0.0197834522, -0.0077490821, -0.0423640236, -0.0065348805, -0.024982905, 0.0441024788, 0.0284210499, 0.0551034957, 0.0265552886, -0.0381417833, -0.0603156202, -0.011725123, 0.0153239863, -0.0453603119, -0.0636228323, -0.0138675468, -0.00759529, -0.0390676968, -0.0764698684, -0.015105919, -0.0446241647, 0.0440390408, 0.0023460665, -0.0086808428, 0.012224392, 0.0332888365, 0.0030806619, -0.0192939546, -0.0323191024, -0.0088135423, -0.0779411048, -0.0225482993, 0.0282018129, -0.05092711, -0.0195710454, 0.0467529818, -0.0384007879, -0.0447352342, 0.0278551914, 0.0269516855, 0.0288195256, 0.0404188335, -0.0051542013, -0.0178006589, 0.0312613957, 0.0573746264, -0.0691331252, 0.0282587763, 0.0634330958, 0.0442460515, -0.0084451251, 0.0252084211, -0.0758958906, 0.0567671284, -0.029792374, 0.0312491562, -0.0085313683, -0.0609575659, 0.0087113595, 0.0552978218, 0.0042001358, 0.0081521524, 0.0248752981, 0.0149998916, 0.0544841997, -0.0005894556, 0.0232373998, -0.0502123274, 0.0064182524, -0.0131856333, -0.0270125754, 0.0225971639, 0.0148444287, -0.0244215578, 0.0196771696, 0.0238779373, -0.0374228135, -0.0164742172, -0.0056388718, 0.0107567711, -0.0105230641, -0.000801302, 0.0385574177, 0.0149411727, 0.0082530072, 0.0036306735, -0.0094909482, -0.0038108355, 0.0059209676, -0.0099646216, 0.0199365821, -0.0877277032, -0.1001562998, 0.0276818015, 0.0430671461, 0.0405189246, 0.0280095004, 0.0468965769, -0.0246654917, 0.0382110626, -0.0430050753, 0.0430287234, 0.0340992995, -0.0259180386, -0.0387402959, 0.0066995556, -0.035280481, -0.0031031903, 0.0562207848, -0.0595546514, 0.0688666627, -0.0424622111, 0.0046345056, -0.0513304658, -0.0509842187, -0.0346924476, -0.0046855109, 0.0201112386, -0.0322369374, -0.0425582901, -0.0289344303, -0.0316769183, 0.0008887121, -0.0021860262, -0.0166482832, 0.0431627966, 0.019418668, -0.0063206013, 0.0028463409, -0.0316975303, 0.0120234052, -0.0362795219, -0.0134314904, 0.0348090567, -6.671868632e-33, -0.0178967677, -0.0051239068, 0.0299636535, -0.0070998757, -0.0399569534, 0.0213197507, -0.0556195453, -0.0039897272, -0.0234980229, -0.0289699193, -0.0390702337, 0.0227296613, 0.0321411677, 0.0035225118, -0.0227571893, -0.0028634847, 0.0816059634, -0.0177524164, -0.0104787378, 0.0560805127, 0.0813356638, 0.0276107509, -0.033434812, -0.0724421516, -0.0436975211, -0.0184340812, 0.0126273455, -0.0087683676, 0.0318309665, 0.0229242388, -0.0065115723, 0.0430847071, -0.0168501735, -0.0268199462, -0.0266904477, 0.034369573, -0.0529184043, -0.0333079174, -0.0133433519, -0.0066250502, -0.0048094848, -0.0350121446, -0.0116966292, 0.0254650749, -0.0387892537, 0.0056060203, 0.0118365474, -0.0138842538, -0.0324959382, -0.0241000038, -0.0633766428, 0.012503067, -0.0358970687, -0.0655581132, 0.004312058, 0.0928729922, 0.0090928646, 0.0065605803, -0.0292677861, 0.0450527668, -0.0259768795, 0.0215393156, 0.0148457391, 0.0158627145, 0.0395432487, 0.0735634267, 0.0710446239, -0.0000571414, -0.0265022777, 0.0057167346, 0.0273442455, 0.0268919393, 0.0142890764, 0.0307802539, 0.0658911839, -0.0646174327, 0.000997415, 0.0221945681, 0.0601383038, -0.000411051, 0.0409274139, 0.0152309621, -0.0378839672, 0.0175741203, -0.0216082819, 0.0362852067, -0.0113271037, -0.0364531465, 0.0303952992, 0.0471790954, -0.0062374091, -0.0640881211, 0.064363189, -0.0212780591, 0.038451992, 0.0441914015, -0.0290617868, 0.0145466095, -0.0064906804, 0.0164298583, 0.0102426773, -0.0253347382, 0.0816877186, 0.0178288519, 0.0395790562, 0.0153587209, -0.0063621877, 0.0444341153, -0.0480143912, 0.0007020804, -0.0371517316, 0.0504257195, -0.0092814993, -0.0329086743, 0.0223254412, -0.0099563235, -0.0213322826, 0.0419909097, -0.0216514524, 0.0108479857, -0.0383474864, -0.0309024304, 0.0065488606, -0.0049675545, -0.0391094126, 0.0293260161, -0.0510109626, 0.015803108, 0.02676351, -0.0694296956, -0.0422897562, -0.0240686052, 3.028e-7, -0.0741809532, 0.0090365699, 0.0251795575, 0.0367001593, 0.0360606015, 0.0432470553, 0.0099327574, 0.0256298725, -0.019220341, 0.0049916268, -0.0136346593, 0.0243327748, 0.0818542764, 0.0167171787, -0.0802821666, -0.0180865414, -0.0954768881, -0.0206357427, -0.0276375841, 0.0020497255, 0.0314936191, 0.0674776435, 0.0182506386, -0.0504739285, 0.0825833157, 0.0114814136, 0.0094029577, -0.0344542414, -0.0005362374, 0.0168507099, 0.0515873544, 0.035609141, 0.025950104, -0.0181836206, -0.0094654104, -0.0376065224, -0.0088419812, 0.0395664237, 0.0291885938, -0.0787735209, 0.0235953312, 0.0034995864, -0.0048677358, 0.0161479749, 0.0619008653, 0.0041800872, -0.0297216438, -0.0577629358, 0.0055020889, -0.0188701544, 0.0157945156, -0.0216843244, -0.0074573336, 0.0162762124, -0.0158365015, -0.0052819946, 0.0166983232, -0.0636825785, -0.001430257, -0.0107015269, -0.0659138113, -0.057037428, -0.0416239016, -0.0029998259, 0.0337957777, -0.0078212386, 0.0186734404, 2.418856052e-34, 0.0297720786, 0.0215634927, 0.0261789728, 0.0711520165, 0.0177611429, 0.0023938171, 0.0763394013, 0.0059383428, 0.0350028723, -0.0207803752, -0.0370906554 ]
http://arxiv.org/abs/2406.18125v1
http://arxiv.org/pdf/2406.18125v1
ResumeAtlas: Revisiting Resume Classification with Large-Scale Datasets and Large Language Models
2024-06-26T07:25:18
2024-06-26T07:25:18
Ahmed Heakl, Youssef Mohamed, Noran Mohamed, Ali Sharkaway, Ahmed Zaky
cs.CL
cs.CL, cs.AI, cs.CY, cs.LG
The increasing reliance on online recruitment platforms coupled with the adoption of AI technologies has highlighted the critical need for efficient resume classification methods. However, challenges such as small datasets, lack of standardized resume templates, and privacy concerns hinder the accuracy and effectiveness of existing classification models. In this work, we address these challenges by presenting a comprehensive approach to resume classification. We curated a large-scale dataset of 13,389 resumes from diverse sources and employed Large Language Models (LLMs) such as BERT and Gemma1.1 2B for classification. Our results demonstrate significant improvements over traditional machine learning approaches, with our best model achieving a top-1 accuracy of 92\% and a top-5 accuracy of 97.5\%. These findings underscore the importance of dataset quality and advanced model architectures in enhancing the accuracy and robustness of resume classification systems, thus advancing the field of online recruitment practices.
ResumeAtlas: Revisiting Resume Classification with Large-Scale Datasets and Large Language Models: The increasing reliance on online recruitment platforms coupled with the adoption of AI technologies has highlighted the critical need for efficient resume classification methods. However, challenges such as small datasets, lack of standardized resume templates, and privacy concerns hinder the accuracy and effectiveness of existing classification models. In this work, we address these challenges by presenting a comprehensive approach to resume classification. We curated a large-scale dataset of 13,389 resumes from diverse sources and employed Large Language Models (LLMs) such as BERT and Gemma1.1 2B for classification. Our results demonstrate significant improvements over traditional machine learning approaches, with our best model achieving a top-1 accuracy of 92\% and a top-5 accuracy of 97.5\%. These findings underscore the importance of dataset quality and advanced model architectures in enhancing the accuracy and robustness of resume classification systems, thus advancing the field of online recruitment practices.
[ 0.07708361, 0.0724467635, -0.0141169736, -0.0418792516, -0.0144148571, -0.0076582255, 0.0064482698, 0.031351652, 0.0182192773, -0.0748523846, 0.0108159399, -0.0397926793, 0.0165927447, -0.0344133675, -0.0087752901, 0.0174208358, -0.0017559879, 0.0306077208, -0.0113873184, 0.0075125573, -0.0320330523, 0.0835129544, 0.0141774965, 0.0845582411, 0.0133139417, -0.0275614168, 0.0461424328, 0.023233043, 0.0235203654, 0.0334212333, 0.0607751831, 0.0487110056, 0.0013408862, 0.0375973433, 0.0000016445, -0.0350695886, 0.0045176046, 0.021321075, -0.0595995337, -0.0362647995, 0.044819098, 0.0116575072, 0.0192066226, 0.0375566632, -0.0287801046, 0.0127738323, 0.0283907466, 0.0000553594, 0.0478639901, 0.0532665178, 0.0065940791, -0.0403595902, 0.0149307931, -0.0135811409, 0.0233756583, 0.0069253868, 0.0516319983, -0.0901454613, 0.0101486687, -0.0125839226, 0.0198007692, 0.0006182001, -0.0773367062, -0.0596111231, 0.0101761576, 0.0078692846, -0.0538816378, -0.0143817542, 0.0347672105, 0.0105268536, -0.0200517718, -0.0357611515, -0.0196959078, -0.0297796093, 0.0355181545, 0.0257023405, -0.0125379264, 0.0365112387, -0.0450215973, -0.0167749319, -0.0241164006, -0.0065687858, 0.0156148449, 0.028177578, 0.0055276346, 0.035886392, -0.0178902969, -0.024807876, 0.0441610478, 0.0154093793, 0.0705379471, -0.0490544625, 0.0539024547, -0.0013305161, -0.0324396975, -0.0093948338, -0.0227651577, -0.0095778313, 0.0528116189, -0.0485737808, -0.0115029672, -0.0120538566, -0.0359246843, 0.0309113823, -0.0080725159, 0.0259983055, 0.0209250208, -0.0173723344, 0.0073001957, -0.012631014, -0.1131870821, -0.0289330408, -0.0196081661, 0.0066140071, 0.0219911505, 0.0396528617, -0.0793964937, 0.0139601193, -0.0335741453, 0.0062694033, 0.0565937422, 0.010652123, -0.0168203562, 0.0323723592, -0.0603781044, -0.0418934338, 0.0633598492, -0.0026746965, -0.0272850636, -0.052237317, 0.0650736019, -0.0075217239, 0.0217477176, -0.0314133428, 0.0322017819, 0.0079276701, 0.0481641777, -0.0317621008, -0.0364521891, -0.0225843024, 0.0266205296, -0.0463854447, -0.0453865752, -0.0118193179, 0.0357884951, 0.0038147795, -0.0193186663, -0.0551080257, 0.0331116319, 0.0407162644, -0.0061299801, 0.0342317037, 0.0071169389, 0.010306525, 0.0760439038, -0.026547147, -0.0091300383, 0.0251969676, -0.0010038512, 0.0784483999, 0.04099999, -0.0318783186, 0.0615539029, 0.0632092282, -0.0285947323, -0.034652181, -0.025116479, 0.0704069659, -0.0056784167, 0.0140807265, 0.013614485, -0.0001952965, -0.0117842136, 0.0029305785, 0.0437153056, 0.0540285967, -0.0082908254, 0.0206346586, 0.0364794023, 0.0619014911, 0.0034805438, -0.0739533231, -0.0118288854, -0.0143007506, -0.0158437286, -0.0506333448, -0.0374540389, 0.0323402397, -0.0348284766, -0.0072458023, -0.0194992535, -0.0379355736, -0.0209021736, 0.0312764943, 0.0036553475, -0.006246774, -0.0330362804, 0.0085776122, 0.0264025889, -0.0389563851, -0.0126954783, -0.0608898588, -0.0069690179, 0.0780083686, 0.0464158021, -0.0332653634, -0.0389366932, -0.076881513, -0.0152183129, 0.0114617208, 0.0549118668, -0.0254142024, 0.0266140103, -0.0288344268, -0.0270601101, 0.0593639687, -0.0099157514, 0.013814101, 0.0309915151, 0.0223261267, 0.032706961, -0.0273974538, -0.0493551791, 0.0406938903, 0.0043889582, -0.0325991213, 0.0712075382, 0.0447423384, -0.0115993628, 0.0162267014, -0.0074320962, -0.0032579985, -0.0067401682, -0.0055232085, -0.0202330519, -0.0341718234, 0.0351369195, 0.0072165374, -0.0206804518, 0.0491166636, -0.0552546158, 0.0350811034, -0.0302447416, 0.0059410469, -0.0409248173, 0.0111608561, 0.0253831986, -0.0614010245, 0.0279366616, -0.0426475406, 0.0302933156, -0.0455610342, 0.0654507875, 0.0400937498, 0.0341591872, -0.0063934745, -0.0675677732, 0.0130473124, -0.1199088097, -0.0092332829, -0.0179551914, -0.0208705254, 0.0104989847, -0.0039742752, 0.0078535061, -0.028593732, -0.0444171093, -0.0799847543, 0.0352074169, 0.0256548356, 0.0233423878, -0.0339502394, -0.0023435103, 0.0060785022, -0.004931896, -0.0468190014, -0.0201174356, 0.0070950417, 0.0210243613, 0.0347952917, -0.0396254808, -0.0349316485, -0.0181827508, -0.0337376744, -0.0012007087, 0.0363801979, 0.0203074925, 0.002072752, -0.0577390753, -0.0326668769, -0.0168197583, 0.0073801666, 0.0327976272, -0.0174463373, -0.0493810885, -0.0152072171, -0.0337271504, 0.0419559702, -0.0029307383, 0.0175158381, 0.0207417067, 0.0609015748, 0.0066564865, -0.0051303934, 0.0085933954, 0.0129506495, -0.0312469341, -0.0422250442, -0.0345822684, -0.0484764352, 0.0191664975, -0.0390079431, -0.0101407785, 0.0263935477, 0.0045951786, 0.0039359913, -0.0739081427, -0.014315797, -0.0106219705, 0.0115475086, 0.0027561297, -0.0123180468, 0.0090434626, -0.0055168895, 0.0266103353, 0.0069475393, -0.0819947794, -0.0420663245, 0.0596692786, -0.0021679588, 0.0096729416, -0.0064578885, -0.0366783403, -0.0174067412, -0.0292504765, -0.0174774919, -0.045695059, -0.0383495763, 0.0215780232, 0.0165894162, 0.0159263927, -0.0318740159, -0.0678297281, -0.0614786148, -0.0280130934, 0.0286224112, -0.0026401361, 0.0020092647, -0.0813598633, -0.0261266548, -0.0287661813, 0.0392150916, -0.0121525545, -0.0470549166, -0.0096350638, 0.0437949412, -0.0002703127, 0.020822145, 0.002808491, -0.0056673884, -0.0268136635, -0.0032269908, 0.0039177099, 0.0252967197, 0.0091195405, -0.0319445133, 0.0464416817, -0.1019066721, -0.0437856279, 0.0043587624, -0.0388846435, -0.0692790374, 0.0231933743, 0.0723366365, 0.0082266433, -0.0538375452, 0.0061183339, 0.0354759432, -0.054525435, 0.0225978103, 0.0655643046, -0.0002813259, -0.0069542904, 0.0012208741, 0.0274883304, -0.0183652975, -0.0041369344, -0.0106943781, -0.0593056716, 0.0843698755, -0.034857586, -0.0278570615, -0.0125420438, -0.0531957559, -0.0138340043, -0.0291220807, -0.0431842469, -0.0403907113, -0.050774157, -0.0108732944, 0.0111581702, -0.034072455, -0.032125663, 0.0621566698, -0.0230819993, 0.0078910338, 0.0376185887, 0.0695437044, -0.0258197132, -0.0332807004, 0.0367858633, -0.0129098529, -0.015750587, 0.0367077552, 0.0487551689, -0.0122565608, 0.0017167539, -0.0040672892, -0.0284549911, -0.0365518481, 0.027125949, 0.0367419124, 0.0815289691, -0.0356759131, -0.0149877584, -0.0274448451, -0.0355940685, 0.0313679352, 0.0752978623, -0.0318198279, -0.0939092636, -0.0020202703, -0.0054091113, 0.0066877007, -0.0040498232, -0.0033561552, 0.0451727435, -0.0202885922, 0.0315871425, -0.0091764741, -0.0426844396, 0.0011044715, -0.0172533523, 0.0251164827, 0.0358743258, -0.0119556105, 0.0055406648, -0.0513303168, 0.0275299754, -0.0246338043, 0.0702169612, -0.044981163, -0.0040973919, -0.0355425589, -0.0560169145, -0.0561506301, -0.0730377808, -0.0602108315, 0.019345317, -0.0262332689, -0.0261391085, 0.0194302443, -0.0228415914, 0.0028449653, -0.0816514269, 0.075087361, 0.0551520139, -0.0140016917, -0.0179263521, 0.0080908872, 0.0077709267, -0.0493435673, 0.0313189812, -0.0379446708, -0.0015805071, 0.0334363058, -0.0602631308, 0.042858839, -0.011945121, -0.0237421691, 0.0133896898, 0.0132134473, 0.0175015721, -0.0354439244, -0.0548627079, 0.0313639268, 0.0012739617, -0.0298514199, -0.0102191186, -0.0177471135, 0.0827161744, 0.0171803776, -0.0122922854, -0.0235101636, -0.0161983427, 0.053119082, -0.0660332739, 0.040443942, -0.0298679564, 0.0593493022, 0.0175285581, 0.0025072426, -0.0273089167, 0.025342511, 0.0494959354, 0.0204370879, -0.023859717, -0.0076586194, -0.0288563352, 0.0370924622, -0.0102802888, 0.0328063443, -0.0125269881, -0.0581302606, 0.0031909815, 0.0311714578, 0.0163287297, -0.0587410629, 0.0148018664, -0.0090392139, -0.0135359, -0.0384740792, -0.0007473134, 0.0375025161, -0.0126505811, -0.0028663916, -0.0064063552, -0.0220537204, 0.040546976, 0.0139072137, 0.0248059649, 0.0183812473, -0.03267207, -0.0226848833, -0.0302132517, -0.008134447, 0.073235862, 0.036830537, -0.0178225655, 0.0197101925, -0.0401826836, -0.0359485932, 0.0383032188, -0.0382933691, -0.0138860522, 0.0134639284, 0.0189532433, 0.0411058515, 0.0075618192, -0.0033008778, -0.0831606612, 0.0054699774, -0.0125844413, 0.0094660409, -0.061737746, -4.962880211e-33, 0.0197722781, -0.0107975286, 0.0112537453, 0.0193401314, 0.0130347582, 0.0246167462, 0.0713646561, -0.0638416335, 0.0099073965, 0.0080993371, -0.0389104001, 0.0224842206, 0.0143954437, 0.0530648082, 0.0267209206, 0.0421146601, 0.0403262228, 0.0110821938, -0.0124655729, 0.0081318486, 0.0191874523, 0.0429796726, 0.0431848653, -0.0525136478, 0.0306133702, -0.0026685407, -0.0213956852, 0.0441343002, 0.008666859, 0.084726207, -0.0111850752, 0.0392722487, 0.0454399884, 0.0077364524, 0.0530812182, 0.0350130424, -0.0318542011, 0.0065820538, -0.0129491398, 0.042426534, 0.0421510972, -0.0281774979, -0.0051494502, -0.0255744811, -0.0136700673, -0.0108626699, 0.083127223, -0.0039603505, 0.0051414645, -0.0222869124, 0.0084234392, 0.0226895008, -0.0343580991, 0.0103361299, 0.0141922729, 0.036176119, -0.0068266806, 0.0572289601, -0.118330203, -0.0071969219, -0.0418592766, 0.0113441423, -0.0289486498, -0.0126951719, -0.0419373401, 0.0499310046, 0.0728637055, 0.037066143, 0.0372406989, -0.0235747378, -0.0159951355, 0.0422528163, -0.0124232881, -0.0019621691, 0.0236355066, -0.0529432073, -0.0356919877, 0.0320405401, 0.0295520872, -0.000251453, 0.014730446, -0.005126643, -0.0246147811, 0.0088520283, -0.0104877139, -0.0351843871, -0.0115660485, 0.0418788604, 0.0343567096, -0.0104096066, 0.0219792537, 0.0063905716, -0.032306876, -0.0406229049, -0.0203737523, 0.0327049531, -0.019091988, -0.0131694013, 0.0191302579, -0.0033460085, -0.0275241695, -0.0239652377, -0.0111112837, -0.042694293, 0.0253842268, 0.0484412387, 0.0534607507, 0.0454282053, -0.0763034225, -0.0052065128, 0.0150036085, 0.0246156827, -0.0253690556, 0.0502885506, 0.0377341323, 0.0062632053, 0.0340712108, 0.0346282534, 0.0497087538, 0.0390627049, 0.0062482147, -0.0856335536, 0.0356162302, 0.0042868629, -0.0002049299, -0.0070865974, -0.0292843077, 0.0208242964, 0.0197644681, -0.0328184776, 0.0175098386, -0.0289859045, 2.238e-7, 0.0452797636, 0.0260626432, -0.0713976845, 0.0025102831, 0.0329795144, -0.0008585604, 0.0056396616, 0.0355271101, 0.0217454117, 0.051150281, 0.0022502469, -0.0044040643, 0.0342289284, -0.0251483526, -0.0877459496, -0.0142313084, -0.0233023539, -0.0276626088, -0.0582424887, 0.0327691585, 0.0781655163, 0.0907522365, 0.0859027281, -0.0602546856, -0.0346714035, -0.0109125869, 0.0236303341, 0.0010042348, -0.0114236409, 0.0430402532, 0.0708121434, 0.0057513998, 0.0030562845, 0.0557001121, -0.047298789, 0.0378251523, 0.0369826667, 0.0540021919, -0.0445369706, 0.0476588421, 0.0294287968, -0.0104031954, -0.0082574077, -0.0113414321, 0.0604013167, -0.0163246654, -0.0098984083, -0.0124168396, -0.0314250849, -0.0155894794, 0.0550047271, 0.0119920606, -0.0121232802, 0.0136262858, 0.0439010821, -0.0329218172, -0.065695174, -0.0668835938, 0.0094823921, 0.0086231157, -0.0325614698, -0.0301640891, 0.0468199737, 0.0101342313, 0.0372786149, -0.0333658569, 0.0133068748, 1.974271461e-34, 0.0178076383, 0.0334892832, 0.0115957074, -0.0156136788, 0.0340426601, 0.0129517922, 0.0638634413, 0.0131323617, -0.0292987451, -0.0330784358, 0.00513697 ]
http://arxiv.org/abs/2406.18122v1
http://arxiv.org/pdf/2406.18122v1
Poisoned LangChain: Jailbreak LLMs by LangChain
2024-06-26T07:21:02
2024-06-26T07:21:02
Ziqiu Wang, Jun Liu, Shengkai Zhang, Yang Yang
cs.CL
cs.CL, cs.AI
With the development of natural language processing (NLP), large language models (LLMs) are becoming increasingly popular. LLMs are integrating more into everyday life, raising public concerns about their security vulnerabilities. Consequently, the security of large language models is becoming critically important. Currently, the techniques for attacking and defending against LLMs are continuously evolving. One significant method type of attack is the jailbreak attack, which designed to evade model safety mechanisms and induce the generation of inappropriate content. Existing jailbreak attacks primarily rely on crafting inducement prompts for direct jailbreaks, which are less effective against large models with robust filtering and high comprehension abilities. Given the increasing demand for real-time capabilities in large language models, real-time updates and iterations of new knowledge have become essential. Retrieval-Augmented Generation (RAG), an advanced technique to compensate for the model's lack of new knowledge, is gradually becoming mainstream. As RAG enables the model to utilize external knowledge bases, it provides a new avenue for jailbreak attacks. In this paper, we conduct the first work to propose the concept of indirect jailbreak and achieve Retrieval-Augmented Generation via LangChain. Building on this, we further design a novel method of indirect jailbreak attack, termed Poisoned-LangChain (PLC), which leverages a poisoned external knowledge base to interact with large language models, thereby causing the large models to generate malicious non-compliant dialogues.We tested this method on six different large language models across three major categories of jailbreak issues. The experiments demonstrate that PLC successfully implemented indirect jailbreak attacks under three different scenarios, achieving success rates of 88.56%, 79.04%, and 82.69% respectively.
Poisoned LangChain: Jailbreak LLMs by LangChain: With the development of natural language processing (NLP), large language models (LLMs) are becoming increasingly popular. LLMs are integrating more into everyday life, raising public concerns about their security vulnerabilities. Consequently, the security of large language models is becoming critically important. Currently, the techniques for attacking and defending against LLMs are continuously evolving. One significant method type of attack is the jailbreak attack, which designed to evade model safety mechanisms and induce the generation of inappropriate content. Existing jailbreak attacks primarily rely on crafting inducement prompts for direct jailbreaks, which are less effective against large models with robust filtering and high comprehension abilities. Given the increasing demand for real-time capabilities in large language models, real-time updates and iterations of new knowledge have become essential. Retrieval-Augmented Generation (RAG), an advanced technique to compensate for the model's lack of new knowledge, is gradually becoming mainstream. As RAG enables the model to utilize external knowledge bases, it provides a new avenue for jailbreak attacks. In this paper, we conduct the first work to propose the concept of indirect jailbreak and achieve Retrieval-Augmented Generation via LangChain. Building on this, we further design a novel method of indirect jailbreak attack, termed Poisoned-LangChain (PLC), which leverages a poisoned external knowledge base to interact with large language models, thereby causing the large models to generate malicious non-compliant dialogues.We tested this method on six different large language models across three major categories of jailbreak issues. The experiments demonstrate that PLC successfully implemented indirect jailbreak attacks under three different scenarios, achieving success rates of 88.56%, 79.04%, and 82.69% respectively.
[ 0.0772217661, 0.0781198516, -0.0214172099, 0.0477085449, -0.0623150133, -0.0286847558, 0.0000200419, 0.0063758772, 0.0782114342, -0.0506623536, -0.0145220095, -0.0086236447, -0.0016957085, 0.0151042137, 0.0510025769, 0.0034853353, 0.0594954826, -0.0140391132, -0.0011365126, 0.0088427942, 0.0268203486, 0.004516012, 0.0296331011, 0.0316657834, 0.0009516911, -0.0046364986, -0.0254577342, 0.0081543149, 0.0023530347, -0.0331723243, -0.0005458574, 0.0355086327, -0.002417295, 0.0577642098, 0.0000022337, -0.0383742638, -0.0417383127, 0.0325405374, -0.0375323966, 0.0194705818, 0.0145690469, -0.0431612246, -0.0364999361, 0.0171253663, -0.0528338067, 0.026053967, 0.0228859391, 0.0570926853, 0.0787853822, 0.0707584396, -0.0034670548, -0.0238053743, 0.0575229079, 0.0004908199, 0.0246034302, -0.0396004803, 0.0218947493, -0.0204555802, -0.0019644981, 0.0428849906, 0.0288392715, 0.0101125399, 0.0205774475, -0.0288677812, -0.0343358181, -0.0269315038, -0.0380487442, -0.0367916077, -0.0009408598, 0.0337115042, -0.0144000798, -0.0174213462, 0.0066060871, 0.0245302971, 0.0238662828, 0.0406904742, -0.0231092349, 0.0455849022, -0.0137010096, -0.0076871011, 0.0456498079, -0.036908336, 0.0630761757, -0.001636013, 0.006330939, 0.0306275561, 0.018851975, -0.047741048, -0.0086393319, -0.0597650968, -0.0034391347, -0.0421912894, 0.0094636343, 0.0052245785, 0.0128101539, 0.0395059213, -0.0727821141, -0.0357332677, 0.0436759926, -0.0592260621, -0.0028177623, 0.0322738588, -0.0099132452, 0.0111786416, -0.005665984, -0.0044835554, -0.0302919131, 0.0093153389, 0.0049364073, 0.0008386634, -0.0190829374, -0.0406738371, -0.0105238874, 0.0510132387, -0.020692734, 0.0073869703, -0.0061808145, 0.0298137758, -0.0317825899, 0.0420953222, -0.0510126203, 0.0200916938, 0.0055945772, 0.0314334892, 0.0197383668, -0.0165612958, -0.0012881043, -0.0128388265, 0.0090439282, -0.075698778, -0.0235153642, 0.0056381468, -0.0610969253, -0.002759055, -0.0205324497, 0.109126471, 0.0188520327, -0.0190074276, -0.1096752584, -0.1184396222, -0.0204553548, -0.0220995676, 0.0070565375, -0.0834543034, -0.0336425006, -0.0201297831, -0.0294649638, -0.067441754, 0.0280806553, 0.0285718627, -0.054835733, 0.0397700667, -0.0131455492, -0.0290302094, -0.0080239298, 0.0455240905, 0.0418973565, 0.0734222233, -0.0041164439, 0.0455645211, 0.0629600808, -0.0034002268, 0.0219669063, 0.051104296, 0.0115199117, 0.0155986873, -0.0298936907, -0.0042409096, 0.0356973521, -0.0212786086, -0.0510126911, 0.0410834998, -0.0183399487, -0.0062683374, 0.0784396455, 0.0441236235, -0.0236152299, 0.0529435202, 0.0155270472, -0.0206386112, 0.0068737231, 0.0141422078, -0.0605730787, -0.015080737, -0.1043861955, -0.0300997403, -0.0109350784, 0.0287431069, 0.0088036004, -0.0083562816, 0.0128213502, -0.0009430199, 0.0114472099, 0.0053725406, 0.0017273746, 0.0694397911, 0.0116594881, 0.0721294433, 0.0226156786, -0.0529062636, -0.0265630465, -0.0777899697, -0.02922149, 0.0677064657, 0.0773630738, -0.0129785314, -0.0053933691, 0.0025595382, -0.0473078713, 0.0186403412, -0.0141311456, 0.0056675072, -0.0166014973, 0.0057602837, -0.001319636, -0.0067323679, 0.0402438119, 0.0036804995, 0.0090840319, -0.0009980204, 0.017724622, -0.0194708146, -0.034626212, 0.0098223584, -0.0231633391, -0.0221504234, 0.0214041099, 0.0221032891, 0.0103460075, 0.0470857359, -0.0256384742, 0.0261714142, 0.0810579732, -0.0030757042, 0.0196585841, 0.002906908, -0.0378588401, -0.0247694943, 0.0201418567, -0.0125438366, 0.0152277527, 0.0706783757, -0.0072745085, -0.0272642933, 0.0692517757, -0.0514582805, 0.0605317131, -0.0237577967, 0.0047115767, -0.0099936137, 0.002738684, -0.0343698673, 0.0150845796, -0.0249221474, 0.005531624, -0.0069555282, 0.0238667876, 0.0092399102, 0.0128920535, -0.0370827951, 0.0213868879, -0.0340651646, 0.0172624812, 0.0418725833, 0.0218163002, -0.0263975244, -0.0480436608, -0.0533848293, -0.0176315568, -0.0273631997, 0.0318500176, 0.0401146039, 0.0427473858, -0.0154124806, -0.0205724593, -0.0254707709, -0.0216878336, 0.0142739033, 0.0286354292, -0.0137975542, -0.0092496369, -0.0513714887, -0.0176295992, -0.0009974125, 0.0118876211, 0.0595924556, -0.0269395877, -0.0286087431, -0.0102320854, 0.0149813052, -0.0142653659, 0.0169329792, -0.0172361098, -0.0316777378, 0.0425400622, -0.0115936249, -0.0197291151, 0.0554675795, 0.0123195061, 0.0668274984, 0.0683733299, -0.0214474853, -0.0449043065, -0.0274570063, 0.0327373557, -0.0346266143, -0.0474865362, -0.0328502506, 0.0272046197, 0.0401318744, 0.0486713126, -0.0056477739, -0.0127784945, 0.0356786475, 0.0068841153, -0.0339573286, 0.007911332, -0.0368805751, -0.0469309613, -0.0109990016, -0.0170177482, -0.0852995813, 0.0549422614, -0.0533379912, 0.0336284824, -0.0367556848, -0.0434847735, 0.0013808964, 0.0554516241, -0.0063557909, 0.0279261582, 0.0442056246, -0.0343519635, -0.0341941379, -0.0050846464, 0.0377185643, -0.0368493833, 0.0330708921, -0.0518167578, -0.0040533151, 0.0032851903, -0.0163059235, -0.0790835395, -0.0445907488, 0.0067392723, -0.0060943342, 0.0545866303, -0.0466876067, 0.0016767107, -0.0515269525, 0.0660791472, -0.0057120221, -0.0023656378, -0.0132938828, -0.0331433751, 0.0444872864, 0.0297161285, -0.0700601786, 0.0052110516, 0.0114161475, 0.0565905944, 0.1019054726, 0.0482188947, -0.0057339757, 0.0183968283, 0.0350144729, 0.0184102673, 0.0132893119, -0.025551606, 0.0282003768, 0.022681348, -0.0495254472, -0.0254455972, 0.0567861684, 0.0479512289, -0.0249067992, 0.0070819478, 0.0000314586, 0.0290999319, 0.0240000319, 0.0360476114, -0.0557943508, -0.0162578542, 0.0297690835, 0.065385282, -0.0539907403, -0.0393729508, -0.0505855419, -0.1165702119, -0.0037841324, 0.0285164751, -0.0171356797, 0.011339305, -0.0439329036, -0.0386499651, 0.0557725988, 0.0162839107, -0.0175854955, -0.0006419133, -0.0032326218, -0.0204108246, -0.0555917956, -0.0038257106, 0.009363655, -0.015108509, 0.0507846661, 0.0877701491, -0.0027022602, -0.0371068269, 0.0411278233, 0.0483115017, -0.0212234799, 0.0000653699, -0.010898185, 0.0025682973, 0.0285223536, -0.0010210797, -0.005148916, 0.0553183071, 0.0010056606, -0.0424784906, 0.0207427219, 0.0511804335, -0.0123773227, 0.0104396259, -0.0237244107, -0.012296618, 0.0409323312, 0.0410162657, 0.0223232768, -0.0378281288, -0.0147736566, -0.0812593922, -0.050318189, -0.0290934965, -0.0493145324, 0.0376231447, 0.0188691244, -0.012350197, -0.0672272891, 0.0097926473, -0.0323055424, 0.0404012129, 0.0475475974, 0.0123333773, 0.0142145827, -0.0157763641, -0.0216623768, -0.0117614018, 0.0008312311, 0.0222911704, -0.0501822568, 0.0182964765, -0.0033407791, -0.0295078047, 0.0052130711, 0.031672135, -0.0420566835, -0.0908769071, 0.0141396103, 0.0519942679, 0.020958595, 0.026414996, -0.0273805466, -0.0477004424, 0.0433322899, 0.0853249505, -0.0755785927, 0.0157614965, 0.0507752672, -0.0107863648, 0.0126800323, 0.0259576626, -0.0563982651, -0.0181053989, 0.0459444113, 0.0056549981, -0.0251953956, -0.0061363797, 0.035940811, 0.0656976402, -0.0046386542, 0.0150762396, 0.0198927391, 0.0160139278, 0.0188209135, -0.0229869243, -0.0486700386, -0.0122608375, 0.0266488846, 0.0044105179, -0.0379691496, 0.0394411236, 0.0252220184, 0.0033042936, -0.0013123502, 0.0023265511, 0.011610223, 0.0218192432, 0.0259122774, 0.0710794479, 0.0209509786, -0.0241183639, 0.0414539576, -0.0281926394, 0.0416890308, -0.0250409693, 0.0476750135, -0.0122832246, -0.0267335903, -0.0222050976, 0.0358532444, -0.0818461627, -0.0291228704, 0.0119659249, 0.0660112053, 0.0016701728, -0.0136453342, 0.0345105156, -0.0248447992, 0.0151807936, -0.0320056044, -0.0234607421, -0.017943209, -0.0822990909, 0.0044725649, -0.0004762195, -0.0133011397, 0.0401517525, 0.0608471334, -0.0052873129, 0.053911332, 0.0154066775, -0.021574866, -0.0370398425, -0.0081908032, 0.0586750992, 0.0179794785, -0.0497716218, -0.0296848193, -0.042868115, -0.0756336078, -0.0156843793, -0.010625666, 0.0155556705, 0.0022385367, -0.0232369956, 0.0645938069, 0.0138030965, 0.0266291443, -0.0666710436, 0.0402312018, -0.0124853877, -0.0692145228, 0.0334314704, -6.170175725e-33, 0.0166801475, -0.0530413873, -0.0352102295, -0.0321737938, 0.0040254318, -0.0112281572, -0.0254024938, 0.0164133459, -0.0201014541, -0.0272017792, -0.0220444091, -0.0028016367, -0.0064955242, 0.0325212441, -0.0319631323, -0.0125184851, 0.0100556621, 0.0116798896, -0.0026493885, 0.0502736568, 0.0911742523, 0.0166339632, 0.0157141928, -0.049533233, 0.0591237508, -0.031827379, -0.023278594, 0.0300790258, -0.0475239679, 0.0638480559, 0.0131624136, 0.043273814, 0.0002638286, -0.0415198281, -0.0147751449, 0.0520897694, -0.0010968641, -0.0089384653, 0.0125291338, -0.0064061191, -0.0047254106, -0.047295846, 0.0408351384, -0.0116914054, 0.0044681537, 0.0179749671, -0.0126246968, -0.0223689936, -0.0198005997, -0.0185992736, -0.0187381487, -0.001821203, -0.0303212125, 0.0028085548, -0.0125679979, 0.0058885464, 0.0259090737, -0.0212643128, -0.040113505, 0.036054723, 0.027554024, 0.0239538532, 0.0236423444, -0.0183312204, 0.0058662235, 0.0631839857, -0.0316084847, -0.01774293, -0.0029101009, 0.0112621346, -0.002275375, 0.0662681311, 0.0461524092, 0.066064015, 0.0540571325, -0.032935068, -0.0612070151, -0.0114771901, 0.0022875087, 0.0427582264, 0.0202731248, -0.0168095082, -0.0330618471, -0.010016541, -0.0242340267, -0.0081276549, -0.0326703824, -0.0431870036, 0.024996398, 0.0132046258, -0.0186409764, -0.0729283914, -0.0072033093, -0.0764710456, -0.012167302, 0.0402573943, -0.0230440721, -0.0266731549, -0.0187382232, 0.0211190283, 0.0041496279, 0.0111414036, 0.0159638561, -0.0692028478, 0.016852593, 0.0089076413, 0.0305685233, 0.016918784, -0.0785080567, -0.0097021991, -0.0133646009, 0.0665919408, 0.0193715896, 0.0027189201, 0.0232425127, -0.0101757888, -0.0024624912, 0.0145754414, 0.0439669862, 0.0113819744, -0.0551995747, 0.0143680489, -0.0261128396, 0.0293659326, -0.011298175, 0.019699201, -0.0139532043, 0.0337553285, 0.042889718, -0.0521861017, 0.0080925561, -0.0072620464, 3.095e-7, -0.0250025988, 0.0017371436, 0.0160642453, -0.0096491771, 0.0386145003, 0.0637819171, 0.0035377408, 0.0556055009, -0.0272361431, -0.0764526948, -0.0692582354, -0.0310749933, 0.0627902523, 0.0451108851, -0.0866604671, 0.0449500531, -0.066691339, -0.0194385238, -0.0586316735, 0.0027045142, 0.0511183254, 0.0786313117, -0.0125252632, -0.0399244837, -0.0260210689, -0.0715056583, -0.0018184468, 0.0102185225, -0.041265212, -0.0334833488, 0.0511624999, 0.0447704941, 0.026198281, 0.0097212251, -0.0540025383, -0.0058379686, -0.0689499304, 0.0080903238, 0.008843231, -0.0074013313, -0.022215005, -0.0395386666, -0.0106105246, -0.0292011872, 0.0828898996, 0.0168857481, -0.0085507818, -0.0148599604, 0.0257914104, -0.0473541394, -0.0112010315, -0.0449175723, 0.0005589658, 0.0119685531, 0.0232180152, 0.0148509452, 0.0406767093, -0.0679224581, -0.0159299057, 0.0027768924, -0.0133468434, -0.0131393885, -0.0283211097, 0.0238071568, 0.0102825519, -0.0332824588, 0.0086727422, 3.26486829e-34, -0.0063121673, -0.0422377586, 0.031441085, 0.0056973309, 0.0371808633, 0.0109885763, 0.0047875852, 0.0374860168, 0.0077152415, -0.0727372169, -0.0285701081 ]
http://arxiv.org/abs/2406.18120v1
http://arxiv.org/pdf/2406.18120v1
ArzEn-LLM: Code-Switched Egyptian Arabic-English Translation and Speech Recognition Using LLMs
2024-06-26T07:19:51
2024-06-26T07:19:51
Ahmed Heakl, Youssef Zaghloul, Mennatullah Ali, Rania Hossam, Walid Gomaa
cs.CL
cs.CL, cs.AI, cs.CY, cs.LG
Motivated by the widespread increase in the phenomenon of code-switching between Egyptian Arabic and English in recent times, this paper explores the intricacies of machine translation (MT) and automatic speech recognition (ASR) systems, focusing on translating code-switched Egyptian Arabic-English to either English or Egyptian Arabic. Our goal is to present the methodologies employed in developing these systems, utilizing large language models such as LLama and Gemma. In the field of ASR, we explore the utilization of the Whisper model for code-switched Egyptian Arabic recognition, detailing our experimental procedures including data preprocessing and training techniques. Through the implementation of a consecutive speech-to-text translation system that integrates ASR with MT, we aim to overcome challenges posed by limited resources and the unique characteristics of the Egyptian Arabic dialect. Evaluation against established metrics showcases promising results, with our methodologies yielding a significant improvement of $56\%$ in English translation over the state-of-the-art and $9.3\%$ in Arabic translation. Since code-switching is deeply inherent in spoken languages, it is crucial that ASR systems can effectively handle this phenomenon. This capability is crucial for enabling seamless interaction in various domains, including business negotiations, cultural exchanges, and academic discourse. Our models and code are available as open-source resources. Code: \url{http://github.com/ahmedheakl/arazn-llm}}, Models: \url{http://huggingface.co/collections/ahmedheakl/arazn-llm-662ceaf12777656607b9524e}.
ArzEn-LLM: Code-Switched Egyptian Arabic-English Translation and Speech Recognition Using LLMs: Motivated by the widespread increase in the phenomenon of code-switching between Egyptian Arabic and English in recent times, this paper explores the intricacies of machine translation (MT) and automatic speech recognition (ASR) systems, focusing on translating code-switched Egyptian Arabic-English to either English or Egyptian Arabic. Our goal is to present the methodologies employed in developing these systems, utilizing large language models such as LLama and Gemma. In the field of ASR, we explore the utilization of the Whisper model for code-switched Egyptian Arabic recognition, detailing our experimental procedures including data preprocessing and training techniques. Through the implementation of a consecutive speech-to-text translation system that integrates ASR with MT, we aim to overcome challenges posed by limited resources and the unique characteristics of the Egyptian Arabic dialect. Evaluation against established metrics showcases promising results, with our methodologies yielding a significant improvement of $56\%$ in English translation over the state-of-the-art and $9.3\%$ in Arabic translation. Since code-switching is deeply inherent in spoken languages, it is crucial that ASR systems can effectively handle this phenomenon. This capability is crucial for enabling seamless interaction in various domains, including business negotiations, cultural exchanges, and academic discourse. Our models and code are available as open-source resources. Code: \url{http://github.com/ahmedheakl/arazn-llm}}, Models: \url{http://huggingface.co/collections/ahmedheakl/arazn-llm-662ceaf12777656607b9524e}.
[ 0.0262527578, 0.0490271971, -0.0210837796, 0.0538118295, -0.055967439, 0.0387913026, 0.038952142, 0.0316529311, -0.016650388, -0.050089296, -0.0571307763, -0.027436018, 0.0602046438, 0.0246781223, 0.0542993881, -0.00724778, 0.0215229988, -0.1006459892, -0.0238489602, 0.0045238519, 0.0314570703, 0.0303136557, -0.0215213113, 0.0067841737, 0.0183318574, -0.0018771151, -0.0282401182, 0.0314758681, 0.009777924, 0.035700541, -0.0264585502, 0.0112458291, -0.0089516854, 0.0520313047, 0.0000020932, -0.0516194664, -0.0434283577, -0.0212647263, -0.0296473522, -0.0156275835, 0.0642890856, -0.0527811795, -0.0441457517, 0.0123044308, -0.0234648753, 0.029096676, 0.0566433296, -0.0031813579, 0.0422043242, 0.0650272667, -0.0004805795, -0.0618168637, 0.0054223617, 0.0007500968, -0.0212130267, -0.0704377666, 0.0096350648, -0.0264314935, -0.0293208044, 0.0115875136, -0.0231632572, 0.0658373162, -0.0286376849, -0.0360164829, -0.0249706693, -0.0257316604, -0.0369495191, -0.0464946926, -0.0320188031, 0.0326999947, -0.0349319465, -0.0181692876, 0.0316563137, -0.0061647776, -0.0317939371, -0.0419376791, 0.0001925464, 0.0217250679, -0.025759032, -0.0043022763, -0.0000683831, 0.0455474518, 0.0180224013, -0.0158078317, -0.0552060939, -0.0140126264, 0.0079175187, -0.0319909006, 0.0174937323, -0.0290297996, 0.0339576639, -0.0218780525, 0.0741899461, 0.0175787713, -0.0169988275, -0.0097710695, 0.0018363973, -0.0338736326, -0.0153901521, -0.0563369095, 0.0266439393, 0.0930487141, -0.0380328223, 0.0317319073, -0.0291113853, 0.0630417317, 0.0090125874, 0.0127707887, -0.0367682762, 0.0329814889, -0.0149497073, 0.033960674, -0.0434847027, 0.0079962481, -0.0468835235, -0.0211269595, -0.0411083326, -0.0143108936, 0.0246604979, 0.0251681115, 0.0084155397, -0.0704152137, -0.0213590339, -0.0103154676, 0.0320355594, -0.0150198201, 0.0389754884, 0.0123123918, -0.0361878462, -0.0680051073, 0.0186286792, 0.0202380978, 0.0091220019, 0.045717936, -0.0228785779, 0.0805774257, -0.0335645825, -0.0247664079, -0.0542191938, 0.0376052223, -0.0128736943, -0.0486669391, -0.0289051645, -0.0307041667, 0.0475138761, 0.0277025867, 0.0149027361, -0.0397808179, -0.0015351108, -0.0149892205, 0.0455989279, 0.0846020505, 0.0311356392, 0.0223107226, 0.0797003359, 0.0032474895, -0.0132089779, 0.0884617344, 0.0688913912, -0.0177942496, 0.0024520305, -0.0042422223, 0.0015978313, 0.0750812292, -0.02495021, 0.0062695481, -0.0127665093, -0.0359948501, 0.0720910206, 0.0460315458, -0.0151880216, 0.0644063354, -0.0218111631, -0.0339938216, 0.0225670859, 0.0332864039, -0.0125960689, 0.0540551208, 0.0185240675, 0.0237595867, 0.0311290286, -0.0085143978, 0.0386804119, -0.006971465, 0.0014051354, -0.0225165822, 0.0423415154, 0.0540704541, -0.0182343181, -0.034014836, -0.0163848605, -0.020108141, 0.0385903381, -0.0071963225, 0.0545155741, 0.0044955239, -0.0678892583, -0.008487328, 0.0081335753, -0.0396737717, -0.0010584086, -0.0754862204, -0.0554154254, 0.0633116215, 0.0123857055, 0.0045179874, 0.0278050639, -0.0289334357, -0.000158976, 0.0589895174, 0.0217928793, -0.0009861892, 0.0383471437, -0.0747429654, 0.0112845954, 0.0595230311, -0.0153285004, 0.0255395379, 0.0051307273, 0.0560206696, 0.0163320899, 0.0261902716, -0.0259866994, 0.0095373001, -0.0755669251, 0.0272377115, 0.0090316441, 0.0060324124, 0.0244572386, 0.0278785974, -0.0187718309, 0.0174104683, -0.0060188086, -0.0330416895, 0.007215919, 0.0308673661, -0.0300305076, 0.0026762949, 0.0339286327, 0.0160060767, 0.0334593616, -0.0017752, 0.0141227199, -0.0702342987, 0.0716628209, -0.0571600832, 0.0754896328, -0.0054071001, 0.0998697951, 0.010202935, -0.0044940729, -0.0079127923, -0.0087454161, -0.0016569917, -0.0307303537, -0.0041526873, -0.0226115044, 0.0116296178, -0.0209475048, 0.0107203014, 0.0427328721, -0.0093473131, 0.0221382789, 0.0131503642, 0.0029000817, -0.0162578393, -0.0815543532, -0.0464226902, -0.0206457246, -0.0243363306, 0.041281525, -0.0300016645, 0.0247667581, -0.0209434889, -0.0119410837, -0.0335014202, -0.0020660951, -0.0194675233, -0.0163356978, -0.005732147, -0.0177565813, -0.0474830456, -0.0351736322, 0.0188314002, 0.0279346853, 0.028129803, 0.0315615013, -0.0483504646, -0.065223515, 0.0033556779, 0.0167296194, -0.0038225842, -0.0049563604, -0.0253565516, 0.0307970215, -0.0213011056, 0.0179605186, 0.0031659072, -0.0191185866, 0.0067483941, -0.0091026062, -0.0220977012, -0.066867575, -0.0324121043, 0.0188697688, -0.0363054425, -0.0495116748, -0.0431493483, -0.0315471329, 0.0292138904, 0.0183576476, 0.0346880667, -0.0468669385, -0.0157287885, -0.0219465811, -0.0097295688, 0.0413544029, -0.0088452771, -0.0006420349, 0.0925773308, -0.0066780266, -0.0319088809, 0.0021999339, -0.0623499341, -0.0099556958, -0.0330291502, 0.0097281989, -0.0518929698, -0.0152598144, -0.0616928488, 0.017731091, 0.0124419341, -0.0730082318, 0.0078800032, -0.005822666, 0.0014407554, -0.0439390242, -0.0156158209, -0.0456580147, 0.0288018044, -0.0103687514, -0.0319793932, 0.0039763441, 0.0278938692, -0.0680010542, -0.0079054739, 0.0472900756, -0.0357431695, -0.0289998986, 0.0548538081, -0.0134495599, 0.0485755876, 0.0225328486, 0.0240553431, -0.0057046302, -0.011641467, -0.0319792069, -0.0018330259, -0.0145999212, -0.0083007915, 0.0483948439, 0.0131278988, -0.005298323, 0.0050937603, -0.0736413598, -0.0118070478, 0.0750631914, -0.0180515982, -0.0209068395, -0.0342428461, -0.0056303386, -0.0177124199, 0.0325865597, 0.082924448, -0.0013967323, -0.0585382916, -0.0208970066, -0.0276640337, 0.0079722349, -0.0008932949, -0.0180269666, -0.0949183926, 0.0118412133, 0.0530186929, 0.051575873, 0.0183644146, 0.0377237611, 0.0182843413, -0.0464012958, 0.0166658983, 0.0447716527, -0.0863273963, -0.0582080483, -0.0113573167, -0.0140602831, -0.0276112948, -0.0192429647, -0.0142224235, -0.0037777962, -0.0207170751, 0.0479144007, 0.0001137523, -0.0079521956, -0.0137616135, -0.0112984236, 0.0496602096, 0.0589570925, 0.0303552654, 0.0088797053, -0.0383678302, -0.0052652974, 0.0041110413, -0.0188542921, -0.0355295688, -0.0131781437, -0.0148562407, -0.0122856451, 0.0315930396, -0.0093945842, -0.0285399035, 0.0101368148, 0.0113455439, -0.0161008164, -0.0319244973, 0.0288085528, 0.0469679832, 0.0327778533, 0.0086210836, 0.0103536304, 0.0019389209, -0.0256812293, -0.0172718, -0.0270769596, 0.0175294448, -0.0169813037, -0.0424363911, -0.0183489844, 0.0199686792, -0.0193259325, -0.0741087198, 0.0136271995, 0.0058312756, 0.0433011092, 0.0536620878, 0.0277549811, 0.0346730426, 0.0449430682, -0.0256609712, -0.0786715075, -0.0219900049, 0.0198631659, -0.0670393333, -0.0448856242, 0.043343123, -0.0864781961, 0.0324151441, -0.0132520888, -0.1108837426, -0.0611112863, 0.0218080897, -0.0000347052, 0.0294723511, 0.0142257269, -0.0229232218, -0.0370574482, -0.0116657605, -0.0643396676, -0.0304472428, 0.0228576362, 0.0552840978, 0.0497291386, -0.03170719, 0.0295382477, 0.009834487, -0.0084186913, 0.0076609184, -0.0014088731, -0.0579329953, -0.0796016082, -0.0302474424, -0.0222646501, -0.0106136706, -0.0109772114, -0.0561327524, -0.0101233199, 0.0387161747, 0.0014108388, 0.0475854464, -0.0316185616, -0.0095199645, 0.0095892502, 0.0318620577, 0.0010190197, -0.027968483, -0.0383842029, -0.0446882136, 0.0074350536, 0.0642526373, 0.0184739903, 0.0112843979, 0.0211162996, 0.0191734079, -0.00572378, 0.0342130885, 0.005621505, 0.0117233712, -0.0191001277, 0.0124599244, -0.0058847712, 0.013396265, 0.0439394042, 0.0450988822, -0.0146306632, -0.050275702, -0.0102154026, 0.0653291866, 0.0281372499, 0.0273133237, 0.1066895053, 0.0369688645, 0.0158271585, -0.0071226647, 0.0469500311, -0.0214320272, -0.0133715803, -0.0354803838, -0.0056756749, -0.0069131046, 0.0663688406, 0.0165334176, -0.0034239066, -0.0154505735, 0.0290556438, 0.0064943205, -0.0133903315, -0.0151393609, 0.0380673409, -0.0436561778, -0.0352335833, -0.0138166239, -0.0161393434, -0.0703889877, 0.0498460978, 0.021218149, 0.0350361243, -0.0256945025, 0.0100957677, 0.0750859976, 0.0023932979, 0.0197061133, -0.0593195073, -0.0068306662, 0.0415423252, -0.0626683831, 0.0025645862, -5.727043007e-33, 0.0150479125, -0.0223798715, 0.0037884505, 0.0393803045, -0.0057604727, -0.0350789241, -0.0354388729, 0.0069612288, 0.0076027303, -0.0613668934, -0.0352018848, 0.0158744883, 0.0430649407, -0.0297256466, 0.035126362, 0.0124088116, 0.0442650132, -0.0126887299, -0.0268464219, 0.0355669372, 0.0206084307, 0.0423748903, 0.0479632467, -0.0574351065, 0.0115088783, -0.0377749838, -0.0348636918, -0.0248650499, 0.0027839167, 0.0551919527, -0.0172473639, 0.0077809249, -0.0095995199, 0.0029071884, 0.0068581421, 0.0011925278, -0.0435031205, 0.064441368, 0.0189677048, 0.0891254842, 0.051880043, -0.0464759432, -0.0016031009, -0.0322374403, -0.0698117688, 0.0370840095, -0.0199582893, -0.0420776717, -0.0183026195, -0.0211046822, -0.0995128378, 0.0506205745, -0.0281153973, -0.0254429076, 0.1115036681, 0.0490916371, -0.0245498959, -0.0303785987, -0.0615510419, 0.0508818813, 0.0303204991, 0.0283626709, 0.0535574555, 0.041160807, -0.0674738511, 0.001255648, 0.0540289171, -0.0430820361, 0.0676037669, 0.0100734122, 0.0400425978, -0.0362707935, 0.0033025818, 0.0137242209, -0.0089048957, -0.0397529602, -0.0276300441, -0.0238545518, 0.0201910865, 0.0135106081, 0.0217299201, 0.0469611734, -0.0543274917, -0.0123760747, -0.0303416532, -0.0657065362, -0.0152552985, -0.0147408536, 0.0110226376, 0.0064356127, -0.0181974191, 0.0376372747, 0.0435870998, 0.0070251222, 0.0096967453, -0.0390694365, 0.0111758085, 0.0143083688, -0.0144947618, 0.0649084598, -0.0328481644, -0.0574868843, 0.0352987275, 0.019643819, 0.0477365516, -0.0301363617, -0.0022800048, 0.0619809851, -0.0463192053, -0.0028697136, -0.0090818275, 0.0072451364, 0.011321798, 0.0237121414, 0.0017514363, -0.0315956548, 0.0114072654, 0.0160382427, 0.0002186362, -0.0134720234, -0.0326943435, 0.0270807967, -0.0159081351, 0.0160332676, -0.0641764924, 0.022992909, -0.0223691203, 0.0144351581, 0.0295289773, -0.023060279, -0.020004198, 0.0118339648, 2.893e-7, 0.0084018307, 0.0268840045, 0.0081435712, -0.0165396947, 0.0105347717, -0.0092800241, -0.0221713874, 0.0087591195, 0.039204061, 0.027450718, -0.0114692114, 0.0014205979, 0.1317580193, 0.0065717562, -0.0841206908, -0.0579683036, 0.0250663552, 0.0348102786, -0.038590204, 0.0068432121, 0.0400593691, 0.1147982627, -0.0594107397, 0.0150020178, -0.0498124175, 0.0111840107, 0.0312670283, -0.0429012068, 0.0190105233, -0.0143860113, 0.0063290466, -0.0338350795, -0.0216306373, 0.0012421897, -0.0145095438, 0.0027364669, 0.0683392212, 0.0829960033, 0.0092576519, -0.0041670627, 0.007165215, 0.0079358807, -0.0342313573, -0.0392084531, 0.0275259446, 0.036874827, -0.0040821745, -0.0063139284, -0.0036792797, 0.0304592922, 0.0016238483, -0.0411856137, 0.015261163, 0.0406896919, 0.0527926795, -0.0103358086, 0.0074714259, -0.0206951927, 0.0034445634, -0.0570639633, -0.0003991894, -0.0228083376, -0.0072336458, -0.0116796335, -0.0003665002, 0.0147634698, 0.0002811846, 2.185472209e-34, 0.0503713489, -0.0337103792, -0.002241628, 0.0610278212, 0.0499623865, -0.0208385494, -0.0009144825, 0.053766977, -0.0045741415, -0.0239347015, -0.0272247549 ]
http://arxiv.org/abs/2406.18118v2
http://arxiv.org/pdf/2406.18118v2
SafeAligner: Safety Alignment against Jailbreak Attacks via Response Disparity Guidance
2024-06-26T07:15:44
2024-06-28T06:06:59
Caishuang Huang, Wanxu Zhao, Rui Zheng, Huijie Lv, Shihan Dou, Sixian Li, Xiao Wang, Enyu Zhou, Junjie Ye, Yuming Yang, Tao Gui, Qi Zhang, Xuanjing Huang
cs.CR
cs.CR, cs.CL
As the development of large language models (LLMs) rapidly advances, securing these models effectively without compromising their utility has become a pivotal area of research. However, current defense strategies against jailbreak attacks (i.e., efforts to bypass security protocols) often suffer from limited adaptability, restricted general capability, and high cost. To address these challenges, we introduce SafeAligner, a methodology implemented at the decoding stage to fortify defenses against jailbreak attacks. We begin by developing two specialized models: the Sentinel Model, which is trained to foster safety, and the Intruder Model, designed to generate riskier responses. SafeAligner leverages the disparity in security levels between the responses from these models to differentiate between harmful and beneficial tokens, effectively guiding the safety alignment by altering the output token distribution of the target model. Extensive experiments show that SafeAligner can increase the likelihood of beneficial tokens, while reducing the occurrence of harmful ones, thereby ensuring secure alignment with minimal loss to generality.
SafeAligner: Safety Alignment against Jailbreak Attacks via Response Disparity Guidance: As the development of large language models (LLMs) rapidly advances, securing these models effectively without compromising their utility has become a pivotal area of research. However, current defense strategies against jailbreak attacks (i.e., efforts to bypass security protocols) often suffer from limited adaptability, restricted general capability, and high cost. To address these challenges, we introduce SafeAligner, a methodology implemented at the decoding stage to fortify defenses against jailbreak attacks. We begin by developing two specialized models: the Sentinel Model, which is trained to foster safety, and the Intruder Model, designed to generate riskier responses. SafeAligner leverages the disparity in security levels between the responses from these models to differentiate between harmful and beneficial tokens, effectively guiding the safety alignment by altering the output token distribution of the target model. Extensive experiments show that SafeAligner can increase the likelihood of beneficial tokens, while reducing the occurrence of harmful ones, thereby ensuring secure alignment with minimal loss to generality.
[ 0.0329220779, 0.0562441871, 0.0122979265, 0.0306081865, -0.0424689762, -0.0171145126, -0.0027163352, -0.033442542, 0.0100423135, -0.0243762583, -0.0006120637, 0.007572067, -0.0240653399, 0.0277761165, 0.0386278965, 0.0258413833, 0.0838408694, -0.060035035, 0.0199744534, 0.0106874006, 0.0144717963, 0.0053707263, 0.0245855805, 0.034367308, -0.0239944849, 0.0017294915, -0.0298899896, 0.0537770651, 0.0273449831, -0.0402674861, 0.0156629477, 0.0491223708, -0.0170114506, 0.0388172753, 0.0000021193, -0.0153985238, -0.0447803773, 0.0391249396, -0.024361968, 0.0298238397, -0.0117243836, -0.0705512166, -0.0282790549, 0.0267735384, -0.028120989, 0.0027082432, 0.042649772, 0.0306845177, 0.0705095977, 0.0380660109, 0.0183906276, -0.0004822353, 0.0497639477, 0.0025328221, -0.024642732, -0.060618069, 0.0050314134, -0.0428864025, -0.0254126098, 0.0355854407, -0.0069314893, -0.037832249, 0.0305904504, -0.0101382844, -0.0215337723, -0.00354604, -0.0490322523, -0.0193919614, -0.0443417951, 0.0126206605, -0.0589703098, -0.0396550782, -0.00133047, -0.0025897077, 0.015998235, 0.0029978075, 0.0161807835, 0.0051751635, -0.0070304018, 0.0024866683, 0.0348242484, 0.0007404395, 0.0412392914, -0.0142644923, -0.0299496856, 0.0600263253, 0.0289309863, -0.0229901113, -0.011909768, -0.0458006747, 0.0081086773, -0.0547060557, -0.0062070913, 0.0065487223, -0.0086776856, 0.007232361, -0.0841607004, -0.0294446871, 0.0372677408, -0.0427519158, -0.0604472533, 0.0505342446, -0.0340279303, 0.0015393321, -0.0212483797, 0.0019317359, -0.0131362658, 0.0456484146, 0.0095867999, -0.0040049669, -0.0070587178, -0.0215247367, -0.0017115725, 0.0524851792, -0.0157894362, 0.0075648208, -0.0428446606, 0.0165679827, -0.0093535241, 0.0187210571, -0.0275984239, 0.0101712411, -0.0038833248, 0.0187269542, 0.0578610301, -0.0054937666, -0.0283436179, -0.0203637183, -0.0066661639, -0.0694847107, -0.0168573558, 0.0029246423, -0.089061752, -0.029424144, -0.0197757315, 0.1395939142, 0.0380621701, 0.0108833723, -0.0912547782, -0.0822156817, -0.0333927721, -0.0227069445, -0.0063161151, -0.0766955912, -0.0359712876, 0.0063410271, 0.0049189231, -0.0677118301, 0.008297598, -0.0157797746, -0.060740266, 0.0268949308, 0.0145677645, -0.0257582292, -0.0349443369, 0.0296838842, 0.0267176628, 0.0561588779, -0.0086704604, 0.0544765964, 0.0431738868, -0.033005435, 0.0357667282, 0.0337465815, -0.0125986189, 0.0125030102, -0.0165014938, -0.0082577374, 0.061969243, -0.0437829196, 0.011607185, 0.0338830613, 0.0086622918, -0.0332982875, 0.0721106082, 0.0809944123, -0.0519204587, 0.0609660335, -0.0328350626, -0.0288189631, -0.0161566045, -0.0282549132, -0.0830745101, 0.0176291279, -0.0743246004, -0.0146407317, -0.0116114765, 0.0471760742, -0.0061423052, 0.0219858773, 0.0055771032, -0.0163285322, -0.0134071307, 0.0128427157, 0.0121120512, 0.0076060272, -0.001227924, 0.0708804354, 0.0270143319, -0.0196742099, -0.0385263264, -0.0838713795, 0.0085249962, 0.0672764927, 0.0208406523, 0.0225931238, -0.0023528645, 0.0172095094, -0.0054503004, 0.0122248204, -0.0188162681, 0.0342852436, 0.0454783328, -0.0197703317, 0.0066914051, -0.0046938467, 0.0375085287, 0.0351262949, -0.0027212824, -0.0083279479, 0.0072354483, -0.0143030668, -0.0408875607, 0.0377769843, 0.0315591991, -0.0041116658, 0.0371082649, 0.0117787961, -0.0001359763, 0.0145707596, -0.0230681878, 0.0274138525, 0.0545127466, 0.0056737927, 0.060172841, 0.0041599781, -0.0360743999, 0.0104401838, 0.03616881, -0.0157778729, 0.0072753816, 0.0611979999, -0.0464350693, -0.0266204048, 0.0901299939, -0.025798481, 0.0418299846, 0.0141708599, 0.0391849317, -0.0045947805, -0.0126393931, -0.0227465108, -0.0017210662, -0.0166530367, -0.0151784923, -0.0336084962, 0.0309115406, 0.0233976748, 0.0204836782, -0.0512280315, 0.0115735708, -0.0649519116, 0.0350783281, 0.0179639924, 0.016903704, 0.0009080497, -0.032944981, -0.0129480781, 0.015058443, -0.0436990149, 0.0448476747, 0.049514275, 0.043169383, -0.0197078977, -0.041891966, 0.0061060064, -0.0356496759, 0.0126570193, -0.0076595121, -0.0144900065, -0.0482394621, -0.0452770479, -0.0315359421, 0.0026981691, -0.0260077585, 0.0450565778, -0.0521281138, -0.0113592036, -0.0245803688, -0.0256411266, 0.0297660325, -0.0062692212, -0.0099328775, -0.0472320095, 0.0512247011, 0.0204118453, -0.0074714068, 0.0375262499, 0.0215167142, 0.0355989486, 0.0684649125, 0.007448399, -0.0426061787, -0.0332288593, 0.0473181717, -0.0573427677, -0.0516026691, -0.071619451, 0.0189493187, 0.0784783661, 0.0293181669, 0.0125566246, -0.0349159017, -0.0086475201, 0.011111578, -0.0257001929, 0.0108964751, -0.0248105787, -0.0459694937, -0.0139906472, -0.0316346884, -0.0683678985, 0.0564050302, -0.0479893014, 0.0313792489, -0.0295182727, -0.049113337, 0.0014619797, 0.0188983772, -0.0374603644, 0.0278232843, 0.0376247391, -0.0389996879, -0.0117104836, -0.0496986695, 0.0036365613, -0.0147884265, 0.0358753018, -0.0436538123, -0.0163484868, -0.0089049982, -0.0188237876, -0.0253822189, -0.0500167012, -0.0368227921, -0.0178867336, 0.0229131989, -0.0402406603, -0.0109157292, -0.0083493805, 0.0086306129, -0.0033134217, -0.0284823943, 0.0027629437, -0.0363442153, 0.0043708063, 0.0144872973, -0.0980419293, -0.0291726273, -0.0073838271, 0.0390824117, 0.0793223232, 0.0420880765, 0.0257733203, -0.050226599, 0.0062331283, 0.0312487632, 0.0007126709, 0.0154655157, 0.0673793629, 0.0645085871, -0.007051365, -0.0438772254, 0.0476291403, 0.0300066043, -0.0017750644, 0.0401856564, -0.0347168744, 0.054876104, 0.008241808, 0.0183016025, -0.0157187656, -0.0232989248, 0.0250646546, 0.0686051697, 0.0070881215, 0.0213748701, -0.0141124995, -0.0480902456, 0.0058373515, 0.0603694096, -0.0459525362, -0.0277277846, -0.0433625802, -0.0334647931, 0.0043937573, 0.0221703816, -0.0002775088, 0.0238372423, 0.0089956354, 0.0235803071, -0.06357418, 0.0057479632, 0.007425935, -0.0604486577, 0.0572500639, 0.0917859375, -0.0282410327, -0.0206988566, 0.0346212685, 0.0255645979, 0.0037817501, -0.0052580857, 0.0304008182, -0.0276579186, 0.0095423758, 0.0189048667, -0.0048541226, 0.0488451496, -0.0217211209, -0.062584199, 0.0418123305, 0.0198212415, -0.000149835, 0.0106608169, -0.0448199101, -0.017762972, 0.0221894905, 0.0190640073, 0.0307470169, -0.0239004474, -0.0476833768, -0.0845146775, -0.0147048179, -0.0923978463, -0.0151554979, 0.0347855724, 0.0366094783, -0.052916687, -0.0086761741, 0.0271571446, -0.0530836582, 0.043675229, 0.0438223071, -0.01310552, -0.0150528178, 0.0010129601, -0.0286122244, -0.0106686512, -0.0162108652, 0.0568645932, -0.0489693955, 0.0485158339, -0.0289242472, 0.0117495907, -0.0338627212, 0.0102197174, -0.0584614798, -0.0953977481, 0.0190826356, 0.077320531, 0.0287050642, -0.0158241522, -0.0701230988, -0.047921963, 0.0559479445, 0.0940593928, -0.075576134, -0.0168864541, 0.0502247363, -0.0582580864, 0.0077762045, -0.022898145, -0.0336509161, -0.0209291317, 0.0911301002, 0.0184883308, -0.0346765928, -0.0079013212, 0.0285806991, 0.0319202468, 0.0078228535, -0.0200359337, 0.015392852, 0.0511965379, 0.0132771404, 0.0042141918, -0.026409192, 0.0343401618, 0.0232218262, 0.0520329326, -0.0338332653, 0.0262136329, -0.0046627345, 0.0038093422, 0.0320505463, 0.0235226136, 0.0235848557, 0.0194423515, 0.0914194211, 0.0444855988, 0.0017273552, -0.0436887331, 0.0453673229, -0.0205965005, 0.0242334735, -0.0219134279, 0.0298388619, -0.0200757701, -0.0021461118, -0.0321026742, 0.0510889143, -0.0711695999, -0.0486836918, -0.034716256, 0.0733842477, 0.0047209393, 0.0393111706, -0.0056505078, 0.0016751245, 0.0141878817, -0.0514266305, 0.0022843361, -0.0061879652, -0.0431888737, 0.018079998, -0.0050885584, -0.0015962588, -0.0183611643, 0.0514820181, 0.0107363472, 0.0476167202, 0.0172269382, -0.0277538467, -0.0277443007, 0.0101959864, 0.0704746544, 0.0325182751, -0.082527101, -0.0568700805, -0.0529034249, -0.0809543356, -0.0075978879, 0.0240782108, 0.0638466552, 0.0286833774, -0.0458517745, 0.0473693386, -0.0118502779, -0.001214777, -0.0462426133, 0.0155775007, 0.0175672527, -0.0456106514, -0.0104057835, -5.533316028e-33, 0.0249495693, -0.0003405141, -0.0671440735, -0.0442890301, 0.0116322441, -0.0258903243, -0.0352495983, 0.0218567606, -0.0035672914, -0.0553897507, -0.0000303578, -0.005669645, 0.0169800892, 0.0127388109, -0.0224511046, -0.0205037985, 0.0175950341, 0.0337216742, -0.0318396389, 0.0404941514, 0.0462759137, 0.0341386758, 0.0008390957, -0.0525264032, 0.0341693088, -0.0338818096, -0.0146381101, 0.0044696662, -0.0323443376, 0.0621551611, 0.0171194226, 0.007661344, -0.0061867405, -0.016662322, -0.0015172489, 0.0542351566, 0.0192361455, 0.0066639162, 0.0140680261, 0.0087803295, 0.0134787206, -0.0767190382, 0.0002194874, 0.0068204631, -0.0191331599, -0.0188650582, -0.0115883946, 0.0100437673, -0.0299634989, -0.012066626, -0.0200152881, -0.0041364548, -0.0371743627, -0.0074595306, -0.0069813584, -0.0055056522, -0.0141932303, -0.0201857779, 0.0071429051, 0.0427652784, 0.0079416428, 0.0063767387, 0.0529024452, 0.0306234062, -0.0583205819, 0.0305468533, -0.0228404924, -0.0635305792, 0.020482365, -0.0155357951, 0.0000574156, 0.0589546897, 0.0515282415, 0.0854632109, 0.0720757172, -0.0464439504, -0.0741080716, 0.0036872234, 0.0074273651, 0.0278656203, 0.038028542, 0.0182457156, -0.0290590525, 0.0123469355, -0.0211712345, -0.0169146154, -0.0091253761, -0.0663071275, 0.0268481988, -0.0179113019, -0.0487904102, -0.0364916995, 0.0141915614, -0.058725521, -0.0036729816, 0.0394917056, 0.0168705694, -0.0143587431, -0.0505110845, 0.040365018, 0.0163612347, -0.0343825854, 0.0371310711, -0.0618870147, 0.0372525081, -0.0157475881, 0.0424485579, 0.0031729147, -0.0637941211, -0.0142226331, -0.0174457841, 0.0308742933, -0.0019251246, 0.0111613134, 0.0052739284, -0.0144958086, 0.0133069931, -0.0104119126, 0.0061045988, -0.0093164127, -0.0348801799, 0.0419562533, -0.0205461103, 0.0392208844, 0.0186854675, 0.0279673692, -0.0039254092, 0.018148236, 0.0045893774, -0.0266068168, -0.0059961132, 0.0161681715, 2.874e-7, -0.0209545232, -0.0208481047, 0.0423695371, -0.0272865146, 0.0056466395, 0.0587951764, -0.0209599417, 0.1151165068, -0.0003444589, -0.0508065335, -0.0416171439, -0.0230766442, 0.0702236667, 0.0512980893, -0.0165706147, 0.054239668, -0.0786758289, -0.0017413431, -0.0572299138, 0.0066718473, 0.039304588, 0.0742088333, -0.026405884, -0.0215310827, 0.0006360617, -0.0460469276, -0.006892777, 0.0131658558, -0.0062569156, -0.0280291401, 0.0082696071, 0.0256065484, 0.0322247669, 0.0406212546, -0.0325119793, -0.0007087584, -0.0600290224, 0.0267071947, -0.0110283829, -0.0093661128, -0.0620296113, 0.0080687376, 0.0059219236, -0.0432957485, 0.0215092395, -0.0026493168, -0.0207814202, -0.0373196788, -0.0156141873, 0.0117148459, -0.0256508496, -0.0192443412, 0.0091224564, 0.0105735138, 0.0497253053, 0.042445641, 0.0190450866, -0.0550328977, -0.0310322046, 0.0163757913, -0.0103799067, -0.0385855362, -0.0108676134, 0.0731900558, 0.0482150018, -0.0278340429, 0.0416632295, 2.693166383e-34, 0.0040296097, -0.0548637025, 0.0220056474, 0.0454160571, 0.0278283898, 0.0215609577, 0.0185771808, 0.0414748825, 0.0083702048, 0.0145669654, -0.0200449917 ]
http://arxiv.org/abs/2406.18116v1
http://arxiv.org/pdf/2406.18116v1
BADGE: BADminton report Generation and Evaluation with LLM
2024-06-26T07:07:52
2024-06-26T07:07:52
Shang-Hsuan Chiang, Lin-Wei Chao, Kuang-Da Wang, Chih-Chuan Wang, Wen-Chih Peng
cs.CL
cs.CL, cs.AI, cs.HC
Badminton enjoys widespread popularity, and reports on matches generally include details such as player names, game scores, and ball types, providing audiences with a comprehensive view of the games. However, writing these reports can be a time-consuming task. This challenge led us to explore whether a Large Language Model (LLM) could automate the generation and evaluation of badminton reports. We introduce a novel framework named BADGE, designed for this purpose using LLM. Our method consists of two main phases: Report Generation and Report Evaluation. Initially, badminton-related data is processed by the LLM, which then generates a detailed report of the match. We tested different Input Data Types, In-Context Learning (ICL), and LLM, finding that GPT-4 performs best when using CSV data type and the Chain of Thought prompting. Following report generation, the LLM evaluates and scores the reports to assess their quality. Our comparisons between the scores evaluated by GPT-4 and human judges show a tendency to prefer GPT-4 generated reports. Since the application of LLM in badminton reporting remains largely unexplored, our research serves as a foundational step for future advancements in this area. Moreover, our method can be extended to other sports games, thereby enhancing sports promotion. For more details, please refer to https://github.com/AndyChiangSH/BADGE.
BADGE: BADminton report Generation and Evaluation with LLM: Badminton enjoys widespread popularity, and reports on matches generally include details such as player names, game scores, and ball types, providing audiences with a comprehensive view of the games. However, writing these reports can be a time-consuming task. This challenge led us to explore whether a Large Language Model (LLM) could automate the generation and evaluation of badminton reports. We introduce a novel framework named BADGE, designed for this purpose using LLM. Our method consists of two main phases: Report Generation and Report Evaluation. Initially, badminton-related data is processed by the LLM, which then generates a detailed report of the match. We tested different Input Data Types, In-Context Learning (ICL), and LLM, finding that GPT-4 performs best when using CSV data type and the Chain of Thought prompting. Following report generation, the LLM evaluates and scores the reports to assess their quality. Our comparisons between the scores evaluated by GPT-4 and human judges show a tendency to prefer GPT-4 generated reports. Since the application of LLM in badminton reporting remains largely unexplored, our research serves as a foundational step for future advancements in this area. Moreover, our method can be extended to other sports games, thereby enhancing sports promotion. For more details, please refer to https://github.com/AndyChiangSH/BADGE.
[ 0.0320475847, 0.0107657006, -0.0020404954, 0.065466173, -0.0138052506, 0.0073774322, -0.0466389842, 0.0276565049, -0.0012263061, -0.0371600688, -0.0129505349, -0.0380311161, -0.00087545, 0.0509290695, 0.0254035108, -0.0734892264, -0.0142227802, -0.0409196801, 0.0533650406, 0.0149339354, 0.0324325338, -0.0007921123, 0.0057924152, 0.0365805961, -0.017352581, -0.0171754304, -0.0576503053, -0.0005709204, -0.0053844485, -0.0738133639, -0.03812702, 0.0258024111, 0.0208065007, -0.0162096657, 0.0000017825, -0.0451840051, 0.0268458333, -0.037584126, 0.0183351859, 0.0327822864, 0.017105978, -0.0150668574, 0.0105390754, 0.0015725178, -0.0313797966, -0.0108535327, 0.0397675857, 0.0130846715, 0.089564316, 0.0696583837, -0.0107130902, -0.0345076099, 0.008478282, -0.0141517334, 0.0168099422, -0.0598715618, 0.0130610913, 0.0293830168, -0.0102650793, 0.0192553382, -0.0188059825, -0.0143630402, 0.0054116859, 0.0046834606, 0.004699056, 0.0278956406, -0.0369678661, 0.02143397, 0.0111069772, 0.0427445397, -0.0319225676, -0.0129202874, 0.0078536561, -0.0128370766, 0.0288680028, -0.0485618114, -0.0556385517, 0.0435305797, 0.0090444041, 0.0234302711, 0.0051063057, 0.007810785, 0.0030901053, 0.0052040527, -0.00251458, 0.013598131, -0.003462685, -0.0289337263, -0.0611007698, 0.0402970575, 0.0359871984, -0.1226657033, 0.0224605817, 0.0454924293, -0.0509754978, 0.0267753731, -0.0049079228, -0.028744163, 0.0057866122, -0.0475796163, 0.0564677939, 0.0204664096, 0.0277262758, 0.0163671058, -0.0187764242, 0.0105178691, 0.0340215862, 0.0059831003, -0.0462815501, 0.0092916116, -0.0512158759, -0.004624574, 0.0039237808, 0.0441223122, -0.0097868079, 0.0009064375, -0.0595693327, 0.0414761193, -0.0045783916, 0.0390369669, 0.0402382538, 0.0019137315, -0.0016431457, -0.0098465364, -0.0408697426, -0.0124586979, -0.0622410141, -0.0059769289, 0.0579181314, -0.0318318903, 0.0222895145, -0.0038456072, 0.0089492276, 0.0145308068, 0.0281698294, 0.0339026377, -0.0085384138, -0.0008687687, -0.0283409432, -0.0347824506, -0.0101867281, -0.0745042711, -0.0284642763, -0.1087467745, 0.0027859583, 0.0005032443, -0.0065784189, 0.0015017346, 0.023007663, 0.0417560674, -0.0029232677, 0.0840640441, -0.0098312087, 0.02457528, 0.0003702998, 0.006702031, 0.0194187295, 0.0385567769, 0.0309641883, 0.0440079011, 0.0082239043, 0.0266464557, 0.0566151962, 0.0390436463, -0.0180681273, -0.0179395508, -0.0480095968, 0.0220784899, -0.0378303342, 0.0527164228, -0.0146017969, 0.0509387739, -0.0015627482, 0.0501830615, 0.0826288313, 0.0562817752, 0.0300517119, 0.0162234344, 0.0728618801, -0.0144579066, 0.0136185931, 0.0049008685, -0.018554043, 0.0394469649, 0.0035209931, 0.0659361109, -0.0054636979, 0.0056301304, -0.0360022672, -0.0866020471, 0.0011525429, -0.0196861587, 0.0172319319, 0.0138236815, 0.0632174388, -0.0126572168, -0.0262722634, 0.0261302441, -0.0192955937, -0.0522502549, 0.0340941399, -0.0505756326, 0.0424727388, 0.0698161796, 0.0137897395, -0.0367758125, -0.000282682, -0.0117805135, 0.0041229944, 0.0347249582, 0.0330137536, -0.0283016358, -0.0095280856, -0.0359135158, -0.0009808285, 0.0091900667, 0.0337437093, 0.0101888543, -0.0222297907, 0.0201112945, 0.0115609253, 0.0040055378, -0.0819970891, 0.0414839461, 0.0360528901, -0.0673437938, 0.0177084766, -0.0168643314, 0.0023643896, 0.0509176664, 0.0080086775, 0.0191950258, 0.0472466946, 0.0058972295, -0.0072041834, 0.0257581938, 0.0018504992, -0.0362183116, 0.0368592106, -0.0470715687, -0.0066578812, 0.0142895542, -0.0412317552, -0.0366781838, -0.022872664, 0.0192956366, 0.0508791506, -0.0365187749, 0.0273650903, 0.0244115293, -0.051162634, 0.0020651433, 0.0678817332, 0.0330670215, -0.0017290682, 0.0040246435, 0.0162695423, 0.0105820578, -0.0669798404, -0.0240692794, 0.0559772179, -0.0147291003, 0.0184551571, 0.0438189246, 0.0504598469, -0.0771990716, 0.001597185, -0.1167905033, 0.0112486919, -0.0167237837, 0.0053911731, 0.0235885028, -0.0016788235, -0.0221663825, -0.0032361355, -0.0592240766, -0.0186169688, 0.0497803986, 0.0056786528, 0.0267459396, 0.0136796366, 0.0356035344, -0.0343314484, -0.0066092764, -0.0068378067, 0.0454222299, 0.0279920213, -0.0061922921, -0.1019657701, 0.0445256941, -0.0576827563, 0.0062008281, 0.0235396195, -0.0008826755, 0.0357237123, -0.0225783493, 0.0287231077, 0.0677774101, 0.0247761887, 0.0067298096, 0.0421341769, -0.0740706325, -0.0152506791, -0.0045710206, -0.0146654416, -0.0330465659, -0.0418997593, -0.0404593013, 0.0639514625, 0.0070482208, 0.0003549599, 0.0144953355, -0.0490114503, 0.0282200724, -0.0129324151, 0.0161181558, -0.0660513639, -0.0337573476, -0.0432097316, 0.0168606993, -0.0111594601, -0.0262380634, 0.0471114926, -0.0013073555, 0.0344583727, -0.0551635586, 0.0018275377, -0.0245010722, 0.1021847352, 0.0097466856, 0.0377510265, -0.0158292893, -0.0359283015, -0.0055557331, -0.0023705373, 0.0385086723, -0.0432326607, 0.0409958772, -0.0195151847, -0.0073758746, 0.0043251477, -0.0427834541, -0.0226429831, 0.0071257376, -0.0234731771, -0.0255620759, 0.0266077649, 0.0117487311, -0.0374290496, -0.0032733744, 0.0153703708, -0.0054718275, -0.0278878007, -0.0400326215, 0.0227669124, 0.0604314357, 0.0219489802, -0.0109067457, -0.0244215112, -0.0538633727, -0.001637943, 0.0520444587, 0.0518056341, 0.0261881426, 0.0258360822, 0.0005130876, 0.0162904877, -0.0393650234, -0.0370734818, 0.0556303859, 0.0672957376, -0.0737315565, 0.0008688028, 0.1401822567, 0.060092885, -0.0752620697, -0.001420392, 0.0107424762, -0.0324174054, -0.0007839212, 0.027355263, -0.0697727129, 0.0003590798, 0.0193999391, 0.0021447877, -0.0254271794, -0.0240140427, -0.0284756068, -0.1265297234, 0.0201496929, 0.0113377413, -0.0183370672, -0.0816176161, 0.0028251954, 0.0040684217, -0.0180407707, -0.0281501561, 0.0328717977, -0.0113946302, 0.0396900661, 0.0113488613, -0.0477682948, -0.0429097041, 0.0265155509, -0.0920128599, 0.0577975959, 0.064962104, 0.0492867008, -0.0221729539, -0.0288518015, -0.012862674, 0.0141509064, -0.009053492, -0.0255303755, 0.050178811, -0.0226453729, -0.0080345552, -0.0446796566, 0.022409169, 0.0287364386, -0.0108955065, 0.0694304109, 0.0705681369, 0.0357195362, -0.0000463261, 0.0142860711, -0.0463926941, 0.0205398723, 0.0538816564, 0.0208401885, -0.0480748489, -0.0079034027, -0.0595581941, 0.0267209746, -0.0466576368, -0.0624156147, -0.0481297486, 0.0198446643, -0.0217066668, -0.0542997532, -0.0645451397, 0.0559596308, 0.0300921649, 0.0093282862, 0.0571386032, -0.0429876037, -0.0147683816, -0.059802305, -0.001492881, -0.0089198695, 0.0022489484, -0.0520632379, -0.0112800067, 0.0445836112, -0.023026621, -0.009430211, 0.0412247963, -0.0136932814, -0.0226570312, -0.0165642779, 0.0564467944, 0.0264643282, 0.0108330771, 0.0155914873, -0.0073121311, -0.0289886575, 0.0201467238, 0.002576153, -0.0113461427, -0.0263869595, 0.013019097, 0.0144486548, 0.0034432074, -0.0102858944, 0.0193827096, -0.0036042356, 0.0598941743, -0.0566798225, 0.0033162662, 0.053723406, 0.0273283366, 0.0563480034, -0.0275849011, -0.0695843771, -0.0044453484, -0.0335351117, -0.0657348931, -0.0224307012, -0.0346332788, -0.0122035788, 0.0027594166, 0.0242383052, -0.0291719195, 0.0079617389, -0.0443342812, 0.0152973281, -0.0585292391, 0.0354721472, 0.0309620313, 0.0072146794, 0.0039435644, -0.0009988174, -0.0128043685, -0.0116331726, 0.0110495929, 0.0054562534, -0.0106211817, 0.01023807, 0.0273568314, -0.0084350742, 0.0351360105, 0.0156315528, -0.0056047388, 0.0587407611, -0.0387804918, 0.0430077501, 0.0995910615, -0.0059940447, -0.0080226045, -0.0293611903, -0.0086532487, 0.001619602, 0.0254054368, -0.006174447, -0.0579059273, 0.0118757291, 0.0181193091, -0.0021231016, -0.0390904993, 0.0453158543, 0.0180781577, -0.0081750751, 0.0101523194, -0.0721260458, -0.0491977669, 0.025093412, 0.0240026545, 0.0225937888, -0.0425535478, -0.0243569743, -0.0730564967, -0.0309187323, -0.0003478161, 0.0395873003, 0.0294046234, -0.0003517928, 0.0223888531, 0.0642774478, -0.0047760033, -0.0028274062, -0.0377717428, 0.0179769713, -0.0222376008, -0.0482708625, -0.0335184075, -5.555675767e-33, -0.0381396301, -0.072247766, 0.0345098861, 0.0467747413, -0.0496342778, -0.001127601, -0.0527932905, -0.0001345383, 0.014854393, 0.0029758252, -0.0215508249, 0.0039025589, 0.011612419, -0.0120712714, 0.0788610876, -0.0528988913, 0.0569659248, -0.0269619599, -0.0290130265, -0.0112311766, -0.0129715428, 0.0362339467, 0.0317151323, -0.0673136264, 0.0505975261, -0.0304099023, -0.0539530702, -0.0094555998, 0.0206161831, 0.0255590547, 0.0042886208, 0.0059585581, 0.0346814953, -0.0167177096, -0.0141493687, 0.0426953621, -0.0314593352, -0.0324528329, -0.0099013969, 0.0087681375, 0.0016470795, 0.0115391901, 0.0569932163, -0.0319999047, -0.0251144152, -0.0189043563, 0.044806812, -0.0492024831, 0.0160079915, -0.0713819265, -0.0570915043, 0.0245146398, 0.0265760813, -0.0016408238, 0.0370069146, 0.001178203, -0.014068312, -0.0052188714, -0.049906224, 0.0579257123, 0.0164854061, 0.0507988557, 0.0362387821, 0.0400622524, 0.0006467405, 0.0088850716, 0.0334673561, -0.0148906009, -0.0491376296, -0.0062072226, 0.0173233412, 0.0297526494, 0.0677384883, 0.0008260196, 0.0349723548, 0.0011060607, -0.0530748591, 0.0312313456, 0.0556139313, 0.0325122699, -0.0409982614, 0.0103248162, -0.0242387652, 0.0096312575, -0.0218139868, 0.0002442191, -0.0285696853, -0.0685239881, 0.0378464162, -0.0149061847, 0.0127238911, 0.0204392411, -0.0015533289, 0.0305389613, -0.057037238, 0.0210096575, -0.0031385552, 0.0866575539, -0.0315625034, -0.0129122613, -0.0673552603, 0.0148661872, -0.0291944332, 0.0406241976, -0.0129929176, 0.0045894454, 0.0100403298, 0.0606395975, -0.0565657988, -0.0002345616, 0.0011463199, -0.0066475654, 0.0080493828, -0.0112917237, 0.0049288706, -0.0068894844, -0.0092852488, 0.0174947698, 0.0179532766, -0.0653284788, -0.0391349234, -0.0050453749, 0.012834915, 0.0182475504, -0.0000578127, -0.0061458391, -0.0380676948, 0.0511689372, 0.0392089002, 0.0025061783, -0.0121431006, 0.0088105993, 2.599e-7, 0.0226467475, 0.0806647092, -0.0529656224, 0.0212521572, -0.0294526145, 0.0259214342, -0.0535425022, 0.031075038, 0.063262552, 0.0450681001, 0.0448268168, 0.0240819026, 0.0243542176, 0.0513830185, -0.1250981539, -0.0222909302, -0.0531049483, -0.0318348669, -0.0582620203, 0.0371239893, 0.0459789634, 0.0294003505, 0.0442433916, 0.0224469267, 0.0270737056, -0.0446717814, 0.065614678, -0.001240171, -0.0202850234, 0.0035571901, 0.0202307981, 0.0045287441, -0.0526576079, 0.0054352283, -0.0280015822, -0.0174607225, 0.0350246467, 0.0288792085, -0.0154441791, 0.0621286221, -0.0052742823, 0.012375026, -0.0283192415, -0.0434546359, 0.0405164585, -0.026799025, -0.0104338098, 0.005506631, -0.0438409857, -0.0203280374, -0.0124557111, -0.0448836274, -0.048523277, 0.0195686799, -0.018585952, 0.0183136966, 0.0141421305, -0.0314878486, -0.0257648677, -0.0183394179, -0.0198736135, -0.0130708553, -0.0305983573, 0.0339985564, -0.0377175957, -0.0795121044, 0.0077252164, 2.244426236e-34, 0.0182666425, -0.0031743152, -0.017087739, -0.0262841973, -0.0053721061, 0.0283770226, -0.012624247, 0.0382108241, 0.037596494, -0.0362787135, -0.0227565244 ]
http://arxiv.org/abs/2406.18108v1
http://arxiv.org/pdf/2406.18108v1
Token-Weighted RNN-T for Learning from Flawed Data
2024-06-26T06:48:11
2024-06-26T06:48:11
Gil Keren, Wei Zhou, Ozlem Kalinli
cs.CL
cs.CL, cs.LG, cs.SD, eess.AS
ASR models are commonly trained with the cross-entropy criterion to increase the probability of a target token sequence. While optimizing the probability of all tokens in the target sequence is sensible, one may want to de-emphasize tokens that reflect transcription errors. In this work, we propose a novel token-weighted RNN-T criterion that augments the RNN-T objective with token-specific weights. The new objective is used for mitigating accuracy loss from transcriptions errors in the training data, which naturally appear in two settings: pseudo-labeling and human annotation errors. Experiments results show that using our method for semi-supervised learning with pseudo-labels leads to a consistent accuracy improvement, up to 38% relative. We also analyze the accuracy degradation resulting from different levels of WER in the reference transcription, and show that token-weighted RNN-T is suitable for overcoming this degradation, recovering 64%-99% of the accuracy loss.
Token-Weighted RNN-T for Learning from Flawed Data: ASR models are commonly trained with the cross-entropy criterion to increase the probability of a target token sequence. While optimizing the probability of all tokens in the target sequence is sensible, one may want to de-emphasize tokens that reflect transcription errors. In this work, we propose a novel token-weighted RNN-T criterion that augments the RNN-T objective with token-specific weights. The new objective is used for mitigating accuracy loss from transcriptions errors in the training data, which naturally appear in two settings: pseudo-labeling and human annotation errors. Experiments results show that using our method for semi-supervised learning with pseudo-labels leads to a consistent accuracy improvement, up to 38% relative. We also analyze the accuracy degradation resulting from different levels of WER in the reference transcription, and show that token-weighted RNN-T is suitable for overcoming this degradation, recovering 64%-99% of the accuracy loss.
[ 0.0362308882, 0.0487010442, 0.0227384418, -0.0304446872, 0.0079037594, 0.0466002263, 0.0093600415, 0.012119432, -0.0194059946, -0.005943906, -0.022276653, -0.0376982912, 0.0363572724, -0.0165620297, 0.0468266308, -0.0176887605, 0.0382353961, 0.0113886753, 0.0390801802, 0.0051302579, 0.0122052282, -0.0090806549, -0.0320183188, 0.0432385989, -0.0901302472, 0.0208073743, 0.0363770053, 0.026948981, 0.0252234191, -0.0180552676, 0.0173941832, 0.0339563191, 0.0084508592, 0.0688890517, 0.0000017785, 0.0017085532, -0.0244845748, 0.003695173, -0.0110483794, -0.0219386034, 0.0271447506, 0.0127182966, -0.0268389434, 0.0350906216, 0.0141715119, -0.0744743571, 0.0802919939, 0.1117039621, 0.0235237908, 0.0686146095, 0.0092601888, 0.0040351725, 0.0009796341, -0.03870482, 0.0352912433, -0.0454126149, 0.0335715227, -0.0261612386, -0.0139082698, -0.0490079485, 0.0004620003, 0.002729615, 0.0470428392, -0.0021898276, 0.0018899768, 0.0258844383, -0.0541889966, -0.0418298244, 0.0006869971, 0.0338506252, -0.015589173, -0.0404396281, 0.0219007023, -0.0145355202, -0.0030746658, -0.0517232791, 0.0084656719, 0.0084745362, 0.0135700488, 0.028091969, -0.0129797729, -0.0044213925, 0.0675635487, -0.0073762191, -0.0263772998, 0.0358191729, -0.0080457442, -0.0520466752, -0.011933051, 0.0024607705, -0.0113373855, -0.0307824928, 0.0282122288, -0.0010359896, -0.0164202359, -0.0025208048, -0.0046005622, -0.0393056162, 0.0379692167, -0.029073745, -0.0042279619, 0.0025656493, -0.0626037046, -0.0008680677, -0.0208193883, 0.0503262207, 0.0359440893, 0.0630021691, -0.0113629485, -0.0143812606, -0.0420595258, 0.0051205773, -0.0256600138, 0.072530888, 0.0125672454, 0.0086640827, -0.087894842, -0.0134560261, -0.0237930324, 0.0084378673, 0.0564702228, -0.0038991857, -0.0734872296, 0.0206561219, 0.0177428592, 0.0022800458, 0.0381684788, 0.0131122489, 0.0142716197, -0.0570505075, -0.0225183088, 0.0160995759, -0.0043766219, -0.0286963843, -0.0052891984, 0.0753670633, 0.042891793, -0.0411908738, -0.0284955092, 0.0095363334, -0.0560180843, -0.0332777537, -0.0594500229, -0.0294388942, -0.0227035005, -0.0284644831, -0.0076220394, -0.0161617063, 0.007170476, -0.0163124762, 0.0154718757, 0.0457533225, -0.0343087167, -0.042002175, 0.0089770369, -0.0045033311, -0.0197673347, 0.0295774527, 0.0351919271, 0.0242304932, -0.0206874162, -0.006992579, -0.0176490843, 0.0072270948, 0.0444102995, -0.0445252955, -0.0045223455, -0.0540647432, 0.0265191123, 0.056991104, 0.0532238409, 0.072292015, -0.0458437987, -0.0006153435, 0.018199319, 0.008475096, 0.014380971, 0.0377728939, 0.0315380804, -0.0150901135, -0.0415789932, 0.026438361, -0.037518464, 0.0157256778, 0.0028653811, 0.020821197, 0.0719321519, -0.0189103857, -0.0255138911, -0.0572325885, -0.0272257496, -0.0019468084, -0.0325968638, 0.011114466, 0.0542924106, 0.0067335311, -0.0519713387, 0.0287050344, 0.0058431919, -0.0409350432, 0.0283061303, -0.0189340487, -0.0150950383, 0.0376313999, 0.0235696826, -0.0755016655, -0.023562137, -0.014569141, -0.0296662413, 0.0343349315, -0.0013206375, 0.0073932428, 0.0253518652, -0.0292950906, -0.009427486, 0.0766085237, 0.0045290985, 0.0564702973, -0.0572945215, -0.0022229685, 0.0270547029, -0.0712190121, -0.0423527211, 0.0442770794, 0.0091609387, -0.0200606231, 0.0372978263, 0.0203888938, 0.0666100755, 0.0412718207, -0.0564293005, 0.0135679897, 0.0391601324, -0.0222002715, -0.042669788, -0.0157225151, -0.0826545805, -0.0455182232, 0.0479548164, -0.0327107534, -0.0367194191, 0.0224426612, -0.030103853, -0.0659774318, 0.036029879, -0.0379868224, 0.0628546849, -0.0711798221, 0.0044101197, 0.0417387448, -0.0284482036, -0.0396255255, 0.0198155195, 0.0069482685, 0.0836338103, -0.0318576656, -0.0744988471, -0.0133144045, -0.08620435, 0.005352289, -0.0052320082, -0.046664577, -0.0047522918, 0.0258369818, -0.0089624757, 0.0080625182, -0.0044455985, -0.067748569, -0.0432228632, -0.0369501114, 0.0221357439, -0.037210051, 0.0307921674, 0.006794468, -0.0214061532, -0.0297612399, -0.0535790883, -0.0202715397, -0.014730148, 0.0180259589, 0.0141964201, -0.0082704769, -0.0167294722, -0.0045612506, -0.0321972892, 0.0028096244, 0.0012049172, -0.0248866435, -0.05833349, -0.0056071798, 0.0063614119, 0.0083617242, 0.0232353378, -0.0326543264, 0.0164046623, 0.0192560442, 0.0158769321, -0.0188683569, -0.0145620741, -0.0077015678, 0.0019603597, 0.0088659218, -0.0488240235, -0.0047855526, -0.0088742916, 0.0369357765, 0.0258997418, -0.0306931231, -0.0007718701, 0.0623232946, -0.003854163, -0.0038576976, -0.0571618825, -0.0252795983, -0.0058594267, -0.0428844988, -0.0012718093, -0.0309123788, -0.0186899323, -0.0067598713, -0.0023340122, 0.0091001531, 0.0760672763, -0.0356137045, -0.0079312446, -0.0358051211, -0.0226197802, -0.0168236308, -0.0388640426, 0.0242790077, 0.0273596309, 0.0588465482, -0.0685063675, 0.0044401241, 0.0031268473, -0.0169303361, -0.0020923261, 0.0133275855, -0.0052646515, 0.0490396433, -0.0103728296, -0.0297408421, -0.0128818899, 0.0038825118, 0.002471234, -0.0036854963, -0.001481354, -0.0333064198, -0.0149083454, -0.0301250871, -0.0096423589, 0.0949647427, 0.0012072303, 0.0150696952, -0.0231167506, 0.0118115572, 0.0187117159, -0.0285351705, 0.0383215696, -0.0142598357, -0.0026544728, 0.030094482, 0.055660814, 0.0051047211, -0.0170964357, -0.0082680266, 0.0179309007, -0.0401739776, 0.0350320637, 0.0368778445, 0.0197800305, -0.0015130051, -0.0307638273, 0.1019982249, 0.0225182064, -0.0374231525, 0.0012666599, 0.000648567, 0.0398019031, 0.0017739314, -0.0075388374, -0.0607115924, -0.0187522545, 0.0442510359, 0.0371737294, -0.0289252531, -0.033267729, 0.0634111464, -0.0552387126, 0.0298121218, 0.0339604802, -0.0971615911, -0.0309606437, -0.0708318129, -0.0315377116, -0.0803041905, 0.0747318417, 0.0291037504, 0.0038141569, -0.0083970986, 0.0960505009, -0.0411277302, -0.0772977546, -0.0116888396, -0.0438100062, 0.0703475624, 0.0870266706, 0.004347235, -0.0091316309, -0.0601003021, 0.0412305035, 0.0023309498, -0.0159535967, -0.0187182259, 0.0187538881, 0.0215722099, -0.0226870235, -0.0144733749, 0.0125609385, -0.0471424162, -0.0078746369, -0.0696426779, 0.071006462, -0.013643737, 0.0059735179, 0.0286652315, -0.0260611605, 0.027947057, 0.0468853749, 0.0467595123, -0.0138460789, -0.0262493156, -0.0248595271, 0.0865135863, -0.0724779889, 0.0140643343, -0.013725847, -0.0340680294, -0.098232843, -0.0916762874, -0.0183655489, -0.0051503526, 0.0457087234, -0.0017541171, 0.0015851553, 0.016061252, -0.0151957478, -0.0208664406, 0.00679194, -0.0187251028, 0.023562951, -0.0686286986, 0.0052634901, 0.0188271515, -0.0104864901, -0.0022438918, 0.0159518309, -0.0383856185, -0.0706171915, 0.0390326865, 0.0190936942, 0.0131953619, 0.0212617833, -0.01383969, -0.0371947922, 0.0660329908, 0.0711485744, -0.0085342871, -0.0313517116, 0.0402490571, -0.0529502109, -0.0215781033, 0.0192843359, -0.0553857498, 0.0187902004, 0.0269539952, 0.0620645881, -0.0314368531, 0.0019541527, 0.0465668663, 0.022139186, 0.0910134539, -0.0117589422, -0.0216983464, -0.0190492254, 0.028472472, 0.0021059918, 0.021856375, -0.0095072165, 0.0222725179, 0.0184299983, 0.0248415396, -0.0364891812, 0.018879462, -0.0583447218, 0.0028088004, 0.0122044627, 0.0227807909, -0.0167831425, 0.046448864, -0.0188449696, -0.0166628323, -0.012060415, 0.0060231164, 0.0342600383, 0.0329050533, 0.0122646438, -0.0405019633, 0.0129418857, 0.0562139899, 0.0191483889, -0.0067919898, -0.0501486994, 0.0443893671, -0.0596479103, 0.0112968422, 0.0521237738, -0.0005593934, 0.0061097564, 0.0066319034, 0.0139836054, -0.0401706807, 0.0046502254, -0.0248771664, -0.0118340952, 0.0227796156, 0.0187959503, 0.0045995116, 0.0126507469, 0.0628253669, 0.0498912074, 0.0817392394, -0.049407877, -0.0149132265, -0.0139671071, -0.0163724516, 0.0037643053, -0.0122319534, -0.0504751839, -0.0364544429, -0.0162214357, -0.1031108052, 0.0546651557, 0.0391484164, 0.0459197797, 0.0133761484, 0.0129572144, -0.0232098792, -0.0011755766, 0.0097357072, -0.0181146562, -0.0218441822, 0.064912729, -0.0556831062, 0.0153190931, -6.61918371e-33, 0.0079588043, -0.0130540319, -0.0135596227, -0.0031363827, -0.0374880619, 0.0135619985, -0.0139698582, 0.0088162031, 0.0257049259, 0.0052051046, -0.0031664835, 0.0014329706, 0.0271674395, 0.0236492362, 0.0080204932, -0.0069888677, 0.0552071109, 0.0076446747, 0.009342582, 0.0018121004, 0.0591177903, 0.0403852127, 0.0551277697, -0.0486352667, 0.0171916503, 0.0479550585, -0.0046707485, -0.0003586568, -0.0041957255, 0.0081595946, 0.0166452546, 0.0262986869, -0.0087789614, -0.0353075638, -0.0222405791, -0.022797456, -0.0738113075, -0.0260084663, -0.0112625184, 0.0476187915, 0.0147311231, -0.054634396, 0.0761420801, -0.0179808177, -0.0712587088, 0.0176430028, 0.0284436382, -0.0066831736, -0.0491484068, -0.0653785244, -0.0241424907, 0.0421834812, -0.0192329604, 0.0006703281, 0.0309505425, -0.0006304674, -0.0102905808, 0.0101247635, -0.101074785, -0.0069315834, -0.0300559904, 0.0620152019, 0.0134684537, 0.0171457306, -0.0254558828, 0.0295965765, 0.0562878698, -0.0249878615, -0.0036761558, -0.0031756978, -0.002647355, -0.029635191, 0.0608660281, 0.0199428089, 0.0823135525, -0.0433922783, -0.0676110983, 0.0326690376, -0.0402201228, 0.0611801483, 0.0027172682, 0.0603875555, -0.0215884149, 0.0257039908, -0.0163104292, -0.0329527408, -0.0357257426, 0.0077956985, -0.0094998851, 0.0288847964, 0.0121121695, 0.0541284829, -0.0071960944, 0.0086754262, -0.0031794945, -0.0223271269, 0.0826940313, -0.008581222, -0.0149996858, 0.0200404041, -0.0635447204, -0.0245858971, 0.0796148628, 0.0293526705, 0.0338009037, -0.0484127775, 0.0799390823, 0.0252367016, -0.0471892767, -0.013255598, 0.017519692, 0.0519488044, 0.0027142973, -0.0274771024, 0.0280438215, -0.0338913202, 0.0081875613, 0.002066883, -0.0097977426, 0.0088227475, -0.010638494, 0.0654050112, -0.0068601714, 0.0136634884, -0.0358153023, -0.0369132496, 0.0421637818, -0.005892213, -0.0055760094, -0.0461649522, -0.0304208696, 0.0302761588, 2.721e-7, -0.0176846422, 0.0238749441, 0.0738258064, -0.064181909, 0.015021408, 0.0306366384, -0.0407317169, 0.0221128911, 0.0115632769, 0.0063706078, 0.043983452, -0.0386808068, 0.0530354977, 0.0109395832, -0.1001160517, -0.0238707047, -0.0667506754, 0.041601073, 0.02546525, 0.0138667757, 0.0078147016, 0.1191440672, -0.0044107656, -0.0530258231, 0.0380376391, 0.0003485159, 0.0337431729, 0.0160816833, 0.0092948386, -0.0153643694, 0.0223176349, -0.0148516977, 0.0248819049, 0.0270376522, -0.0340344198, -0.0303787589, 0.0460862443, -0.0190221425, -0.0393383615, -0.0523584634, 0.0129055614, 0.0645990074, 0.0486536473, 0.0013537869, 0.0105021307, -0.0069654034, -0.014913247, -0.0195273701, -0.0366306044, 0.0254740883, 0.0246583316, -0.0160406809, -0.0201363768, 0.0387352444, 0.0330406651, 0.0154455518, -0.0110861911, -0.0350678042, 0.0143179568, 0.0080329431, -0.0412808172, 0.0092325546, -0.0401182175, 0.0456983857, 0.0052300347, -0.0527311154, -0.0144658731, 1.964326365e-34, 0.0334860869, -0.056579601, -0.0105541628, -0.0007173607, 0.0223046467, 0.0191776268, -0.0167505555, 0.0278294887, -0.0217719283, -0.0299798343, -0.0354099944 ]
http://arxiv.org/abs/2406.18094v1
http://arxiv.org/pdf/2406.18094v1
Shimo Lab at "Discharge Me!": Discharge Summarization by Prompt-Driven Concatenation of Electronic Health Record Sections
2024-06-26T06:10:20
2024-06-26T06:10:20
Yunzhen He, Hiroaki Yamagiwa, Hidetoshi Shimodaira
cs.CL
cs.CL
In this paper, we present our approach to the shared task "Discharge Me!" at the BioNLP Workshop 2024. The primary goal of this task is to reduce the time and effort clinicians spend on writing detailed notes in the electronic health record (EHR). Participants develop a pipeline to generate the "Brief Hospital Course" and "Discharge Instructions" sections from the EHR. Our approach involves a first step of extracting the relevant sections from the EHR. We then add explanatory prompts to these sections and concatenate them with separate tokens to create the input text. To train a text generation model, we perform LoRA fine-tuning on the ClinicalT5-large model. On the final test data, our approach achieved a ROUGE-1 score of $0.394$, which is comparable to the top solutions.
Shimo Lab at "Discharge Me!": Discharge Summarization by Prompt-Driven Concatenation of Electronic Health Record Sections: In this paper, we present our approach to the shared task "Discharge Me!" at the BioNLP Workshop 2024. The primary goal of this task is to reduce the time and effort clinicians spend on writing detailed notes in the electronic health record (EHR). Participants develop a pipeline to generate the "Brief Hospital Course" and "Discharge Instructions" sections from the EHR. Our approach involves a first step of extracting the relevant sections from the EHR. We then add explanatory prompts to these sections and concatenate them with separate tokens to create the input text. To train a text generation model, we perform LoRA fine-tuning on the ClinicalT5-large model. On the final test data, our approach achieved a ROUGE-1 score of $0.394$, which is comparable to the top solutions.
[ 0.0526900701, 0.0091706309, 0.0040654461, -0.0788339972, -0.0088894488, 0.0238967966, 0.0178263281, 0.0143379914, 0.0208864622, -0.0010184563, 0.0253491066, -0.0280800853, 0.0347143188, 0.0916703269, 0.0308186188, -0.0687616244, 0.0215592198, 0.0221156627, 0.019597413, 0.0142513709, 0.0008011576, 0.0003485927, -0.0324145667, 0.0454928987, 0.0357282721, -0.010457512, -0.0162106622, 0.0361627713, 0.0431146733, -0.0662010983, 0.007109344, 0.013907304, -0.0215670113, -0.0371257104, 0.000002051, -0.0500863791, -0.0205292515, 0.0317828842, -0.0765237361, -0.0309715718, 0.0111156385, -0.0432166494, 0.0021442368, 0.0265400726, 0.0139664477, -0.030836815, 0.0188405439, 0.0485063493, 0.0699084178, 0.0686235428, -0.0203492921, 0.0328480601, -0.0185815096, -0.0440675355, 0.0412376486, 0.0210196767, -0.0071970061, -0.0267977826, 0.0481531732, -0.0627637729, 0.0022305057, 0.0216512922, 0.0193554405, 0.017376259, 0.0370422676, 0.0741263479, 0.0126859266, 0.0101659847, -0.0174546484, 0.047538843, 0.036052756, 0.0204478726, 0.0706761479, -0.0155828418, 0.0033538661, -0.032323949, -0.0567960516, 0.0188109241, -0.0391932316, -0.0100877527, 0.0055577182, 0.0203836989, -0.0042668097, -0.0267347358, -0.0097598629, -0.0236285739, 0.0221472215, -0.0426869243, -0.0043526636, -0.0049557872, 0.0252385363, -0.0687348992, 0.0554096587, 0.0598707385, -0.0534607507, -0.0195295922, -0.0320246667, -0.0309699569, 0.0893804133, 0.0207013637, 0.0272793639, 0.0010100156, 0.0455362126, 0.0101166032, -0.0136372326, 0.0249028411, 0.0009832272, -0.006933731, -0.0380297638, 0.0274065714, -0.0291278716, -0.0223366208, -0.0134431422, 0.0471406654, -0.0232536085, -0.0035417839, -0.0958931521, -0.0234276094, 0.0272497889, 0.0362249054, 0.0633023605, 0.022719834, -0.0367258899, 0.0215800349, -0.003000201, 0.0039587417, -0.0159163885, -0.0238269158, 0.0208486207, -0.0247011427, -0.0034362962, 0.0132860178, 0.010660707, 0.0039470959, 0.0109191053, 0.0659028217, -0.0234725215, 0.0001649133, -0.0092913248, -0.0743093416, -0.0316682123, -0.0335638821, 0.0103551205, -0.0479487814, 0.03660018, -0.0082618417, -0.0135079138, -0.0243723933, 0.0205107313, 0.0366950557, -0.0161849987, 0.0097413519, -0.0813126639, -0.0043628281, 0.0476348735, 0.0053829029, 0.0066720708, -0.0184281636, 0.03476597, 0.033698611, -0.0206445567, -0.0039027734, 0.0277976077, 0.0164115857, 0.0188949946, -0.047194086, 0.0492929518, -0.0068987012, 0.0020641517, 0.041890718, 0.0073208455, 0.0806138813, -0.0093301125, 0.056733802, 0.0281715952, 0.0585553944, 0.0310546942, -0.04644344, 0.049077712, -0.0194376651, -0.0215004031, -0.0065586311, -0.0222889632, 0.0175832342, 0.0414033122, 0.0442897826, 0.0554969758, 0.0199907571, -0.0406498723, -0.058399681, -0.0108566554, -0.0165639967, -0.0260402206, -0.0062432042, 0.0210353024, -0.0020844564, -0.054848101, 0.0536584146, 0.0474949032, -0.0620487891, 0.0137288822, -0.0624069646, 0.0049833115, 0.1050842479, 0.0519478805, -0.0073301541, -0.0511750057, -0.0532207415, 0.0024117809, 0.0355822705, 0.0321267769, -0.0114914859, 0.0037502479, -0.0649753809, -0.0255633574, 0.0220513642, -0.0185713843, -0.0051851557, -0.0231904853, -0.0048866863, 0.0250100419, -0.0083989901, -0.0724226832, 0.0505623743, -0.0707886219, -0.0267195478, 0.020575745, 0.0054862462, 0.0363248587, 0.0432441756, 0.0058219936, 0.047599256, 0.0433350988, -0.030364776, 0.0065152864, 0.0204812791, -0.0377328098, -0.0325538442, 0.0075065512, -0.0288937185, 0.0148385847, -0.0321586616, -0.0125959832, -0.0360847563, -0.0440476872, -0.0553393029, -0.0064890203, -0.0241921302, 0.0399600416, 0.088540189, -0.0330794938, -0.0392199308, 0.0401280075, -0.0107752457, 0.0245842244, 0.0421747491, -0.0069693066, 0.0116901835, -0.0509630069, 0.0092618549, 0.0045601898, -0.0220662057, -0.0127118668, 0.0280879233, 0.0308278278, -0.0478241481, -0.0243210923, -0.1214301437, -0.0019833057, 0.0059969267, 0.0183707662, 0.0340129621, 0.0179956127, -0.0170496237, -0.0180425458, -0.047608234, -0.0337396227, 0.0383244529, 0.0109057892, 0.0339390039, 0.0363003239, 0.0042694132, -0.0082468009, -0.0192724466, -0.0276938267, 0.0247220416, 0.0338374227, -0.0156110311, -0.031048879, 0.0014540224, 0.0046101329, -0.0350593068, 0.0134587362, -0.0124019887, 0.0256175846, -0.0030853201, 0.103658773, 0.0227864385, 0.0183621775, 0.0152157331, 0.0242102351, -0.0148660038, -0.0268566478, 0.0114663858, -0.0234285612, 0.018397795, 0.0330442451, -0.0895222276, 0.0480909497, 0.038069915, 0.0443445556, 0.0360996574, -0.0442432202, -0.0336199775, -0.0171537157, 0.0011136367, -0.0195770431, 0.0102347219, -0.0511793457, 0.0208214, -0.0134924063, -0.0167201161, 0.0601608679, -0.009040649, 0.0261368547, -0.0531200618, -0.0145147648, -0.0465098508, 0.0282021314, -0.0158512574, 0.0145580396, 0.0465187952, -0.0506377779, 0.0132411709, 0.0062118205, -0.028492175, -0.0360284634, 0.0345715284, -0.0254469644, -0.001416527, -0.0099649355, -0.0493812189, -0.0213390924, -0.0238307379, 0.0024136298, 0.0013209647, -0.007010079, 0.0150249172, -0.0637357384, 0.0228670873, 0.0085343737, 0.0637105703, -0.0121967196, -0.0139738536, 0.0015892051, 0.0483440682, 0.0559330173, -0.0231720693, -0.0090779243, -0.087413989, 0.0099800564, 0.0586033762, 0.0410543866, 0.0485689044, 0.0394446515, -0.0063852188, -0.016543109, -0.0891852677, -0.0279623102, 0.0457064137, 0.0607407354, -0.0406982116, -0.0104028517, 0.1153815836, 0.0272074342, -0.0402863845, 0.0196831692, 0.0092933057, 0.0110617178, 0.0069099013, 0.0220881216, -0.0204110593, -0.0074854312, 0.0144107211, 0.0807753727, -0.0260564163, -0.0295544062, -0.0064700651, -0.0607184395, 0.0322703086, 0.0008790356, -0.0409258679, -0.0565805845, -0.0288599022, -0.0054776534, 0.0068384381, -0.0266644284, 0.0163729619, -0.0389782488, 0.0067215073, 0.0398140177, -0.0528490283, -0.0371150933, -0.0140912151, -0.019737782, 0.0537859835, 0.0777901709, 0.0214632303, -0.0043603005, -0.0090101175, 0.0174633935, -0.0390649997, 0.0142891193, -0.0071667871, 0.0895120502, 0.008478079, 0.0051426119, -0.0653801486, 0.0546246916, 0.0012179802, -0.0430984385, 0.0330033861, 0.0265908167, -0.0156545546, -0.0356810465, -0.0425574668, 0.0497300848, 0.0147004845, 0.0460061915, 0.0350789465, -0.0239881482, -0.0424940996, 0.0028509197, 0.0435001813, -0.0323235132, 0.0273884051, -0.021489922, -0.054751832, -0.0300197583, -0.0699997172, -0.0603102259, 0.0904851332, 0.0241327938, 0.0061120563, -0.0077181961, -0.014363002, 0.0031310138, -0.0517140515, 0.0116812708, 0.0006566132, 0.000276422, -0.0930821076, -0.0594230071, 0.0176924691, 0.0081133414, -0.0184140466, 0.0486291684, -0.0322133638, -0.0195351727, 0.0115953889, 0.0451603867, 0.0276352037, -0.0125933709, 0.0333197229, 0.0156837013, 0.0464100651, 0.0606394857, -0.030623395, -0.0127347084, -0.0512566976, 0.0258779768, -0.0289003775, -0.0229565371, -0.0235215072, 0.012196104, 0.0228788629, 0.0544353127, 0.0076408866, -0.0164886601, 0.0142270979, 0.0602730624, -0.0119376238, -0.0526969954, -0.0497865044, -0.0582147911, 0.0133321891, -0.0166430101, 0.0077006542, -0.0039116549, 0.0042162868, -0.0277889874, -0.0015852005, 0.02060722, 0.0306877438, -0.0465821251, -0.006807561, -0.0765924975, -0.0262494199, 0.0092719272, 0.0260417573, 0.0310283266, -0.0378456265, -0.0272148512, 0.0217161234, -0.0115505755, 0.0065066023, -0.0167311374, 0.0553234331, -0.0060565895, 0.0374161154, -0.0006737719, 0.0581591949, -0.0443218015, -0.0081650028, -0.0389062464, 0.024634853, 0.0118052876, 0.0351436585, 0.0254801866, -0.0309138559, 0.0113403024, -0.0382355638, 0.0367911048, -0.0039098281, 0.0215360057, -0.0003976954, 0.0131815141, -0.0011445237, -0.0522876494, 0.0733997971, -0.007511762, 0.0353958309, 0.0255650971, -0.0165406559, -0.0811165422, -0.0147928745, 0.0411823615, 0.013874013, -0.0350584127, -0.0413166881, -0.0701471716, -0.0097787958, 0.0722705424, 0.0212419555, -0.0046504852, 0.0162738487, 0.0692903772, 0.000001406, -0.0126702143, -0.0235635471, -0.025645474, 0.0463756882, 0.0164917726, -0.0178612918, 0.0042860964, -5.866654267e-33, -0.0414762981, -0.0440915674, 0.0162643567, 0.0084170075, -0.0320844613, 0.0081495047, -0.0645942464, -0.0186737664, 0.0530131646, -0.009295282, -0.0262719169, -0.0374815278, 0.0297494009, 0.0370640717, 0.0181382149, -0.0122301504, 0.0615423582, 0.009380023, -0.0555003732, 0.0541429669, -0.0076278374, 0.0456401668, 0.0185144562, -0.0249604098, 0.0326539017, -0.009036934, -0.0175951868, -0.0383341908, -0.0057742749, 0.0460758395, -0.037866313, 0.0177634414, 0.046950534, -0.0314445309, -0.0227323454, 0.0383689478, -0.0853726119, -0.0239573587, 0.0084985662, -0.0650314316, -0.0539853647, -0.0921362638, 0.0405841805, -0.0296383761, -0.0209818929, -0.0271728262, 0.0373443663, 0.0016403544, -0.018679373, -0.0795834586, -0.0174921397, 0.0299680158, -0.0139144817, -0.0165751558, -0.0349873342, 0.0336709246, -0.0365568362, -0.0103911199, -0.0284054559, 0.0367484689, 0.0476100631, 0.0351299085, 0.0096235257, -0.019792052, -0.0095709376, 0.0479252748, 0.0263791233, 0.0151325408, -0.0349038467, 0.0359021202, -0.0465108156, -0.0101995822, 0.0242105611, -0.0062704291, 0.0482445136, -0.0249417201, -0.0702722371, 0.0769570991, -0.045338124, 0.0313485526, 0.0372812338, -0.0512282439, -0.0248556659, -0.0225379895, -0.0116890026, -0.0069387089, -0.0094300993, -0.032601282, 0.0347655676, 0.0289040152, -0.0208774135, -0.0215320494, -0.0193389114, 0.0298085921, 0.0415044166, -0.002475085, 0.0619120225, 0.0147973243, -0.0466707088, -0.0332808755, -0.0413828939, -0.0058174971, 0.0527070388, 0.0015111043, 0.0347376764, 0.0283620339, -0.0106009552, 0.0035744687, -0.0712746531, 0.0081207044, 0.0355698727, 0.0574583635, -0.0026777559, 0.0160400588, 0.019783061, -0.0029273934, 0.0072781299, 0.0412927642, -0.0216114577, -0.0266415868, -0.0330283344, 0.0044356333, 0.0480505601, 0.0303908624, -0.0288863331, 0.0323655307, -0.0493007228, 0.0621190965, 0.0169118661, 0.0211402308, 0.0141963707, 0.0148258042, 2.681e-7, 0.0296506956, 0.0168570932, 0.0659810007, -0.0112160724, -0.0068841903, -0.0167132858, -0.0129079204, 0.002033964, -0.0012003026, 0.0088021299, 0.0486744866, -0.0204163194, 0.0402696095, 0.0237623919, -0.0898196623, -0.0366513655, -0.0577238724, -0.0394220464, -0.0446254499, -0.0051769861, 0.0750921741, 0.0270319153, -0.0440722853, -0.030597657, -0.0066029304, 0.0374347419, 0.0093280161, -0.0061967229, -0.0260856561, 0.0300570037, 0.0111118592, -0.0573782921, -0.0120603461, 0.0071119261, -0.0041607618, -0.0460798778, 0.0039681154, 0.1038277149, -0.0080846902, -0.0086989515, 0.0439593159, 0.0052215164, 0.0121077644, 0.0057909703, 0.0404599532, -0.0318405814, -0.0296264756, 0.0100665418, -0.0628347993, -0.0104613323, 0.0464184098, -0.0225112289, -0.051510293, 0.0389146022, -0.003148979, 0.0053259414, -0.0181029327, -0.0420134589, -0.0263235979, 0.0064032096, -0.0476345494, -0.0456703044, 0.0084269308, 0.0051920707, -0.0301435441, -0.0883204713, -0.0103505701, 2.822837644e-34, 0.0315141939, -0.0052652904, -0.0034024806, 0.0016502483, 0.0184517764, -0.0280244015, 0.0061663347, 0.0369039178, 0.027128851, -0.0794867724, -0.0471280552 ]
http://arxiv.org/abs/2406.18088v2
http://arxiv.org/pdf/2406.18088v2
LLM-Driven Multimodal Opinion Expression Identification
2024-06-26T05:52:47
2024-06-29T09:55:50
Bonian Jia, Huiyao Chen, Yueheng Sun, Meishan Zhang, Min Zhang
cs.CL
cs.CL, cs.AI, cs.SD, eess.AS
Opinion Expression Identification (OEI) is essential in NLP for applications ranging from voice assistants to depression diagnosis. This study extends OEI to encompass multimodal inputs, underlining the significance of auditory cues in delivering emotional subtleties beyond the capabilities of text. We introduce a novel multimodal OEI (MOEI) task, integrating text and speech to mirror real-world scenarios. Utilizing CMU MOSEI and IEMOCAP datasets, we construct the CI-MOEI dataset. Additionally, Text-to-Speech (TTS) technology is applied to the MPQA dataset to obtain the CIM-OEI dataset. We design a template for the OEI task to take full advantage of the generative power of large language models (LLMs). Advancing further, we propose an LLM-driven method STOEI, which combines speech and text modal to identify opinion expressions. Our experiments demonstrate that MOEI significantly improves the performance while our method outperforms existing methods by 9.20\% and obtains SOTA results.
LLM-Driven Multimodal Opinion Expression Identification: Opinion Expression Identification (OEI) is essential in NLP for applications ranging from voice assistants to depression diagnosis. This study extends OEI to encompass multimodal inputs, underlining the significance of auditory cues in delivering emotional subtleties beyond the capabilities of text. We introduce a novel multimodal OEI (MOEI) task, integrating text and speech to mirror real-world scenarios. Utilizing CMU MOSEI and IEMOCAP datasets, we construct the CI-MOEI dataset. Additionally, Text-to-Speech (TTS) technology is applied to the MPQA dataset to obtain the CIM-OEI dataset. We design a template for the OEI task to take full advantage of the generative power of large language models (LLMs). Advancing further, we propose an LLM-driven method STOEI, which combines speech and text modal to identify opinion expressions. Our experiments demonstrate that MOEI significantly improves the performance while our method outperforms existing methods by 9.20\% and obtains SOTA results.
[ 0.0072719012, 0.1124599501, 0.0066968342, 0.0374459438, -0.0312733315, -0.0080780406, 0.0389023423, -0.0135465711, -0.0154487574, -0.0608288907, -0.085143283, -0.0352993272, -0.019370662, 0.051806204, 0.0128053511, -0.020640431, 0.0198089499, 0.0084927361, 0.0331099853, -0.0095264548, 0.0504361093, 0.0480298847, -0.053929396, 0.033380948, -0.0624211356, -0.0431364998, 0.0210451297, -0.0028608204, 0.015960468, -0.0448594578, 0.0179249328, 0.0385620482, 0.03774257, -0.0201501343, 0.0000018578, -0.0129407393, 0.0051810984, -0.0005693269, 0.0262156893, -0.0075017167, 0.1086449176, 0.0118598752, 0.0029273112, 0.0085407319, -0.012283789, -0.0150202103, 0.0528000481, 0.0833410025, 0.0862571448, 0.0612698682, -0.0297413189, 0.0024556536, -0.0107261427, 0.0086778635, -0.007323124, 0.0083643151, 0.0040501324, -0.0034706865, -0.0070334929, -0.027116945, 0.0048352019, 0.0453994684, -0.0555704311, -0.0128432279, -0.0286155492, -0.0079324059, 0.0182686523, -0.0157915708, -0.034095943, 0.1191934645, 0.0631604865, 0.0223746076, -0.0132952938, 0.016337784, -0.0441832431, 0.0082662459, -0.0234378036, 0.0212193895, 0.0087369997, 0.038461335, -0.0232923776, 0.0444888063, 0.0116069755, 0.0074279313, 0.0093517872, 0.0297613814, -0.0093538249, -0.0195826627, 0.0256874785, -0.005218612, -0.0532736517, 0.0206964351, -0.0041371407, -0.0041337078, 0.0050168503, 0.0064344862, -0.0151767535, -0.0306314249, -0.0298356656, -0.0357425511, 0.041561991, -0.0022837142, -0.0809681565, 0.05018837, 0.0030968608, 0.000395305, -0.008734501, -0.0163454004, -0.0029809903, -0.0258755274, -0.064778313, 0.0011787652, 0.0401250087, 0.0408999883, -0.0689450875, -0.0728444681, -0.0417098179, -0.0133905141, -0.0521825813, 0.0016536592, 0.0121823922, 0.0146626486, 0.0431311503, 0.010388121, -0.0266163126, 0.0592984259, 0.0099868793, 0.0707460344, -0.0510088429, -0.0433575138, 0.029374199, 0.0729838908, 0.0236383807, 0.0003780291, 0.0362601317, 0.0595694967, -0.001842249, -0.0293886121, -0.0301366355, -0.0305659659, -0.0607493371, -0.0494510271, 0.0204507019, -0.0436782725, 0.0336115621, -0.0082956739, -0.0147432918, -0.0358184576, 0.025963489, -0.0104708811, 0.0062096673, 0.0678063706, -0.0002697519, 0.0395276025, 0.0035314604, -0.0070415945, 0.0271373279, 0.0468157381, 0.0369824469, 0.0172858592, 0.045360513, -0.0187817384, 0.0343205705, 0.0504183993, -0.0230092127, 0.0081437062, -0.0031289668, -0.027208101, -0.0557313785, 0.0308320466, 0.0451812819, 0.0871569291, 0.014444856, -0.0088943727, 0.0936070606, 0.031172052, -0.0345647633, 0.0304912739, -0.0192888137, 0.0127074681, 0.0025002602, -0.027840415, 0.0160889309, 0.0335014388, -0.0153015703, -0.0187781006, 0.0455505699, 0.0098410817, -0.0354214348, -0.0303329416, -0.038490206, -0.0233729761, -0.0312711373, 0.024233697, 0.0185807366, -0.01381227, -0.0459964015, 0.0196575578, 0.0322170667, -0.0805181041, -0.0161553007, -0.0370059386, 0.0466858372, 0.0439445674, 0.0112800226, -0.0309717506, -0.0052923788, -0.0288127791, 0.001025549, 0.066819489, 0.0175840091, -0.0129123796, 0.0217903946, -0.0485947616, 0.0139206173, 0.0060939118, -0.0029661753, -0.0058369036, -0.0125673916, 0.0112403799, -0.0149142388, -0.060380917, 0.023410311, 0.0143835796, 0.0378773771, -0.0321793109, 0.0238175951, 0.0222996585, 0.050148014, 0.0370458029, -0.049788624, -0.0062009441, 0.0287372135, -0.0332802609, -0.0308809485, 0.031247355, -0.0308840815, -0.0470481813, 0.0172522627, -0.0345560424, -0.0056475955, 0.0384352766, -0.031044215, -0.0100280941, 0.0315768383, -0.0215370059, -0.007357107, -0.0704443529, -0.0264923275, 0.0190512426, 0.0285382383, -0.0142028611, 0.0252125263, -0.0047863475, -0.032600183, 0.0057521281, -0.0165284071, 0.0105157681, 0.0053891623, -0.0540769994, -0.0047041499, 0.0248680394, 0.0112032546, -0.0124386065, -0.0128605412, -0.004833316, -0.0465805568, -0.1202741638, -0.0253150128, 0.0182390139, 0.0175816845, 0.0213732049, -0.0145738525, -0.0102913855, 0.0263141301, -0.0501531251, -0.0141023016, 0.0473311953, 0.0111181913, 0.014887657, -0.020041585, 0.0192783456, -0.0384548381, 0.029441908, -0.023083616, 0.0348935649, -0.0324474387, -0.0300347209, -0.0125293974, -0.0082197171, -0.0577827953, 0.0494112223, -0.0051707197, -0.0318789668, -0.0012874271, -0.0091100223, 0.0503557213, 0.0117940241, 0.0051355632, 0.0412769057, 0.0342140645, -0.0032159274, -0.0194188487, 0.0153692868, -0.0270754918, 0.0172924679, -0.0145417796, -0.0357508138, 0.0099546975, 0.0321917385, 0.0446875654, 0.0164563861, 0.0047695818, -0.0043386305, 0.0153445853, -0.0370939784, 0.0073118689, -0.0097708823, -0.0304362141, -0.0058006141, 0.0138727613, -0.0211311821, 0.0339379571, -0.0340780877, -0.0255069677, -0.0554249994, -0.0179076008, -0.0592725985, 0.0674446002, -0.0700670779, 0.0158040617, -0.0040140445, 0.0136193857, -0.0108227143, -0.029637564, -0.03624513, -0.0449567288, 0.0243846998, -0.0081947437, 0.0146482661, 0.0361085497, -0.0477250144, -0.0055151968, -0.0630450249, -0.0435716547, -0.0183894504, 0.0111250738, -0.0227749906, -0.0018569086, -0.0004288312, 0.0274257734, -0.0500889905, 0.0092613436, 0.0251804125, 0.0097863721, 0.0603391044, 0.0748944432, 0.0305489991, 0.0070060659, -0.0074718464, -0.0018702849, 0.0383147784, 0.0064748023, 0.0175793748, -0.0474992543, -0.0225504693, 0.0266137961, 0.0664098933, 0.0110313697, -0.0258315485, 0.0415224768, 0.0056385295, 0.048188027, 0.1423469484, 0.0008534009, -0.055335667, -0.0030189797, 0.0194755457, -0.015210011, 0.0134041058, -0.0072363988, -0.0610465705, -0.0712637827, -0.0286118593, 0.0267266184, -0.0222633965, 0.0112336306, 0.0043240483, -0.1195890382, 0.0267268606, -0.0147948368, 0.0243684556, -0.029825259, -0.0090432214, 0.0276333597, -0.0091181556, -0.0041281837, -0.0439240746, -0.0082593244, -0.0146664018, 0.0266005285, -0.0470330939, -0.0573520139, 0.0558075309, -0.0131011233, 0.0609622002, 0.0347296186, 0.0468997993, -0.0295640603, -0.0454146564, 0.041595567, 0.0088969441, -0.0070222579, -0.0145002026, 0.0546632893, -0.0564313754, -0.0198966432, -0.0083917119, 0.0364424139, -0.0174223706, 0.0241100639, -0.0097960122, 0.0448738746, -0.0309011023, -0.0221007764, -0.0349569134, 0.0543337986, 0.0446290597, 0.0668992475, 0.0009180293, -0.0746736377, -0.0129771987, 0.0276563205, 0.0154840443, -0.0021178967, -0.0177253969, -0.0297815204, 0.0113528278, 0.0158001464, -0.0745669976, -0.0115976259, 0.0071950518, 0.0221936963, 0.0200409498, -0.0148662841, -0.0039290995, 0.021729771, -0.0436985046, 0.0431544185, 0.0062390426, 0.0326106139, -0.0310782772, -0.0048573781, 0.0673584789, -0.0194189679, -0.020209223, 0.028650213, -0.0627255216, -0.0390376821, -0.0219694506, 0.015708372, 0.0054808203, -0.0230544247, 0.041984275, -0.0457027406, 0.0382543951, 0.0469585843, -0.0062306663, 0.0315985568, 0.0308662355, 0.0617684461, -0.0106681064, 0.0185226072, -0.0141999023, 0.0602062158, 0.0095717981, 0.0222507138, 0.0052280868, -0.0420147814, -0.0279912017, 0.0102349641, 0.0579135977, 0.0072303261, 0.0095739514, -0.0540818907, -0.0267609134, 0.0081234202, -0.0089504505, -0.0739803165, -0.0008609743, 0.0144422781, 0.0333617665, -0.0560191646, -0.0138265491, 0.0069217104, 0.0425491519, -0.033824347, 0.0712455437, 0.0283722691, -0.0418254919, -0.0351020694, 0.0221791025, -0.0362125859, 0.0304752849, -0.0181107707, -0.0378109366, 0.0140982866, 0.0597739443, -0.000508438, -0.0343174376, -0.0171638448, 0.0374290124, 0.018585328, 0.0142912241, -0.0351532511, 0.0393917784, 0.018613195, 0.0006319028, 0.040588364, -0.0259276796, 0.0133013008, -0.0477426797, 0.0390841179, -0.0504703782, 0.0013119993, 0.0483208038, -0.0078806998, -0.0265857112, -0.0079603763, 0.0472768806, -0.0401286855, 0.0305406507, 0.0099333087, -0.0026683242, -0.0574891195, -0.0218089074, 0.0450610109, 0.0044186241, 0.0027301088, 0.0329694934, -0.018843146, -0.0533618405, 0.0558591485, -0.0297282841, -0.0062887971, -0.0258744191, 0.0147298183, 0.0412069634, 0.0094366036, 0.0440227315, -0.0571139604, -0.0327807963, -0.0503371209, -0.0217522364, -0.0283640064, -5.780324493e-33, 0.0029278041, -0.0224609002, -0.0245398656, 0.022764165, -0.0342773907, 0.0138999745, -0.0178196598, 0.0332518071, 0.008962038, 0.0252256934, 0.0265891701, -0.0284049567, -0.0041528987, 0.048580952, -0.009681467, -0.0100218691, 0.0327992253, -0.019227013, -0.0102747427, 0.005805843, 0.0568935387, 0.0221917424, 0.0537698641, -0.1214790195, 0.0182237215, 0.0127689419, -0.0516611189, 0.0055976696, 0.0388636962, 0.0593744777, -0.0538077019, 0.0969142765, 0.010444981, 0.0178297143, 0.0138917966, 0.0476092324, -0.0135025429, 0.0211546365, 0.0239708051, -0.0082435794, -0.0468607917, -0.0013818711, 0.0021486504, -0.0029920007, -0.0823326334, -0.0032754294, 0.0494966283, -0.0162878335, 0.0396687873, -0.0657447949, -0.0461233035, 0.0109129352, 0.0035196147, -0.0395448916, 0.0414665751, 0.0377902947, 0.0010221801, 0.0189865921, -0.0423243865, 0.0528868623, -0.0644011199, 0.0008030849, 0.0259257462, -0.0026355388, 0.0235135425, 0.1047230363, 0.0163335204, -0.0414873324, -0.044907134, 0.0187376402, 0.025679443, 0.0388610102, 0.0217294022, 0.0599501692, -0.0293571614, -0.0234479178, 0.000494985, -0.006368184, 0.0340191163, -0.0268710144, 0.0778041556, -0.00988416, -0.0357642397, 0.0077421782, -0.0256632362, -0.0142275207, -0.0438267179, -0.0062780008, 0.0327212624, 0.0005136147, -0.0115272934, 0.0006191151, 0.0397799164, -0.0283101499, -0.0174200032, 0.0091134096, -0.0212465897, 0.0686006323, 0.0062991856, 0.0035312171, -0.0519187786, -0.004554362, 0.0233309232, 0.0070549026, 0.0505002365, -0.0328150205, 0.0053310823, 0.0434202105, -0.0735940859, -0.0078166341, 0.0045972466, 0.0042750509, -0.0267467629, 0.0405583382, -0.0255791321, -0.0087512163, -0.0306032635, 0.0161346998, -0.0247039087, 0.073980093, -0.0187010486, -0.0703170002, -0.0141909635, 0.0056335563, -0.0542483367, -0.0091351308, 0.0186892282, -0.0797385052, -0.0356206633, -0.0352941528, 0.0186368879, 0.0264358725, 2.665e-7, 0.0239152927, 0.0464603193, -0.0182615258, 0.0115199136, 0.0319429301, 0.0357935391, -0.0140015483, 0.0111280465, 0.0113490717, -0.0457224548, 0.025605401, 0.000466418, 0.0732664838, 0.0021185393, -0.0814221501, 0.0257291049, -0.008621648, 0.0169932786, -0.0347990841, 0.0342534482, 0.080194816, 0.0920856595, 0.0042910166, -0.0340316296, -0.0179228764, -0.0102820825, 0.0609023795, -0.022673225, -0.0252840314, 0.0044413181, 0.0264422745, -0.0299696214, -0.0035079108, -0.0340619311, 0.0052138008, -0.0144539224, 0.0176758785, 0.0539858192, 0.0204932131, -0.1086676344, -0.0107276682, 0.0747688711, -0.0129180048, 0.0266028717, 0.0673382208, -0.00944228, -0.039919924, -0.1144355685, -0.0070520984, -0.0124243861, 0.018833939, -0.0228108633, -0.0047175097, 0.0108030131, 0.0153530268, -0.0529563837, -0.0566843934, -0.0665062442, -0.0060220403, 0.0061299875, -0.072869055, -0.0351963565, 0.0099664507, -0.0082260435, 0.0147998584, -0.0659708753, -0.0120607084, 2.144498884e-34, 0.0392255448, -0.0311476495, -0.0207118671, 0.0241423938, 0.0330202468, 0.0041369526, -0.0454349481, -0.0348587781, 0.0192001257, -0.0569789745, 0.0062893052 ]
http://arxiv.org/abs/2406.18087v1
http://arxiv.org/pdf/2406.18087v1
EHR-Based Mobile and Web Platform for Chronic Disease Risk Prediction Using Large Language Multimodal Models
2024-06-26T05:51:08
2024-06-26T05:51:08
Chun-Chieh Liao, Wei-Ting Kuo, I-Hsuan Hu, Yen-Chen Shih, Jun-En Ding, Feng Liu, Fang-Ming Hung
cs.SE
cs.SE, cs.AI, cs.CL
Traditional diagnosis of chronic diseases involves in-person consultations with physicians to identify the disease. However, there is a lack of research focused on predicting and developing application systems using clinical notes and blood test values. We collected five years of Electronic Health Records (EHRs) from Taiwan's hospital database between 2017 and 2021 as an AI database. Furthermore, we developed an EHR-based chronic disease prediction platform utilizing Large Language Multimodal Models (LLMMs), successfully integrating with frontend web and mobile applications for prediction. This prediction platform can also connect to the hospital's backend database, providing physicians with real-time risk assessment diagnostics. The demonstration link can be found at https://www.youtube.com/watch?v=oqmL9DEDFgA.
EHR-Based Mobile and Web Platform for Chronic Disease Risk Prediction Using Large Language Multimodal Models: Traditional diagnosis of chronic diseases involves in-person consultations with physicians to identify the disease. However, there is a lack of research focused on predicting and developing application systems using clinical notes and blood test values. We collected five years of Electronic Health Records (EHRs) from Taiwan's hospital database between 2017 and 2021 as an AI database. Furthermore, we developed an EHR-based chronic disease prediction platform utilizing Large Language Multimodal Models (LLMMs), successfully integrating with frontend web and mobile applications for prediction. This prediction platform can also connect to the hospital's backend database, providing physicians with real-time risk assessment diagnostics. The demonstration link can be found at https://www.youtube.com/watch?v=oqmL9DEDFgA.
[ 0.0094233928, 0.0665947944, -0.0188552868, -0.0418096893, 0.0295001585, 0.0113684693, 0.0204817895, 0.0261826105, 0.0461694375, 0.0025784771, 0.0272550154, -0.0282444526, 0.0181926396, 0.0545988977, 0.0075100753, -0.0575620793, 0.0049890801, 0.0035744223, 0.0249345209, 0.0252640061, 0.0185168739, 0.0052686781, -0.0079486202, 0.0673931465, -0.0401181839, -0.0296636764, -0.0139587577, 0.0022222514, 0.0061533018, -0.0772711486, 0.0108107915, 0.0010102604, -0.0101139015, -0.037421748, 0.0000022071, -0.0499078147, -0.051514864, 0.018017767, -0.0173578076, -0.0818092003, 0.070620358, -0.0365174413, 0.0115593104, 0.0154199759, -0.025244901, -0.0216382332, 0.0414696112, 0.0477089584, 0.0362592861, 0.0192200188, -0.0302297585, 0.0214072131, 0.0474454723, -0.0057841083, 0.0276172366, -0.0473318882, 0.0121920733, -0.0153932776, 0.0280046649, -0.062810041, -0.0051736855, 0.0278762728, 0.0029608961, 0.0731501728, -0.0251417924, 0.0217688233, 0.0314383358, -0.0329856239, -0.0343316533, 0.0600177161, -0.0069495193, 0.0047148177, 0.028672386, 0.0140798837, -0.0235600471, 0.0261516161, -0.0238612015, -0.018895369, -0.0489333384, 0.0002410515, 0.0007320795, -0.0242638662, 0.0339598507, 0.0054589566, -0.003908196, 0.0207892749, 0.016745979, -0.0121998712, -0.0092833629, -0.049008213, 0.0433278792, -0.0300857108, 0.0193040632, 0.0540589169, -0.0679940209, -0.0235621408, 0.0015929397, -0.0039951378, 0.065166004, -0.0532102734, 0.048033081, 0.0270119235, 0.0014701828, 0.0257909391, 0.0134075992, -0.0112501532, -0.0395045094, -0.0093854424, -0.0297751538, -0.0115558449, -0.0309041291, 0.0007762522, 0.0108918622, 0.0086597875, 0.0101817958, -0.0323109515, -0.0939862058, 0.0040591415, 0.0048608407, -0.0018873963, 0.0143823689, 0.0432689935, 0.0041157161, 0.028867593, -0.0573059916, 0.040880505, 0.0293985046, 0.0070379255, -0.0006619756, -0.0517884679, 0.0835328996, 0.0033516802, 0.0163157023, -0.0168030448, 0.0339688882, 0.0900921971, 0.0234844778, 0.0002393892, -0.0209198371, -0.0629403368, -0.0087787146, -0.0311166048, -0.0012960203, -0.0649519786, 0.0146460794, 0.0148114637, 0.0228658747, -0.0491491035, 0.0269050766, 0.0140351132, -0.0133733712, 0.0320971012, -0.0511009879, 0.0065901494, 0.0449003577, 0.0121140396, -0.0009285935, 0.0384558812, 0.0357572138, -0.0004126181, 0.0198149085, -0.048390571, 0.0319310129, 0.0462091267, 0.0044114981, -0.0508226529, -0.000538972, -0.0234746765, -0.011565865, 0.0397191197, -0.0532743037, 0.0414066017, 0.0346185341, 0.0219073463, 0.0747167021, -0.0011637381, 0.0362760536, 0.0099760657, 0.0002830335, -0.0524871908, -0.0285619944, -0.0676690415, -0.0025091146, 0.0463889278, -0.0437476039, -0.0191098861, -0.004084493, 0.0177961364, -0.0313555002, -0.0471633077, -0.0387680903, -0.0420492068, 0.0300844144, -0.0006389436, 0.0417578928, -0.0573589616, 0.0061046295, 0.0295847896, 0.0354318619, -0.0100693023, -0.0420671366, -0.0551148988, -0.0108079202, 0.0915206894, 0.0031763301, -0.0093406262, -0.0269933604, -0.006131486, -0.0089041907, -0.063252531, 0.024064837, -0.0074461633, 0.0389751196, -0.0240972657, -0.0157982856, 0.0593845583, 0.0294019263, -0.0402178429, -0.0244144294, 0.0747066736, -0.0236494225, -0.0376629122, -0.0140944524, 0.0043854672, -0.063719064, 0.0108517706, 0.0193704944, 0.0538601279, 0.04599116, 0.0190043394, 0.0349327885, 0.0203779172, 0.0229036659, -0.0413220339, -0.035367921, 0.0291189719, 0.0177273788, 0.020688517, 0.0371691249, -0.0071471934, 0.0316739529, -0.0219414625, -0.0238474067, -0.0158398133, -0.0084285717, -0.0553764999, 0.0106053604, 0.0023906056, -0.0198161136, 0.0758506358, 0.0362218842, -0.0080981022, 0.0012818453, -0.0034311106, -0.0254021417, 0.002742596, -0.056403175, -0.0039161113, -0.0303863026, -0.0041799494, -0.0094893379, 0.0006687263, 0.0235086828, 0.0063574766, -0.010873789, -0.0272385795, 0.0247511379, -0.1405617148, -0.0039416431, -0.0030746593, -0.0160580054, 0.0261847135, 0.0481407382, -0.0072659613, -0.0489144959, -0.0013769431, 0.0259424839, 0.0015049884, 0.0267201997, 0.0193797853, 0.0443655998, 0.0239493344, 0.0015125956, 0.0124644367, -0.0660087988, 0.0494134761, -0.0046680616, -0.0182927772, 0.0115012759, -0.0809887424, -0.056662973, 0.009128009, 0.023211034, -0.0125843408, -0.0231626071, -0.0413142927, 0.0314200558, 0.0343814529, -0.0224417988, 0.0409754999, 0.0330043845, -0.0124200182, -0.0158901177, 0.0202643014, -0.0399771556, -0.0029789554, 0.0072757225, -0.0577345416, -0.0242815297, 0.0527098291, 0.0461149551, 0.0383353047, -0.0038846831, 0.0135836927, -0.0189278219, 0.0322556309, -0.0715892836, 0.073499538, -0.0307952613, 0.0421740711, 0.0006612078, -0.0209558569, 0.0559929125, 0.015283633, 0.0549998842, -0.0222514737, -0.0594786778, -0.0307697244, -0.0082515916, -0.0032657662, 0.0328695141, -0.0254363641, 0.0141130453, 0.0120136812, 0.0022985511, -0.0301129092, -0.0581663288, 0.0191790257, -0.0247566234, -0.0174582526, 0.0127359452, -0.0317616686, -0.0068685599, 0.0000563976, -0.0595021658, 0.0085668983, -0.0510501675, 0.0454617105, -0.0351931229, 0.0109960288, -0.0240919441, 0.061853461, 0.0194484629, -0.0139561919, 0.0334455557, 0.0155525943, 0.0545473769, -0.0033576617, -0.0381574146, -0.0237887446, 0.04253437, 0.0397386067, -0.0065435441, 0.0163991936, -0.0388669372, 0.0424469486, -0.0032236872, 0.009950377, -0.0248387083, 0.0599295385, 0.0107113961, -0.0541534759, 0.0614897646, 0.0675610527, 0.0001159666, -0.0532887205, 0.0114470199, -0.0046321498, -0.0327150449, -0.0076407683, -0.0098147588, -0.0723430589, -0.0590632632, -0.0036444038, 0.0110880984, -0.0065961145, -0.0277620554, 0.0389967524, 0.0180235952, 0.0214920174, -0.0032642793, -0.0311995298, -0.0207052417, -0.0378070883, 0.0292603094, -0.0300831292, -0.0232231095, -0.038070444, -0.0137924915, -0.0235740058, -0.0019707677, -0.1160960495, -0.0611282587, 0.0282019731, -0.0045900908, -0.011139039, 0.0793213248, 0.0535249263, -0.017662257, -0.0332511067, -0.0077151619, 0.0410546288, 0.0422160514, -0.0606771931, 0.0520031564, 0.0198731832, 0.0093245842, -0.0570106059, 0.0550966449, -0.0262791943, 0.0023582957, 0.0226715747, 0.1148285493, -0.0307454821, -0.0272182655, -0.0788403004, -0.0241199434, 0.0600740574, 0.0140745714, -0.0196879692, -0.0756338835, -0.0473859981, 0.0100318985, 0.0345559269, 0.0527323037, -0.0045455405, 0.0277595241, -0.030965697, -0.0446278602, -0.031338647, -0.0082874158, 0.0145103261, -0.0409800783, 0.0245818775, 0.0555013977, 0.0412272438, 0.025995722, -0.047225859, 0.0095907534, 0.0502090454, 0.0577076189, -0.0571266189, -0.0110962689, 0.0429659598, -0.046346046, 0.0190111659, -0.0088487221, -0.054464031, 0.0335375331, -0.0133999242, 0.0586158186, 0.0148296887, -0.0202076938, 0.0107331825, -0.0252140686, 0.0120739462, 0.0109752733, -0.0372366831, -0.0223659296, -0.0237706285, 0.0010610817, -0.0160379205, -0.0067058648, -0.024776062, 0.0019552251, -0.0179175977, -0.0123622147, -0.0212758463, -0.0201454517, 0.0152897891, 0.0432234667, 0.0762950405, -0.0170203689, -0.0557433218, 0.0058077346, 0.0366120413, -0.0062170499, -0.0001906722, -0.0271243546, 0.0136389304, -0.0185161792, 0.0281332433, -0.0016262393, -0.0153259495, 0.0276092067, -0.071098946, -0.0659194291, 0.1026349738, 0.0184827708, 0.0038187096, 0.0474079326, 0.0076330109, 0.0230599679, 0.0381328091, -0.0100027211, 0.0001789541, -0.0187418405, 0.0766709447, -0.0425990187, 0.0153507497, -0.0029180462, 0.0593837388, 0.0118623665, -0.0304684453, -0.0136866029, 0.031701155, 0.0410243981, 0.0286302082, 0.0138506349, -0.0239157192, 0.0084544132, -0.059764266, 0.0262493417, -0.0099113239, -0.0036074677, 0.066027455, -0.0165453125, -0.0162367728, -0.019684568, 0.0672523975, 0.0090156496, 0.0292395297, 0.0091013284, -0.0537161008, -0.0452565216, 0.0358718447, 0.0888553783, 0.0212689862, -0.0842134058, -0.0031032958, -0.03041091, -0.0767690614, 0.0629572719, -0.025307294, -0.0272291861, -0.0245184805, 0.0181871094, 0.0389981717, -0.0011746099, 0.0141809918, -0.0474887639, 0.0268675648, -0.0643187016, -0.007881552, 0.0030849443, -6.255163967e-33, 0.0298932437, -0.0227635149, 0.0086105801, -0.0373248979, 0.0325947367, 0.004732816, -0.0210984293, 0.0054365289, 0.0260168687, 0.0073251608, 0.0111053512, -0.0455717146, 0.008483408, 0.0419822931, 0.014274206, -0.0378712378, 0.0129305702, 0.01862658, 0.0076516634, 0.0066917618, -0.0054638828, 0.0442052819, -0.0016114194, -0.0899773613, -0.038247671, 0.025788974, -0.0278599206, 0.003043887, -0.0318807885, 0.0774253383, -0.0361325778, 0.0807348788, 0.0483530127, -0.0117157074, 0.0015147143, 0.0316162221, -0.0633951351, -0.0302138682, -0.02673853, -0.0019128438, -0.0376966186, -0.0331185162, 0.0416199118, -0.0316053256, -0.0326265916, 0.0034122884, 0.0552369282, -0.0542231277, -0.0145104183, -0.0553100742, 0.026215503, 0.0267419182, 0.0003217878, 0.0227114018, -0.0038472333, 0.0216161981, -0.02968763, -0.0038610757, -0.0624775887, 0.0011650926, -0.0265841838, -0.0154752852, -0.0104906438, -0.0089878868, -0.0260917339, 0.0467277691, 0.0092428895, -0.0795503929, -0.0268284827, 0.0265465621, -0.054869134, 0.0622159205, 0.0017805403, 0.026023645, 0.004512331, -0.0487642065, -0.0870495811, 0.0666209161, -0.0304429643, 0.0084882844, 0.0200596675, -0.0172988772, -0.0253441073, 0.0093262335, -0.0241715368, 0.0240141898, -0.0557100624, -0.0432617664, 0.0342156142, 0.0120622003, 0.0392237194, 0.0170640741, 0.0201667808, -0.0286778305, -0.0370222852, 0.0120548718, 0.034465611, -0.0043862872, -0.0130802039, 0.0132514667, -0.0177021399, -0.0011202835, -0.0255542751, -0.0122442637, 0.0463158339, 0.0036312744, 0.0183580406, 0.0076526743, -0.0516001061, -0.0061828922, 0.0469514616, 0.0024642427, -0.0025491114, 0.0143898251, 0.0472973771, -0.0127326772, -0.0103764012, 0.0921255574, 0.0028941147, 0.0296970159, -0.0752333626, -0.062882565, 0.0075676823, -0.0186212771, -0.0515211932, -0.0042330241, -0.0746605247, -0.0008249083, 0.0190772079, -0.0570211373, -0.0084534446, 0.0034752402, 2.915e-7, 0.0451162122, 0.0489340052, -0.0549967363, -0.0350021422, 0.0507118814, 0.017476907, -0.030440405, 0.0044411737, 0.0034981889, 0.0036092808, 0.0425951369, -0.0083710263, 0.041768983, 0.0497654006, -0.0086344955, -0.0089391703, -0.0485891439, 0.0165059231, -0.0359885767, 0.0564474352, 0.0575483255, 0.0840215981, -0.0206630472, -0.0404542796, 0.0470031537, 0.0456312113, 0.0269196369, -0.0319447927, 0.0082322173, -0.0240108706, 0.003008228, 0.011235889, 0.0236666594, 0.0209109765, -0.0327487364, -0.0426436029, 0.00659568, 0.0644153431, -0.0633197874, 0.0341795161, 0.0397028886, 0.0019229815, 0.0197320171, -0.0190123674, 0.0986128673, -0.0069391257, -0.0524650551, -0.0190333389, 0.0006583302, 0.0090987328, 0.0338397436, -0.0127608869, -0.0488320254, 0.0102677755, 0.0302790664, 0.0357052349, -0.0091559933, -0.0345263481, -0.0345324241, 0.0334925726, -0.0065406687, -0.0361333862, 0.0135677774, -0.0462416969, 0.0109151937, -0.0960729048, 0.0089720665, 2.510464714e-34, 0.0166832004, 0.0502730682, -0.0054398435, -0.0020394633, 0.0290638451, -0.0375955664, 0.0513047688, -0.007081551, 0.0017303657, -0.0550935604, -0.0498853475 ]
http://arxiv.org/abs/2406.18085v1
http://arxiv.org/pdf/2406.18085v1
Multilingual Knowledge Graph Completion from Pretrained Language Models with Knowledge Constraints
2024-06-26T05:46:35
2024-06-26T05:46:35
Ran Song, Shizhu He, Shengxiang Gao, Li Cai, Kang Liu, Zhengtao Yu, Jun Zhao
cs.CL
cs.CL
Multilingual Knowledge Graph Completion (mKGC) aim at solving queries like (h, r, ?) in different languages by reasoning a tail entity t thus improving multilingual knowledge graphs. Previous studies leverage multilingual pretrained language models (PLMs) and the generative paradigm to achieve mKGC. Although multilingual pretrained language models contain extensive knowledge of different languages, its pretraining tasks cannot be directly aligned with the mKGC tasks. Moreover, the majority of KGs and PLMs currently available exhibit a pronounced English-centric bias. This makes it difficult for mKGC to achieve good results, particularly in the context of low-resource languages. To overcome previous problems, this paper introduces global and local knowledge constraints for mKGC. The former is used to constrain the reasoning of answer entities, while the latter is used to enhance the representation of query contexts. The proposed method makes the pretrained model better adapt to the mKGC task. Experimental results on public datasets demonstrate that our method outperforms the previous SOTA on Hits@1 and Hits@10 by an average of 12.32% and 16.03%, which indicates that our proposed method has significant enhancement on mKGC.
Multilingual Knowledge Graph Completion from Pretrained Language Models with Knowledge Constraints: Multilingual Knowledge Graph Completion (mKGC) aim at solving queries like (h, r, ?) in different languages by reasoning a tail entity t thus improving multilingual knowledge graphs. Previous studies leverage multilingual pretrained language models (PLMs) and the generative paradigm to achieve mKGC. Although multilingual pretrained language models contain extensive knowledge of different languages, its pretraining tasks cannot be directly aligned with the mKGC tasks. Moreover, the majority of KGs and PLMs currently available exhibit a pronounced English-centric bias. This makes it difficult for mKGC to achieve good results, particularly in the context of low-resource languages. To overcome previous problems, this paper introduces global and local knowledge constraints for mKGC. The former is used to constrain the reasoning of answer entities, while the latter is used to enhance the representation of query contexts. The proposed method makes the pretrained model better adapt to the mKGC task. Experimental results on public datasets demonstrate that our method outperforms the previous SOTA on Hits@1 and Hits@10 by an average of 12.32% and 16.03%, which indicates that our proposed method has significant enhancement on mKGC.
[ 0.0349740088, 0.0532990061, -0.0066379798, 0.0490576811, -0.0209717844, 0.0038852312, -0.0274775401, 0.0042125843, 0.0645103529, -0.004391463, -0.0102475761, -0.0177923702, -0.0320902541, -0.0026482814, 0.0283546876, -0.0203196555, 0.0326710753, -0.0067801583, -0.0157916341, -0.0782895684, 0.009977066, 0.0608416162, 0.0024265666, -0.005058093, -0.0093408078, -0.0170939919, 0.0197431017, 0.0068735853, -0.0066484683, 0.0236618165, 0.0036283238, 0.0485676974, -0.0151965916, -0.0212498102, 0.0000020955, -0.0495357923, -0.0306080319, -0.0137691684, -0.0098915016, -0.0386386476, -0.0227538701, -0.0202017203, 0.0475576185, 0.0139258048, -0.0465896428, -0.0109171914, 0.0205734428, 0.0172992535, 0.0195273086, 0.0335325524, -0.0076100943, -0.0251571741, 0.000690236, -0.0358927213, 0.0284394566, -0.0614462532, 0.0325913392, 0.033497896, 0.0190871377, -0.0378740169, -0.0348534845, 0.0435616896, 0.0119278366, 0.0129039446, -0.0266409088, 0.0070985756, -0.0157252774, 0.0050114375, 0.0503958017, 0.0645901486, -0.0242640972, -0.0061333152, -0.0149283148, 0.0085947551, -0.0251626801, 0.0155915646, -0.0151056601, -0.0186854191, -0.0297181308, 0.0030663419, -0.0116682108, -0.0271218698, 0.0327781364, -0.0063891066, -0.0542298853, 0.1073396206, 0.0276048668, -0.0072018025, 0.0194850899, 0.031734433, -0.0155742895, -0.0343286246, 0.0144684874, 0.0150720216, -0.0407751724, 0.0588528737, -0.0169465374, -0.0173565783, 0.0236826129, -0.0454554409, -0.0585674644, 0.024668688, -0.0095021222, 0.072476685, -0.0286084712, 0.0113230953, -0.0113630714, 0.0043311524, -0.0617361814, -0.0769086853, -0.0282945205, -0.0149090178, -0.049319692, 0.0148503948, -0.0487422198, 0.0086339023, -0.0412069522, 0.0144321509, 0.0012802088, 0.020699909, 0.0279266071, -0.0000139701, -0.0190520734, -0.0409646668, -0.0334364772, 0.0472030491, -0.0509135202, -0.000338435, -0.0492150448, -0.0208970774, 0.0065619834, 0.0063171545, 0.0229435377, -0.0521341823, 0.0330552086, 0.0376267098, -0.0137336664, 0.001318855, -0.0658495054, -0.0280597564, 0.0121124135, -0.0597241148, -0.0133926738, -0.0635003448, 0.0971385539, 0.034989018, 0.002332686, -0.0362643525, 0.0075982586, -0.0076530771, -0.0529282764, 0.0698792711, -0.0211148337, 0.0199746881, 0.0069157444, -0.0113806194, 0.0053208661, 0.0880841315, -0.0039061303, -0.0078126285, 0.0741216242, -0.0138889132, -0.0235187672, 0.0540807769, 0.0555519871, 0.0170087218, 0.0131875053, 0.0271342322, -0.0159803834, -0.0087585282, 0.0265585855, 0.0394343324, -0.0186812859, -0.0030538868, -0.0073402254, 0.0039163511, 0.0164570566, 0.0163326059, 0.0308327749, -0.023648316, -0.0256932173, 0.0207333248, -0.0150111234, -0.0071030906, -0.0385546722, 0.0301727932, 0.0065343822, 0.0483242646, -0.0299414694, -0.0210068878, 0.0315845795, -0.0220457446, -0.0037841566, 0.0692328587, 0.1067611799, 0.0107365903, -0.0228170417, 0.0484333411, 0.0382713415, -0.0333966762, 0.0612264238, 0.0217367616, 0.0027201783, 0.0463138297, 0.0332356617, -0.0331035145, -0.0532476082, 0.0255166721, -0.0171997398, 0.0079165408, 0.0106693879, -0.0288981795, 0.0092208209, -0.0424164496, -0.0470230691, 0.0289222375, -0.0238253362, 0.0364281721, 0.0300383214, 0.0493851565, 0.0171743389, -0.1188129634, 0.0077314852, -0.0041931388, -0.0410678238, 0.0252706893, 0.0133366277, 0.0341742188, -0.0101987431, 0.0364737958, 0.0186170898, 0.0114933848, 0.0074775908, -0.0354780592, -0.0222659372, 0.0148956552, -0.0049985717, -0.0611703582, 0.0441695489, -0.0209793653, -0.0019538198, -0.036008358, -0.0133199235, -0.0624751449, 0.0092469845, 0.0182543229, 0.0690114275, -0.0501006581, 0.0053801103, 0.0101035153, -0.0290150195, -0.0172193497, 0.0772388056, 0.0132401912, 0.0534531176, -0.0035228203, -0.0469488241, 0.0608909279, -0.0235912707, -0.0509389937, 0.0312908366, 0.0200661439, 0.0087646348, 0.0490529537, -0.0030199322, -0.0518482886, 0.0311180968, -0.0991975293, -0.0104346042, -0.0056535881, 0.0219499804, -0.0237296578, 0.0418726057, -0.012909228, -0.046563793, -0.0588370189, -0.0160751976, 0.0308969878, 0.0371374227, 0.0010829619, 0.0084089255, 0.0142996088, -0.0165581573, -0.0274253488, -0.001166482, 0.0365778394, -0.0232212674, -0.0209470298, -0.0130452374, 0.0465558693, -0.0748832598, 0.058556132, -0.0246697795, -0.0008355185, -0.0092229275, -0.0220484734, 0.0509155504, 0.0038543129, -0.0352318883, -0.0580407307, 0.031642057, -0.0100562116, -0.0589306466, 0.0308253411, -0.0487652496, -0.0250212718, -0.0361059718, -0.0237676576, -0.0101536429, 0.0161212794, 0.0050563719, -0.0607752092, -0.0817093849, -0.0375119559, -0.0038149257, 0.0012635286, -0.0623532534, 0.0355964266, 0.0484654605, 0.0560654663, -0.0207180828, 0.005641168, 0.0913973898, -0.0134570617, -0.0069274935, -0.0112002548, 0.029466942, -0.0294052884, 0.0487308316, 0.0171823651, 0.0494058914, -0.0035252289, -0.0373045206, -0.0556426719, 0.0085212914, 0.034306068, -0.0155062163, -0.0003693408, -0.0035192273, 0.0096791806, 0.0214017667, -0.0142756375, -0.016422743, 0.011432902, -0.0138930865, -0.0021480042, 0.0339511298, 0.0011740102, 0.0126637956, 0.0145994574, 0.0321456678, -0.0106741413, 0.0043678307, 0.0150748221, -0.0164431632, 0.017045157, 0.0565425791, -0.0163555294, -0.025977239, -0.0135932779, -0.0057537733, 0.024977766, 0.0416745916, -0.0779228359, 0.0433885865, 0.0174833126, 0.0638986155, -0.0417973027, -0.0436115302, -0.0270598717, 0.0076573309, 0.0085454835, 0.0374259464, 0.0803328827, 0.0131961647, -0.0270971861, -0.0145512456, -0.0127620045, 0.0123857977, 0.0418216661, 0.0632108226, -0.1204875186, 0.008499587, -0.030192269, 0.0118541056, 0.046081841, -0.0543271601, -0.0123713613, -0.1227808446, -0.019916635, 0.000103225, -0.0502054505, -0.050622113, -0.0270957258, -0.0244496521, -0.0134504056, -0.0278069675, 0.0006733878, 0.032309927, 0.0070077018, 0.0035885428, -0.0103672389, -0.0390328839, 0.0311484728, -0.0501674339, -0.0065762908, 0.0166611504, 0.0306381099, -0.0019955011, 0.0011898865, -0.022716403, 0.0028377825, -0.0176840201, -0.0142139094, -0.0004261043, -0.011424901, -0.0296241939, 0.0233584028, -0.0126556326, -0.0463041924, 0.0044541443, 0.03307027, 0.1185687631, 0.0284527261, -0.0197583288, 0.004018175, -0.0293708872, -0.0009064889, 0.0603931583, 0.001801089, 0.000338329, -0.0123134758, -0.0276326519, 0.0392116681, -0.0458787568, 0.0223274976, 0.0475841425, -0.0508951396, 0.0101131527, -0.0326031148, -0.0588655099, 0.0702436045, 0.0205073953, 0.0101075089, 0.0508431569, 0.0241949707, -0.0116226096, -0.0225504544, -0.0090884389, -0.0465949178, -0.0021004337, -0.0390368588, -0.0094783381, 0.0146095436, -0.0365180671, -0.0139185498, -0.0264459588, -0.0143916607, -0.0550218932, -0.0696898326, 0.0380874015, 0.0091482149, 0.0564110056, -0.0633673072, -0.0537009947, 0.0026704338, 0.0219549481, -0.0341186672, -0.0232662968, 0.0269423779, -0.0170643087, -0.0123545229, 0.0524557978, -0.0603917129, 0.0098380735, 0.0026882696, 0.0891207978, -0.0487664752, -0.0456990376, 0.0414617248, 0.0539470315, -0.0012837782, -0.0337659642, 0.002155592, -0.0601475351, 0.0184569173, -0.045893345, -0.041414652, -0.0298374239, -0.0165438689, 0.0250697639, -0.0343980268, 0.018359093, -0.0504548028, -0.0101775862, 0.0106954202, -0.0278006401, -0.0264059901, 0.0404067338, -0.0152809452, 0.0772735178, 0.0104956161, 0.0059340643, 0.0256129634, 0.0309697278, 0.0268087536, -0.0408263579, 0.0201513134, 0.0061861398, -0.0167141836, -0.0141254738, -0.0381352901, 0.0022264188, -0.0249940604, -0.0135126309, 0.0272949003, 0.1060412377, -0.0047783367, 0.0296749752, 0.0007424897, -0.0166602954, -0.0067950371, 0.0350318886, 0.0490367338, -0.0467612073, 0.0332076736, 0.0174380541, -0.0126419626, 0.0333596393, 0.1049819216, 0.0186093152, 0.05313766, -0.0114663858, -0.0430725627, -0.0181761291, -0.0381678678, -0.0035263819, 0.0238676053, -0.0450468808, 0.0432758965, -0.0297998972, -0.1151567101, -0.0086307144, -0.0111285541, -0.0596453361, 0.0263967123, 0.0263526682, 0.0854268447, 0.0313847363, -0.04069978, -0.0541383997, 0.0186976735, -0.0259185154, -0.0001421842, 0.0543057173, -6.050193017e-33, -0.0112870838, -0.0214096494, 0.0354948342, -0.0149877509, 0.0265157484, -0.017906785, -0.0238292236, -0.0077121314, -0.0541688837, 0.0075757727, -0.0311458912, -0.011318963, 0.0117198639, 0.0283475108, 0.0024698311, 0.0110700633, 0.0044679451, 0.0191102549, 0.0419547409, 0.0206315871, 0.0470594727, 0.0522955097, 0.092309542, -0.0356390849, 0.0459534153, -0.0351662077, -0.0012047829, 0.0499063395, -0.0087126372, 0.0544862375, -0.0132218674, 0.0613787286, 0.0227824692, -0.0162556823, -0.0035840995, -0.0350886956, -0.1030373126, -0.0676927865, 0.049854178, 0.0381526686, -0.0273021441, -0.0066391751, 0.0250837207, -0.0330177359, 0.0526209176, 0.0143221337, 0.0391539335, -0.0108848484, -0.0551600717, -0.0566191971, -0.0350765809, 0.019423591, 0.0133863306, 0.0246674679, 0.0274384264, 0.0682310686, 0.0094992816, -0.0254801512, -0.0403311141, 0.0369030274, 0.0065263109, 0.0042739292, 0.0466133691, 0.0012358271, 0.0114708608, -0.0169685055, -0.0044011055, -0.0103700068, -0.0094013168, 0.0393044762, 0.0174088534, 0.0329232812, -0.0075297682, 0.065587759, -0.0364536382, -0.0241162442, -0.082329832, 0.0225293655, -0.0192624312, 0.0250227936, -0.0114283869, -0.019046735, -0.0381626301, -0.075261116, -0.049579531, -0.0166435409, -0.0371452682, -0.0447048023, 0.0135155721, -0.0144773573, -0.0067843171, -0.0154931992, 0.0126528516, -0.0389319882, -0.0312342159, 0.0629898906, 0.0136238439, -0.0215613395, 0.00230952, 0.0874244124, -0.041640792, -0.0127712842, -0.0203029551, -0.0661217719, -0.0277087297, 0.0307815634, -0.0014870628, 0.0064155762, -0.0133676119, 0.0058881501, -0.030598972, -0.0010191527, -0.0260933433, -0.0110648992, 0.0489130393, 0.0298067648, 0.0159236919, 0.0342183821, 0.0706083775, 0.0108887376, -0.0242243335, -0.0073625688, 0.0255014356, 0.0307778344, -0.0095583145, -0.0046457145, -0.0095082689, -0.0195015669, 0.0234573577, -0.0024386758, -0.0023333374, 0.0137743549, 2.789e-7, 0.0638517514, -0.0076374304, 0.0056281011, 0.0594260208, -0.0310088359, 0.006489581, 0.0031495627, -0.0018709025, 0.0214295294, 0.0017130157, 0.0360360183, 0.0007272869, -0.0049557751, 0.0630613863, -0.0917111486, -0.0044757044, 0.010178973, -0.01750491, -0.0423787609, -0.0030114313, 0.0376700945, 0.0248803012, 0.0207152385, 0.0116788112, -0.0505203791, -0.0559968911, 0.068879202, -0.0021693171, -0.0483097248, 0.0478205383, 0.0337735936, 0.0391662195, 0.0184253156, 0.0700895637, -0.0328594856, 0.0176624022, 0.0501450226, 0.0332739875, 0.0076082624, -0.0305155143, 0.0139324041, -0.021031199, 0.004786477, -0.0694415197, 0.0939709023, 0.0068076323, 0.0250457134, 0.0616995357, -0.0271456167, -0.0165904686, 0.0151889026, -0.032659512, -0.0355139077, -0.0108821662, 0.0453190655, 0.013973698, -0.0080354372, -0.0550016314, 0.0054923329, 0.0407043546, -0.0091925431, -0.0237102695, 0.0260705911, -0.0349688381, -0.0190747101, -0.022045223, 0.0250083357, 2.368350432e-34, 0.0292665213, 0.0148844337, 0.0079022935, 0.0111244088, 0.0483023711, -0.041250933, -0.0419603698, 0.0244580545, 0.0211398322, -0.0732031912, -0.0498618782 ]
http://arxiv.org/abs/2406.18082v1
http://arxiv.org/pdf/2406.18082v1
Octo-planner: On-device Language Model for Planner-Action Agents
2024-06-26T05:40:10
2024-06-26T05:40:10
Wei Chen, Zhiyuan Li, Zhen Guo, Yikang Shen
cs.CL
cs.CL, cs.HC
AI agents have become increasingly significant in various domains, enabling autonomous decision-making and problem-solving. To function effectively, these agents require a planning process that determines the best course of action and then executes the planned actions. In this paper, we present an efficient on-device Planner-Action framework that separates planning and action execution into two distinct components: a planner agent based on Phi-3 Mini, a 3.8 billion parameter LLM optimized for edge devices, and an action agent using the Octopus model for function execution. The planner agent first responds to user queries by decomposing tasks into a sequence of sub-steps, which are then executed by the action agent. To optimize performance on resource-constrained devices, we employ model fine-tuning instead of in-context learning, reducing computational costs and energy consumption while improving response times. Our approach involves using GPT-4 to generate diverse planning queries and responses based on available functions, with subsequent validations to ensure data quality. We fine-tune the Phi-3 Mini model on this curated dataset, achieving a 97\% success rate in our in-domain test environment. To address multi-domain planning challenges, we developed a multi-LoRA training method that merges weights from LoRAs trained on distinct function subsets. This approach enables flexible handling of complex, multi-domain queries while maintaining computational efficiency on resource-constrained devices. To support further research, we have open-sourced our model weights at \url{https://huggingface.co/NexaAIDev/octopus-planning}. For the demo, please refer to \url{https://www.nexa4ai.com/octo-planner}.
Octo-planner: On-device Language Model for Planner-Action Agents: AI agents have become increasingly significant in various domains, enabling autonomous decision-making and problem-solving. To function effectively, these agents require a planning process that determines the best course of action and then executes the planned actions. In this paper, we present an efficient on-device Planner-Action framework that separates planning and action execution into two distinct components: a planner agent based on Phi-3 Mini, a 3.8 billion parameter LLM optimized for edge devices, and an action agent using the Octopus model for function execution. The planner agent first responds to user queries by decomposing tasks into a sequence of sub-steps, which are then executed by the action agent. To optimize performance on resource-constrained devices, we employ model fine-tuning instead of in-context learning, reducing computational costs and energy consumption while improving response times. Our approach involves using GPT-4 to generate diverse planning queries and responses based on available functions, with subsequent validations to ensure data quality. We fine-tune the Phi-3 Mini model on this curated dataset, achieving a 97\% success rate in our in-domain test environment. To address multi-domain planning challenges, we developed a multi-LoRA training method that merges weights from LoRAs trained on distinct function subsets. This approach enables flexible handling of complex, multi-domain queries while maintaining computational efficiency on resource-constrained devices. To support further research, we have open-sourced our model weights at \url{https://huggingface.co/NexaAIDev/octopus-planning}. For the demo, please refer to \url{https://www.nexa4ai.com/octo-planner}.
[ 0.0264776889, -0.0079462109, -0.0062274989, -0.0298971944, -0.0138132945, -0.0038048706, 0.0347914882, -0.0178565271, 0.0226947907, -0.0345070362, -0.0328278318, -0.0028891801, -0.0246406626, 0.0449838564, -0.0049496614, -0.028868638, 0.0251017977, -0.0022547676, 0.0047547161, -0.0082420502, -0.0118044196, -0.0010042727, 0.0179650635, 0.0547613837, 0.0010865119, 0.0050190478, -0.0532017462, 0.0420856848, 0.0177546553, -0.0404550172, 0.0167460181, 0.0139630055, 0.0340750292, 0.0218974017, 0.0000021932, -0.0375405513, -0.0455953069, -0.0248884279, -0.0009219166, -0.0526607186, 0.0196819697, -0.0168853682, -0.005509648, 0.0539219528, -0.075888373, 0.0180725809, 0.0119723445, 0.0460641198, 0.0253705475, 0.0969111919, -0.0103668403, -0.0023312559, 0.0138240438, -0.016344076, -0.0207110494, -0.0503280051, -0.0286923535, -0.0577487573, 0.0153292688, -0.0246186387, 0.026888283, -0.0143567203, 0.044886332, 0.0533123799, -0.0473230034, 0.0005358216, -0.019894341, -0.0293945391, -0.0311984085, 0.0308032464, -0.0476471037, -0.0071096797, 0.0690247715, 0.0013034885, -0.0329591222, -0.0245677345, -0.0103018098, -0.0196563657, -0.0140412664, -0.0142023787, -0.0180670507, -0.0358477272, 0.0266209953, 0.005348213, -0.0255944449, 0.0992538854, -0.0103590181, -0.0123572145, 0.0081166374, -0.0311312638, 0.0014499432, -0.0669741482, 0.0291941892, 0.0873601437, -0.0780169591, -0.0044215154, -0.0393590704, 0.0053214333, -0.012716095, -0.0296726897, 0.0671341419, 0.0374198556, 0.037961565, -0.0039754021, -0.0066301934, 0.038388852, 0.0008322199, 0.0185570512, -0.0404874831, 0.0317991152, -0.0426954404, 0.0313302092, -0.0292858705, 0.0486864373, -0.0637522042, -0.0277531594, -0.0168418158, 0.0129299136, 0.0037633611, 0.0166280977, 0.0143073425, 0.0218773149, -0.0036095395, -0.0279084817, -0.0469820648, -0.0187006108, -0.0360212773, 0.0588790663, -0.0049270457, -0.0496880263, -0.0000052873, 0.0097600417, -0.0114878844, 0.0016857146, -0.0004595996, 0.0772571117, -0.0126566114, -0.0544723123, -0.0399162173, -0.0670689568, 0.0412502177, -0.0978202596, 0.025216205, -0.048826009, 0.032996878, -0.0421866812, -0.0135959862, -0.0216264967, 0.024459213, 0.0275388733, 0.0436586253, 0.0281755049, -0.0089461822, 0.0278856419, 0.0326754041, 0.0068254904, 0.0046390146, 0.0739231855, 0.0681904703, 0.0519359857, -0.0174672119, 0.0100811627, -0.0109540466, 0.0345012732, 0.0080720307, -0.020349253, 0.0517412834, 0.0051416634, -0.0061726938, 0.039156042, -0.0230398476, -0.003896825, -0.0163924433, 0.0184958111, 0.0352323353, 0.0216975696, 0.0252820179, -0.0057205344, 0.0200205315, 0.0103586474, -0.0097501101, -0.0709176213, -0.072817333, 0.036158666, -0.0134575339, 0.012147638, -0.015913412, 0.0281111486, -0.017380422, -0.0475246422, -0.0121437693, -0.0262319297, -0.0051923045, 0.0350796133, 0.0101804519, 0.0008366537, -0.0912872776, 0.0233707447, -0.0154050617, -0.0321342982, -0.0386187397, -0.0692011043, 0.0488058999, 0.0757978186, 0.0187125187, -0.0393645354, -0.0782057941, 0.009999766, 0.0290416144, 0.0174665917, -0.0000242395, -0.0156810321, 0.0360821448, -0.0496046729, -0.0486197956, 0.0061862799, 0.0164703671, 0.0164526459, 0.0434346348, 0.0231246147, -0.0139564108, -0.0125488583, -0.0132973231, 0.0178330988, -0.016075043, -0.0546735823, 0.0208228342, 0.0338170491, 0.0321542397, -0.0034406325, 0.0022768914, -0.0195088349, 0.0477534495, -0.0323736593, 0.0225938279, -0.0310743693, -0.0439648442, 0.0354081988, 0.0399640389, 0.0172114205, 0.0191959087, 0.0480798781, -0.0406489968, -0.0620720163, -0.0077620451, 0.005127877, 0.0417138115, 0.0092659835, 0.0154508362, 0.0404693, -0.0166833811, 0.0087365694, 0.0352224335, 0.0156008247, -0.0325894989, 0.0060236254, 0.0288739484, 0.0261916108, -0.029627895, 0.0380302258, -0.0168548059, -0.0162420757, 0.0232960656, 0.0153730651, 0.0519357175, 0.0031459993, -0.0506373681, -0.1046575457, 0.0227008332, -0.065262951, -0.0016453764, 0.0005823061, 0.0377703086, 0.0073969122, -0.0270252377, 0.0105935475, 0.0516496859, -0.0179866683, 0.0276535228, -0.0338680074, 0.019132996, -0.04375007, -0.0411939658, 0.0168518759, -0.0154641559, -0.0130440062, 0.0023157692, 0.0128269522, -0.0226587206, 0.0083504934, -0.0432672091, 0.0231417343, -0.0174843892, -0.0287491884, 0.0420951918, -0.0191788189, -0.0030389915, 0.0521168709, -0.0215749182, -0.0097777992, 0.0666146725, -0.008245986, -0.0578168407, 0.0523239896, -0.027749449, -0.0251964349, -0.0706527159, 0.0138715459, 0.0081358086, 0.0534457229, -0.0290043447, 0.0239412673, -0.0535220988, -0.0352444276, 0.0307913851, -0.0526660532, 0.0361146219, 0.0350226052, -0.024677055, 0.0670785755, -0.0176932737, -0.022405263, 0.0249032751, 0.0156702548, -0.0012477476, -0.0213515777, -0.0409236401, -0.0265822578, 0.0748452544, 0.0036777009, 0.0166922379, 0.007066512, -0.0190027803, -0.0261653811, -0.0121343229, -0.0122985691, -0.031193817, 0.0351840146, -0.0564367287, 0.007213877, -0.0113175791, -0.0779563934, -0.0045607262, 0.0137204034, -0.0381662324, 0.0225015711, -0.0524043702, -0.0335089639, -0.0608281754, -0.006934721, -0.0163472313, 0.0879355296, -0.0187204629, -0.0090845712, -0.0348425433, -0.0086192489, 0.0523827001, 0.0721799061, -0.0452063531, 0.0308905095, 0.0281397421, 0.0059860139, 0.0821255445, -0.0290765241, -0.064318791, 0.058566235, 0.0495574139, -0.0196559653, -0.0064414376, 0.077516228, 0.0640368983, -0.0559461601, -0.0153552229, 0.0811363831, 0.0184234604, -0.0279861167, -0.0003522524, -0.03786394, 0.0477366187, 0.0292972401, 0.0582767874, -0.0264285263, -0.0615446083, 0.0119230645, 0.0196359511, 0.0196579453, -0.0160150956, -0.0161085203, -0.096290648, 0.0381409526, -0.0046212412, 0.0142981801, -0.0030303416, -0.0464058705, 0.0049862498, -0.0167448428, 0.0055692764, 0.0010163322, 0.0051554614, -0.0199299883, 0.0002283953, -0.0215294529, -0.0108569441, 0.0380123183, -0.0046113352, -0.0063369344, 0.0884395316, 0.0219982509, 0.0202486273, -0.0055291429, 0.0310880356, 0.0543860272, 0.0221819356, -0.0301882066, -0.030695159, -0.0091994414, 0.0217330791, 0.007258466, 0.0365116671, -0.032950338, -0.0442931578, 0.0336903743, 0.0971688032, -0.0129398992, 0.0189018827, -0.0271295384, -0.0126250153, 0.0355293341, -0.0049335314, -0.0457732342, -0.0195859671, -0.0737339258, -0.0071588489, -0.0442462228, -0.0363545306, -0.0080525968, -0.0146905789, -0.0055715637, -0.0237704311, -0.0364800468, -0.0330775939, 0.0620614477, 0.0411644615, 0.0423097052, -0.038904652, 0.0088070873, 0.0664222687, -0.0404829532, -0.0232895371, 0.0131433029, 0.0347394086, -0.045874659, -0.0061007636, 0.0515221097, -0.0499988906, 0.0134402988, 0.051169876, -0.0496981703, -0.0513163432, -0.0050860718, 0.079187654, 0.0244494881, 0.0226705093, -0.0113004427, -0.0167519674, 0.0360246375, 0.062736243, -0.0593932644, 0.0438033938, 0.0267286822, -0.0019267006, 0.0031810021, 0.0527645051, -0.0942999944, 0.0160373412, 0.0034765408, 0.0520120896, -0.0988137946, -0.029965248, 0.0187228806, -0.0143562239, 0.0201737676, -0.0125268875, -0.0250121672, 0.0072186063, 0.0510280766, 0.0201923028, 0.017001858, -0.0483775139, 0.0157283135, -0.0447189063, -0.0292561725, 0.061992377, 0.0284045674, 0.0354284756, -0.0411471538, -0.0083808135, 0.0160383172, 0.0075952811, 0.0068322294, 0.0049466593, 0.0157261621, -0.0006055619, 0.0301162414, 0.0130080236, 0.0147638498, -0.0227754936, 0.0123386895, -0.0133070322, 0.0352594927, -0.0467884354, 0.0199112594, -0.0376435146, -0.1035066471, -0.0009035363, 0.0073643094, 0.0316772386, 0.0109326961, 0.0164039005, -0.0498343818, 0.0141672567, -0.0254441779, -0.0113950158, -0.0291513279, 0.0101156887, 0.0384441465, 0.0242148414, 0.0018716191, -0.0215162933, 0.0524806194, 0.023097543, 0.0836546272, 0.0141002918, 0.0055143978, -0.0624106489, -0.0624362119, 0.041277986, 0.0515958145, -0.0000329606, -0.0332290344, -0.0179844499, -0.0274409112, 0.0211860258, 0.0086457571, -0.0072336164, -0.048330307, 0.0444117524, 0.0320971906, -0.0154691488, -0.0515380315, -0.019032184, -0.0829264596, -0.0563989356, -0.0068133329, 0.008241985, -5.8301908e-33, -0.0110687111, -0.0131678749, 0.0708522201, -0.0322159231, 0.0062051686, -0.0221150257, -0.0166237, -0.0053673857, -0.0334828198, -0.0142089678, -0.0553187728, 0.0163784567, 0.0263239592, 0.0022900682, 0.0070111793, -0.0011842801, 0.0392646007, -0.0426855609, -0.0407771654, -0.0043822699, 0.0244451743, 0.0288075656, -0.0055935793, -0.0094522759, -0.0525445901, 0.0274208076, -0.0401064903, 0.0015002416, -0.0446660146, 0.0383721218, 0.0055060592, 0.0012277098, 0.0160663426, -0.0068332893, -0.0121530062, 0.0327094495, -0.0585012026, -0.0602377951, 0.0119472118, 0.0249908864, -0.0487901345, -0.0460730866, 0.0364243314, 0.0237053558, -0.0677404925, -0.0044434252, 0.0367059782, 0.0123578953, -0.0506939366, -0.0204391014, -0.0334437788, 0.0113938274, 0.0357430652, -0.0469603129, 0.0500465967, 0.0273000635, -0.0218906756, -0.0205070246, -0.0542725287, 0.028406458, -0.0667303726, 0.0464828908, 0.0175960939, 0.0229295008, -0.0061011021, 0.0507411174, 0.0413183495, -0.0054394514, -0.001641516, 0.0317189023, -0.0196114685, 0.1138628572, -0.0317578427, 0.0170754511, -0.0033481834, -0.0323898867, -0.0513649583, 0.0253376644, 0.0070867869, 0.0359379463, 0.0108913509, -0.0046486934, -0.0230485257, 0.0170611031, -0.0053246906, 0.0007152594, -0.0267137475, -0.0255418681, 0.0371720903, 0.0565355793, 0.0187132936, -0.0191667583, 0.03594964, -0.0154240718, 0.0466208868, -0.0123383971, 0.007668904, 0.0347765461, 0.017898431, 0.037329413, -0.0410984233, -0.0625697151, -0.0288432334, -0.0040803412, 0.0369532853, -0.002036646, -0.0145633854, 0.1006619632, -0.0482902452, -0.0114287101, -0.0317448117, 0.0348135829, 0.0189720038, 0.0185138825, 0.0512962751, -0.0199463367, 0.0033984969, 0.0421093181, -0.0530693009, -0.0899900943, 0.0022594328, -0.0009921381, 0.0184999108, 0.0175025202, -0.0249412935, -0.0064460356, -0.0852368474, 0.0142809674, -0.0176289827, -0.0111721372, -0.0195463374, 0.0636078119, 3.093e-7, -0.063514784, 0.0051243743, -0.0466932654, 0.0462157056, -0.0027402304, 0.0445016399, 0.0312095955, 0.0217256919, -0.0238654725, 0.0055606216, 0.0184412878, 0.032633394, 0.0328839049, -0.0052626315, -0.084608607, 0.01170089, 0.0254891273, 0.0082353158, -0.0525064021, 0.0130423848, 0.0172422584, 0.0875333175, 0.0198981557, -0.0118394867, 0.0346406661, -0.0065155323, 0.0087954439, -0.007851515, 0.0289163142, -0.0059367856, -0.0456995443, 0.0054308609, -0.0554930046, 0.0019350155, -0.0075181723, 0.0248340853, -0.0308142155, 0.1014172286, -0.0146191996, -0.0545680299, 0.0126664937, -0.0364210159, -0.0086581241, -0.0296861771, 0.0461439937, 0.0434548147, -0.0217524748, 0.0351475142, 0.034449093, -0.0196388122, -0.0210181866, 0.0086210035, 0.0224974975, -0.00984977, -0.0077804201, -0.0256533865, -0.0069721323, -0.0663466081, 0.0343239009, 0.0427815504, -0.0128173893, -0.0386489071, -0.0628535151, 0.0375520773, 0.0390415564, -0.0667079613, -0.0070551662, 2.871255384e-34, -0.0095579941, -0.0127337584, 0.0132132685, 0.0789709464, 0.0145887043, 0.0245988108, -0.0022592642, 0.0267441068, 0.0509871356, -0.0625177845, -0.0281194095 ]
http://arxiv.org/abs/2406.18078v1
http://arxiv.org/pdf/2406.18078v1
Self-Training with Pseudo-Label Scorer for Aspect Sentiment Quad Prediction
2024-06-26T05:30:21
2024-06-26T05:30:21
Yice Zhang, Jie Zeng, Weiming Hu, Ziyi Wang, Shiwei Chen, Ruifeng Xu
cs.CL
cs.CL, cs.AI
Aspect Sentiment Quad Prediction (ASQP) aims to predict all quads (aspect term, aspect category, opinion term, sentiment polarity) for a given review, which is the most representative and challenging task in aspect-based sentiment analysis. A key challenge in the ASQP task is the scarcity of labeled data, which limits the performance of existing methods. To tackle this issue, we propose a self-training framework with a pseudo-label scorer, wherein a scorer assesses the match between reviews and their pseudo-labels, aiming to filter out mismatches and thereby enhance the effectiveness of self-training. We highlight two critical aspects to ensure the scorer's effectiveness and reliability: the quality of the training dataset and its model architecture. To this end, we create a human-annotated comparison dataset and train a generative model on it using ranking-based objectives. Extensive experiments on public ASQP datasets reveal that using our scorer can greatly and consistently improve the effectiveness of self-training. Moreover, we explore the possibility of replacing humans with large language models for comparison dataset annotation, and experiments demonstrate its feasibility. We release our code and data at https://github.com/HITSZ-HLT/ST-w-Scorer-ABSA .
Self-Training with Pseudo-Label Scorer for Aspect Sentiment Quad Prediction: Aspect Sentiment Quad Prediction (ASQP) aims to predict all quads (aspect term, aspect category, opinion term, sentiment polarity) for a given review, which is the most representative and challenging task in aspect-based sentiment analysis. A key challenge in the ASQP task is the scarcity of labeled data, which limits the performance of existing methods. To tackle this issue, we propose a self-training framework with a pseudo-label scorer, wherein a scorer assesses the match between reviews and their pseudo-labels, aiming to filter out mismatches and thereby enhance the effectiveness of self-training. We highlight two critical aspects to ensure the scorer's effectiveness and reliability: the quality of the training dataset and its model architecture. To this end, we create a human-annotated comparison dataset and train a generative model on it using ranking-based objectives. Extensive experiments on public ASQP datasets reveal that using our scorer can greatly and consistently improve the effectiveness of self-training. Moreover, we explore the possibility of replacing humans with large language models for comparison dataset annotation, and experiments demonstrate its feasibility. We release our code and data at https://github.com/HITSZ-HLT/ST-w-Scorer-ABSA .
[ 0.0422653221, 0.0595722422, 0.0237011481, 0.0324650854, 0.0167903099, 0.0061978824, 0.0052596061, 0.0012423536, -0.0138576375, -0.0188793503, -0.1055848822, -0.0061940043, -0.0445337892, 0.0576477163, -0.0167377405, 0.0001211011, 0.0315147564, 0.0216851495, 0.009749148, -0.0115196751, 0.0135789001, 0.0216412731, 0.0063019851, 0.0083876513, -0.0281840619, -0.0124060847, 0.027655853, 0.0057278504, 0.0263148099, -0.0282158107, 0.0365205966, 0.0717884153, -0.0048821238, 0.0241961684, 0.0000020307, -0.0123394588, -0.0180421341, -0.0254055336, 0.0401036516, -0.0092133693, 0.0678401142, -0.0075273709, 0.0257630926, 0.0141714513, -0.0626992583, -0.0846990794, 0.0245875549, 0.0835018232, 0.0456900373, 0.0565823466, -0.0301335063, -0.0357008204, -0.0216805451, -0.0540342294, -0.0095917284, -0.0707109421, -0.0040708994, 0.0186315831, -0.0346367769, 0.083947733, 0.0090790438, -0.005383777, 0.0303110648, 0.0009777856, 0.0325636864, 0.0269062445, 0.0156797804, 0.0345663354, -0.0043008551, 0.0521791279, 0.0426378772, -0.0433583632, 0.0006771907, -0.0132910674, -0.0007999793, 0.0064684483, 0.0280842204, -0.0264562201, 0.0107110962, 0.0475769565, 0.0497290269, 0.0553464107, 0.0126085142, 0.0031476247, 0.0375483744, 0.0315482765, -0.01366679, -0.033113949, 0.0262107197, 0.0825556591, 0.0333446562, -0.0471424721, -0.0025945543, -0.0171383191, -0.0199066531, -0.0020857868, -0.0153020276, 0.0362939648, 0.0134589029, -0.0593392327, 0.0390451811, -0.0274448507, -0.0113985231, 0.0370157994, 0.0089452509, -0.01047544, 0.0424560979, 0.0090400251, -0.0511255935, 0.0404473282, -0.0940198228, 0.0179662239, -0.0100969505, 0.028298391, -0.0100963861, 0.0357882939, -0.0879221708, -0.0234795976, -0.0253532752, -0.0031070861, 0.0607837327, 0.0032823898, -0.0171645693, 0.0170662161, -0.0329979956, 0.031324219, -0.0462625623, 0.034710627, -0.0475374088, 0.0170108099, -0.0040608151, -0.0015917587, -0.0108915158, -0.0349569768, 0.0206183791, 0.0183250662, 0.0267376136, -0.0399668142, 0.0136871981, -0.0222945362, -0.0179855991, -0.0623498447, -0.0292552784, -0.0783535838, 0.0446127728, -0.0238539074, -0.0135284001, -0.0150367757, 0.0112262229, 0.0182946753, 0.0101730023, 0.0600642189, -0.0519475341, -0.0229930431, -0.0166466422, 0.001457066, 0.0223048404, 0.0497879609, -0.0418870747, 0.006730638, 0.0152233271, -0.0029561978, 0.0480912589, -0.0046973778, -0.0019025324, 0.0075775301, 0.0072059599, -0.0059187319, -0.0162023418, 0.0084745176, 0.0303892568, 0.0048361346, -0.0066139149, -0.0636623949, 0.0174484272, -0.0234181266, 0.0088547999, 0.0699818283, -0.0007663002, -0.0004754872, -0.000328239, -0.0054600071, -0.0069143977, 0.0421466641, -0.0193107314, -0.0204700921, 0.0195483211, -0.0085016219, -0.0460577086, -0.0716313347, -0.0101736067, -0.0071740439, -0.0252133273, -0.025941737, 0.0680926591, 0.0396923274, -0.0264208857, -0.0435536802, 0.0463383384, -0.0359863974, -0.0316292457, -0.0436525419, 0.0230175219, 0.0728832558, 0.0209862217, -0.0420925543, -0.0236805286, 0.0007349902, -0.059160769, 0.0215664655, 0.0603536777, -0.0437365957, -0.0314178169, -0.0116178142, -0.0195386391, -0.0114920419, 0.0012020895, -0.0029234628, 0.0149088334, 0.0362501591, -0.0174615793, -0.0538821109, -0.0330384746, 0.0178106185, 0.0080769975, -0.0383513048, 0.0463835336, -0.0087342849, 0.0385831296, 0.0531579591, -0.0073507591, -0.0062184865, 0.0536476485, -0.0288187731, -0.0701742321, 0.0150496326, -0.0360337123, -0.0548265874, 0.0317809843, -0.003361146, -0.0057544126, 0.0234636515, -0.0659897253, -0.0267595444, 0.0193060972, 0.0190346632, 0.011877941, -0.0363517925, -0.0335589685, 0.0348317139, -0.0267020389, -0.0098522883, 0.0691522956, -0.005452401, 0.0284789987, -0.0828629211, -0.0136917429, -0.0033667409, 0.0292183664, -0.0313057713, -0.0061342595, 0.0139998971, 0.0203887206, 0.0350817852, 0.0185604673, -0.0253333747, -0.0250313804, -0.1138856113, -0.0002747276, 0.0223461855, 0.0242977273, 0.0489088185, -0.0430713072, 0.0094349924, 0.0060460484, -0.0411826819, -0.0151294991, -0.0161660891, 0.0048297658, 0.0335620716, 0.0364769623, 0.0441634096, -0.062564902, -0.0238918699, -0.0278718732, 0.0362419412, 0.0172193907, -0.0014563923, -0.0371680818, 0.0409112163, -0.0287554841, 0.0120242862, 0.0156404991, -0.0272691846, -0.0230382103, 0.0056689866, -0.0147083811, 0.0117952535, 0.0158296246, 0.0145060997, 0.0415231921, 0.0160756297, -0.0299120899, 0.0218127556, -0.0600366145, -0.0178443529, -0.0130506782, -0.0406082496, -0.0115348492, 0.0904004946, -0.0073065376, -0.0052618189, -0.0512643717, 0.0368412808, -0.0348918214, -0.0180323552, 0.0057237586, 0.0223933198, 0.0197548009, 0.0073402068, -0.0068756687, 0.0115858121, 0.0696220398, 0.0292071346, 0.0346962102, -0.0644206107, 0.018317841, -0.0227369368, 0.0839392394, -0.0408132747, 0.016079532, -0.0026120935, 0.0516868345, 0.0079582464, -0.0460135788, 0.001292065, -0.015925521, 0.0013811913, -0.0021388466, 0.0276025105, 0.0133848442, -0.0695085377, 0.0107674915, -0.0300309043, -0.0569472574, -0.0300562643, -0.0433668345, -0.0001428816, 0.031181559, -0.0042778659, -0.0491957664, 0.0276949164, -0.0275690351, -0.0310307853, 0.0339417495, 0.0516763404, 0.0499630496, -0.0723486617, 0.0091027403, -0.0270607546, 0.0045491238, 0.05122637, -0.0144621199, 0.0067942166, -0.0026576156, -0.0335898213, 0.0287057664, 0.0015029204, -0.010047893, -0.0441732705, 0.0686792657, -0.0422184914, -0.0285716057, 0.1158523709, 0.0457840078, -0.0620263107, -0.0147794187, 0.0214142576, 0.0011064801, 0.0426735096, 0.029105546, -0.0865515769, -0.030559713, -0.035681244, 0.0469736643, -0.0066949609, -0.0042350744, 0.0075400756, -0.1030430049, 0.0256777126, -0.0291042347, -0.052469071, -0.0638263822, -0.041691523, 0.008105102, -0.0669238344, -0.0359859578, -0.0135095483, 0.0069037247, 0.0377499014, -0.0042101499, -0.0101290829, -0.0167838503, 0.0714009777, -0.0173891019, 0.0545777678, 0.0411274433, 0.0446625575, -0.0065019466, -0.0478139706, 0.0263233446, -0.0695869178, -0.0062205591, -0.0291514713, 0.0942812711, -0.0401410386, -0.0204420965, -0.0108275507, 0.0281669069, -0.0434155464, 0.0064704553, 0.0092314454, 0.0072506899, 0.0209541768, -0.0319956169, -0.0000100136, 0.0307705104, -0.0020576753, 0.0654572099, -0.0157373194, -0.0745690018, -0.0165867973, -0.0058021676, 0.0192910321, -0.0691798627, -0.0759335309, -0.0247448739, 0.0120407743, -0.0776337683, -0.0829526857, -0.0492354594, -0.0066797007, 0.0356417671, -0.0167363603, 0.0127701228, -0.0098397052, -0.0025681695, -0.0433806628, 0.0144900205, -0.0212788451, -0.0166292097, -0.0049540228, -0.031863559, 0.1068058982, -0.0178554468, -0.0178240556, -0.0012679427, -0.0510276332, -0.024182966, -0.005901352, 0.0721665174, 0.003730569, 0.0279489215, 0.0035245761, 0.0102911582, -0.0378894992, 0.0735487044, -0.01749097, -0.0066551417, 0.0526315793, 0.0256601274, 0.0242985487, 0.0354929194, -0.0241873637, 0.0641696304, 0.0304291341, -0.0051693623, -0.0431886427, -0.0238609202, 0.0380347371, -0.0338712148, 0.0629597008, 0.0027528242, 0.0024679455, -0.0452448018, -0.0077980473, -0.0203561112, -0.0274548009, -0.0312676616, 0.0136799244, 0.0058145691, 0.009063215, -0.0493041091, 0.0407392383, -0.0316469483, 0.0636477023, -0.0542318784, 0.0258017592, 0.0132606057, 0.0062005352, -0.0549830459, -0.004444737, -0.0052499077, -0.044707574, 0.0356592648, 0.026300827, 0.0034076741, 0.0375586189, 0.0389980748, 0.0334679633, -0.0001352169, 0.02615151, -0.0338168927, 0.0479466245, -0.0599843189, -0.012121547, 0.0108117992, -0.0234010611, 0.0276353639, -0.0328335278, -0.0096605588, -0.0067667416, -0.0006433271, -0.0362575427, 0.0079517374, 0.0266793165, 0.0117017664, -0.008713793, 0.002965071, 0.0690578818, -0.0526600555, 0.0373041779, 0.0412649103, -0.0081266072, -0.0126323551, 0.0050938493, -0.023767164, -0.0207222942, 0.0197802074, 0.0017393164, -0.052174136, -0.0441982299, 0.0328378603, 0.011459955, 0.0134457592, 0.037731763, 0.0277720019, 0.035616748, 0.0235749502, 0.041093342, -0.0654729679, -0.0317849927, 0.0298497286, -0.0268302765, 0.0002499812, -6.594017846e-33, -0.0332791582, -0.0212365966, -0.0164337792, 0.000277534, -0.0017412431, -0.0006792599, -0.0080971634, 0.0095598996, 0.0298664421, 0.0413148999, 0.0081985798, 0.0123852277, -0.0036117651, 0.010812765, 0.0518092848, -0.0068993797, 0.0548592135, -0.02624413, 0.0146714747, 0.0438466519, -0.0089692585, 0.0061340146, 0.0647484213, -0.1208339334, 0.0153138731, 0.002107705, -0.0427136309, 0.0125826867, 0.0013824842, 0.0410630852, -0.0120138824, 0.0384172685, 0.0340081714, -0.0422804691, 0.0056401649, -0.0190186128, -0.0451351069, -0.0166400541, -0.0187865682, -0.0174092166, -0.013932744, -0.0061868192, 0.0074826558, -0.0136670386, -0.0648724809, -0.0119533278, 0.0259423982, -0.0196350608, 0.0261266623, -0.112810187, -0.0101784868, 0.01703896, 0.0002284049, -0.0279458109, 0.0128735825, 0.0658759028, 0.0255155619, 0.0113106305, -0.0505003482, 0.0442554466, -0.061012771, 0.0606400408, -0.005256529, -0.000783799, 0.0194892045, 0.0693907365, 0.0184683092, 0.01887412, 0.0052403156, 0.0144274784, -0.0085454797, -0.0038495595, 0.0193072222, 0.0423794203, 0.0473996215, -0.0524369217, -0.0384364538, 0.0426064506, 0.0052173259, -0.0630603358, 0.0478635468, 0.0048634359, -0.0354594365, 0.0055071223, 0.0034982921, -0.0623812787, -0.0376775153, -0.0124211712, 0.0016351605, 0.0298754722, -0.0261685569, -0.0028731485, -0.0357977301, -0.0651186183, -0.0220397692, -0.0117637692, -0.0391593687, 0.0262813959, 0.0102711553, -0.0068148985, -0.0637215748, -0.0209004357, -0.0047337241, 0.0437232293, 0.0256061759, -0.0214766543, 0.1002990529, -0.0108553283, -0.106661655, -0.0041892403, -0.0148385549, 0.0437242575, -0.0072497427, 0.0217315052, 0.008762042, 0.0215859488, -0.0190987904, 0.0892428607, 0.0141646806, 0.0441772677, -0.0255969148, 0.0271957517, -0.004741285, 0.0264034681, 0.0340344943, -0.0086777611, 0.0127792051, -0.0039747804, -0.042704314, -0.0372603089, -0.0362763256, 0.0271688234, 2.728e-7, 0.0541928671, 0.0562123619, -0.0019009059, -0.0210205931, 0.0129417647, 0.0646097884, -0.0000107995, 0.0283404496, 0.0237247907, -0.0209513791, 0.0510800593, 0.0278780721, 0.0121797891, 0.0469248071, -0.0734753311, 0.068652302, -0.0321081765, 0.0048154406, -0.0123966001, -0.0235239211, 0.0666132793, 0.0908819884, 0.0260796696, 0.0046249344, 0.0024061657, 0.0381977074, 0.0362499356, -0.0027956641, -0.0299145393, 0.0039561093, -0.0211212803, 0.0047089783, -0.0207566433, -0.0078246426, -0.0311813913, 0.0113616567, 0.0045990217, 0.0378099307, 0.0041309074, -0.0756051019, 0.0392828211, 0.0523861125, 0.027071923, -0.004357195, 0.050992161, -0.0302545298, -0.0283144545, -0.0841678753, -0.0269563161, 0.0478014387, 0.043587476, -0.0001122754, -0.0293627847, 0.0580527633, -0.0094518401, -0.0306499917, -0.0392244533, -0.0283056255, -0.0153897768, 0.0673245639, 0.0165621042, -0.0275807567, 0.0115400841, 0.0203385167, -0.0464690402, -0.0299071353, 0.0192401372, 2.098169484e-34, 0.0569510832, 0.0029475153, 0.0052931877, -0.0332600623, 0.0155766951, -0.0084165325, -0.0630563945, -0.0160094909, 0.0140610626, 0.0295967348, 0.0157380439 ]
http://arxiv.org/abs/2406.18069v2
http://arxiv.org/pdf/2406.18069v2
Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals
2024-06-26T04:54:45
2024-06-27T03:58:25
Zengding Liu, Chen Chen, Jiannong Cao, Minglei Pan, Jikui Liu, Nan Li, Fen Miao, Ye Li
eess.SP
eess.SP, cs.AI, cs.CL
Large language models (LLMs) have captured significant interest from both academia and industry due to their impressive performance across various textual tasks. However, the potential of LLMs to analyze physiological time-series data remains an emerging research field. Particularly, there is a notable gap in the utilization of LLMs for analyzing wearable biosignals to achieve cuffless blood pressure (BP) measurement, which is critical for the management of cardiovascular diseases. This paper presents the first work to explore the capacity of LLMs to perform cuffless BP estimation based on wearable biosignals. We extracted physiological features from electrocardiogram (ECG) and photoplethysmogram (PPG) signals and designed context-enhanced prompts by combining these features with BP domain knowledge and user information. Subsequently, we adapted LLMs to BP estimation tasks through fine-tuning. To evaluate the proposed approach, we conducted assessments of ten advanced LLMs using a comprehensive public dataset of wearable biosignals from 1,272 participants. The experimental results demonstrate that the optimally fine-tuned LLM significantly surpasses conventional task-specific baselines, achieving an estimation error of 0.00 $\pm$ 9.25 mmHg for systolic BP and 1.29 $\pm$ 6.37 mmHg for diastolic BP. Notably, the ablation studies highlight the benefits of our context enhancement strategy, leading to an 8.9% reduction in mean absolute error for systolic BP estimation. This paper pioneers the exploration of LLMs for cuffless BP measurement, providing a potential solution to enhance the accuracy of cuffless BP measurement.
Large Language Models for Cuffless Blood Pressure Measurement From Wearable Biosignals: Large language models (LLMs) have captured significant interest from both academia and industry due to their impressive performance across various textual tasks. However, the potential of LLMs to analyze physiological time-series data remains an emerging research field. Particularly, there is a notable gap in the utilization of LLMs for analyzing wearable biosignals to achieve cuffless blood pressure (BP) measurement, which is critical for the management of cardiovascular diseases. This paper presents the first work to explore the capacity of LLMs to perform cuffless BP estimation based on wearable biosignals. We extracted physiological features from electrocardiogram (ECG) and photoplethysmogram (PPG) signals and designed context-enhanced prompts by combining these features with BP domain knowledge and user information. Subsequently, we adapted LLMs to BP estimation tasks through fine-tuning. To evaluate the proposed approach, we conducted assessments of ten advanced LLMs using a comprehensive public dataset of wearable biosignals from 1,272 participants. The experimental results demonstrate that the optimally fine-tuned LLM significantly surpasses conventional task-specific baselines, achieving an estimation error of 0.00 $\pm$ 9.25 mmHg for systolic BP and 1.29 $\pm$ 6.37 mmHg for diastolic BP. Notably, the ablation studies highlight the benefits of our context enhancement strategy, leading to an 8.9% reduction in mean absolute error for systolic BP estimation. This paper pioneers the exploration of LLMs for cuffless BP measurement, providing a potential solution to enhance the accuracy of cuffless BP measurement.
[ -0.0173057653, 0.0278067756, 0.0101025198, -0.0018825373, -0.0260041375, 0.0000627678, 0.0539831407, 0.0162267387, 0.0177215301, 0.0160769317, 0.0041099424, -0.0347788855, 0.047895994, 0.0474164784, -0.0098463614, -0.0460868813, -0.0131180696, -0.0059706736, 0.0888855308, 0.0122124115, 0.0323126055, -0.0218210481, -0.0316994116, 0.0451612286, 0.0125694769, -0.0136921341, -0.0062760888, -0.0265506692, 0.0015232938, -0.0815142244, -0.0401954502, 0.0472004935, -0.0301773753, 0.0018188067, 0.0000021141, -0.0390884019, 0.012304225, 0.0547824092, -0.0271714311, -0.0618515015, 0.0651283562, -0.0253428929, 0.0204171035, 0.0052712909, -0.0397703461, -0.0497206114, 0.0548223294, 0.0555569567, 0.0203446429, 0.0142401792, -0.0272569358, 0.0711912662, -0.0365852155, -0.0604253672, 0.0434567221, -0.0605624169, 0.0334353186, -0.0554103069, 0.0843917057, 0.0295609981, -0.0189089142, 0.0529132187, -0.0054718354, 0.0605217405, 0.0312668085, 0.0158047024, -0.0189264752, -0.0228690282, -0.0303223096, 0.0521109998, 0.0067158425, 0.0230996311, 0.0362385623, 0.0084821628, -0.0011066053, -0.0163050033, -0.0740846172, -0.0409244001, -0.0255478639, 0.0084294491, -0.0443021916, 0.0103005553, 0.0230914261, -0.0375718959, -0.0375940464, 0.0382445008, 0.0078749098, -0.0129767591, 0.0153566562, 0.0180931687, -0.0348822437, -0.0576621778, 0.0419350043, 0.019614609, -0.0331697427, -0.0180721767, -0.0365557782, -0.0029521321, -0.0021394691, -0.0428192653, 0.0160803273, 0.0461491011, 0.0235692561, 0.0370970629, 0.0065845069, 0.0177761745, -0.0283690672, -0.0314156078, -0.0430182666, 0.0183225144, -0.0310540106, -0.0036896188, 0.0102256089, 0.0131731872, -0.0315491855, -0.0137250638, -0.0902895555, -0.032216955, -0.014955366, 0.0066604665, -0.0320931002, 0.0134013528, 0.0203173421, -0.0088123018, -0.0064248098, 0.0268825144, 0.0356421284, 0.0162073132, -0.0158886258, -0.0500417911, 0.0271006972, -0.0189621281, 0.0397616215, 0.0102420105, -0.0031408847, 0.0402937867, -0.0020213514, 0.0245020315, -0.0271514915, -0.0501912832, -0.0200611521, 0.0266802758, -0.0147378687, -0.0522070453, 0.0392488204, 0.017922055, 0.0210453961, -0.0295597631, 0.0163419079, 0.0317382701, -0.0280739237, 0.0616874211, -0.0490184166, 0.0124614462, -0.0102224778, 0.0086442968, 0.0178354047, 0.0497665964, 0.0386555158, -0.0681994408, 0.0272680596, -0.0299870446, 0.0006585316, 0.0554592721, 0.0095102945, -0.0706239194, -0.050091397, -0.0163982566, 0.0448117182, 0.0353304781, -0.0068843025, 0.0493330397, 0.0418495238, 0.0326640494, 0.0776662529, 0.0584226772, 0.0480747595, 0.0226426851, 0.016472131, -0.0394567624, 0.0049611791, -0.0309561007, -0.0652100667, -0.0217372309, -0.0319871269, 0.0209042784, -0.0396840274, -0.026394764, -0.0288213994, -0.0566797741, -0.0554745756, -0.0363186523, -0.0113016386, 0.0073372638, 0.0808561742, -0.0512578934, -0.0019361334, 0.0295189358, 0.0230895076, -0.0401762687, -0.0042919889, 0.0152134029, 0.0029364948, 0.0774594098, 0.0162472464, -0.021041153, 0.0660103336, -0.0216138158, 0.0415743403, 0.0173582733, 0.0231034737, -0.0006054355, 0.0256258808, -0.0537689477, -0.0340796076, 0.0595481507, -0.0220635887, -0.001562268, -0.0607123189, 0.0504047051, 0.0173924249, -0.001243889, -0.0323934257, 0.024055142, -0.0775434375, -0.0413869433, 0.0271347836, 0.0193843916, 0.0182155501, 0.0248031933, 0.012372301, 0.0256638434, 0.0164282005, -0.0651352108, -0.023805093, 0.0032077834, -0.0370579027, 0.0134560205, 0.0436875671, -0.0270443112, 0.0409862138, -0.050530985, 0.0011430548, -0.0410253741, -0.0488617979, -0.029158311, 0.0459729284, -0.0044212211, -0.0191345979, 0.0199857131, 0.002905075, -0.0162419695, -0.0280140135, 0.0417447612, 0.0197790079, 0.0165693015, -0.0127471313, 0.007749462, -0.0259126425, -0.0350992233, -0.0188774653, -0.0183228087, 0.0070531759, 0.0258624703, 0.0389695019, -0.0106820725, -0.0005291713, -0.1385159492, -0.009402642, 0.0268030744, 0.0146459434, 0.0171286333, 0.0783084631, -0.0182905793, -0.0117792645, -0.0308416691, -0.0375247337, -0.0016869566, 0.0240295269, -0.0242501404, 0.0160783138, 0.0082607372, 0.0210801233, -0.0103301965, -0.0527630411, 0.0518046543, 0.0024759059, -0.003428424, 0.0323288776, 0.0020058593, -0.0325159691, 0.0227950048, 0.0404689722, 0.0045330357, 0.0193689615, -0.0166849699, 0.0385157056, 0.0358670205, -0.0155439675, -0.0239879973, 0.0328499004, -0.0197530556, -0.0095213819, -0.0053009903, -0.0457798764, 0.0188664459, 0.0175980851, -0.0295314174, -0.0335536823, 0.0650045052, 0.0392942205, 0.0148322089, -0.0127037931, 0.0159681812, -0.0463252217, -0.0058095199, -0.1048693284, 0.0303607453, -0.0309415776, 0.0405401886, 0.005883038, -0.0335800275, 0.0404391699, -0.0215494223, 0.0072982684, -0.0010854868, -0.032491032, -0.0209787264, -0.013788905, -0.0301484764, 0.0056418777, 0.0117942225, -0.0042584948, 0.0273788869, -0.0014109819, -0.0233034752, -0.0637856871, 0.0156652275, -0.0521756038, 0.0106313126, 0.0138075659, -0.0408994369, -0.0502813049, -0.0043928097, -0.069597587, 0.0046275104, -0.0290797576, 0.0131763667, 0.0101330644, 0.0377697721, 0.0088955453, 0.0202645306, -0.0014461493, -0.0095038172, -0.0083111245, 0.0095787439, 0.0356337391, -0.0263821874, -0.0006395709, -0.0075383941, 0.0590672232, 0.0422851369, 0.0593782179, -0.0012108287, -0.0294780619, 0.0121077579, -0.0092604291, -0.0012138464, -0.0243048221, 0.0335581005, 0.0161916111, -0.0142461266, 0.0139101176, 0.1510569602, 0.026906481, -0.0590974838, -0.0025730438, -0.0198634509, -0.0263238214, 0.001707449, -0.0483182408, -0.0464836843, -0.0318883508, 0.0141231734, 0.0889017805, 0.0058016875, -0.020590296, -0.0350683667, -0.0391650163, 0.0088029047, 0.0427169502, 0.0017183766, -0.0020561784, -0.0672787726, 0.0666219145, -0.043568857, 0.0217811596, -0.0114985071, -0.0321708806, 0.0098065576, 0.0246045366, -0.1024015024, -0.0658092648, 0.0130864317, 0.0039759954, 0.0151313292, 0.0611108243, 0.0102589075, -0.0128209507, 0.0061403788, 0.0226942617, -0.0256031528, -0.0141484896, -0.0563300513, 0.0069155935, 0.074044697, -0.007062946, -0.0278264359, 0.0998318568, -0.0469388068, 0.0118070357, 0.0346028246, 0.0575090721, -0.0025528176, -0.0157617424, -0.0435881689, 0.0072063613, 0.0421673991, 0.0570903569, 0.0076290574, -0.0707682222, -0.0112006059, -0.0135716368, -0.0092530111, 0.0405080952, -0.0560623072, -0.026163049, -0.0279264692, -0.0474140868, -0.0175311416, -0.0155518893, 0.0371764973, 0.0685540438, 0.0464579165, 0.0604566075, 0.1131873429, 0.0193315893, -0.0247620549, 0.0352815092, -0.0242531765, 0.0401446708, -0.0313166194, 0.0532634482, 0.0095768925, 0.0119668515, -0.0255662072, 0.058686506, -0.0378289111, -0.0196478236, 0.0344220325, 0.0563199669, 0.0321770534, -0.0209380351, -0.055099342, -0.0245749131, 0.0199829321, 0.0492016226, -0.0404185504, 0.0409679525, -0.0386089943, 0.0348146968, -0.0004372565, 0.0071063838, 0.001750715, 0.028289929, -0.0185241867, 0.0377079286, -0.0100533403, -0.0585217588, 0.0260092765, 0.0443439148, 0.0480768979, 0.0183376987, 0.0014319402, 0.0060407817, -0.0177135281, -0.0649955794, -0.0056130919, -0.0395221338, 0.0439100191, -0.0364417359, 0.0213223323, 0.0099444222, -0.0444750302, -0.05181133, -0.0519687533, -0.0740346536, 0.0540996939, -0.0066032042, 0.0224261973, 0.0406789556, -0.0189605914, 0.0125376256, 0.013533894, 0.0139104705, -0.0051562851, 0.0087115653, 0.0727377683, -0.0237688944, -0.0022486076, -0.0322857015, 0.0177595094, -0.03249586, 0.0442245752, 0.0142691769, 0.0861818269, -0.0035626236, 0.0034176663, -0.0113430955, 0.0151592707, -0.0027804514, -0.0744198337, 0.0282889996, -0.0045691989, 0.0474155694, 0.0673976392, -0.0038364858, -0.0086571453, -0.0006556126, 0.0702353939, -0.0111874314, -0.0174911562, 0.0183328521, -0.0938715041, -0.0179159995, 0.0372881405, 0.0498534925, 0.0113660004, -0.0580226257, -0.0036113574, -0.072769329, -0.0486393273, 0.0736775398, 0.031139167, -0.0287285745, -0.0453687347, 0.010401424, 0.0371327661, -0.0711670071, 0.0049307193, 0.0095925368, 0.0589340553, -0.0570530109, -0.0315939561, -0.0354544967, -5.963967935e-33, 0.011468973, -0.0400455445, -0.005348267, -0.0250823237, -0.0026341865, 0.0593895838, -0.0400548242, 0.0423107184, 0.0383125767, 0.0365659073, -0.0172914695, -0.0505026281, 0.0371840596, -0.0080511719, 0.0183138922, -0.0510893278, 0.0352227762, 0.0142560489, 0.0110148378, 0.0197143238, 0.0322666503, 0.0348792449, -0.0183747746, -0.10104727, 0.0107925963, 0.0159517862, -0.0324594192, -0.0078692837, 0.0229927078, 0.0363852195, -0.0317194089, 0.0732770339, 0.0130627099, 0.0142614814, -0.0161612444, 0.0659156367, -0.0648334175, -0.0164687019, 0.0275269095, -0.0163154379, 0.0342766568, -0.0197011176, 0.028766619, -0.0117814681, -0.0585114248, -0.0103147812, 0.0224146098, -0.0779429227, -0.0217650533, -0.0256147273, 0.0353488848, 0.0027546191, 0.0027073715, -0.0038430202, 0.0294583123, 0.0161763299, -0.0537859537, -0.0296681989, -0.0655071139, 0.0075446549, 0.0643187836, -0.0014853715, 0.0127880089, 0.0270909648, 0.0014253814, 0.0030115347, 0.0098999087, -0.0578640103, 0.001813427, 0.0456488989, -0.0129893925, 0.0461588092, 0.0598059855, 0.0280103777, 0.0123384334, -0.0232001878, -0.0699160323, 0.0244866721, -0.0159031749, -0.041169513, -0.0131920027, 0.0025927206, -0.0464042909, -0.0214776658, -0.0032557782, -0.0032200601, -0.0425360836, -0.025039807, 0.0112982923, -0.0090422826, 0.0079167886, 0.0131869381, 0.0113811409, -0.0329508856, 0.0085652061, -0.0311106108, 0.0454660356, -0.0209986139, 0.0183310956, -0.0450188778, -0.0326140411, 0.021705687, 0.0271914732, 0.0490107499, 0.0341334976, -0.0265174899, -0.0178216435, 0.0176570304, -0.0388804078, -0.0315367766, 0.0109699033, 0.0094416784, -0.0196691155, -0.0244259257, 0.0677987039, -0.0337337218, -0.0009879953, 0.0836257935, -0.0279290248, -0.0127173439, -0.0052198619, 0.0327930041, 0.016722355, 0.0221768487, -0.0414147712, -0.0307509229, -0.0174985602, 0.0185830947, 0.0253423043, 0.0064782295, -0.0249670502, 0.0038264839, 2.881e-7, 0.0121904286, 0.0096639944, 0.0043290728, -0.0356729887, 0.0194921959, 0.048264429, -0.0376808718, -0.012317026, 0.0141706727, -0.0368341655, 0.0219701789, -0.0094549684, 0.0673421919, 0.0183375571, -0.0204253867, -0.0672367662, -0.0054810308, -0.0033610447, -0.0311477613, 0.0009177297, 0.0659635812, 0.0244868845, -0.0312347226, -0.0153741241, 0.0310790967, -0.0006551773, 0.0312539302, -0.0405957922, -0.0268808138, -0.0191222541, 0.0026134241, -0.0042325864, 0.0346760675, -0.0005417228, 0.0035355408, -0.0167937018, 0.0375806652, 0.0857304037, -0.0415812656, 0.0503547713, 0.0158828478, 0.0282067116, 0.0194343626, -0.0302469581, 0.0820963159, -0.0465124287, -0.0369208306, -0.0615930073, -0.0190175157, -0.0019293618, 0.0096350927, -0.0032816979, -0.0252425838, -0.0142543567, 0.0141823459, 0.0140109444, -0.030232653, 0.0247811917, -0.0201659054, 0.0409721509, -0.0103307245, -0.0359166376, -0.009098256, 0.0622215793, -0.006577333, -0.0860429853, 0.0061742645, 2.461122191e-34, 0.0093995649, -0.0108976001, -0.0126961023, -0.0234171785, 0.0145884026, -0.0291305408, -0.0249172021, 0.02598105, -0.0001382885, -0.0928258747, -0.0085662426 ]
http://arxiv.org/abs/2406.18067v1
http://arxiv.org/pdf/2406.18067v1
Exploring Energy-Based Models for Out-of-Distribution Detection in Dialect Identification
2024-06-26T04:52:48
2024-06-26T04:52:48
Yaqian Hao, Chenguang Hu, Yingying Gao, Shilei Zhang, Junlan Feng
cs.CL
cs.CL, eess.AS
The diverse nature of dialects presents challenges for models trained on specific linguistic patterns, rendering them susceptible to errors when confronted with unseen or out-of-distribution (OOD) data. This study introduces a novel margin-enhanced joint energy model (MEJEM) tailored specifically for OOD detection in dialects. By integrating a generative model and the energy margin loss, our approach aims to enhance the robustness of dialect identification systems. Furthermore, we explore two OOD scores for OOD dialect detection, and our findings conclusively demonstrate that the energy score outperforms the softmax score. Leveraging Sharpness-Aware Minimization to optimize the training process of the joint model, we enhance model generalization by minimizing both loss and sharpness. Experiments conducted on dialect identification tasks validate the efficacy of Energy-Based Models and provide valuable insights into their performance.
Exploring Energy-Based Models for Out-of-Distribution Detection in Dialect Identification: The diverse nature of dialects presents challenges for models trained on specific linguistic patterns, rendering them susceptible to errors when confronted with unseen or out-of-distribution (OOD) data. This study introduces a novel margin-enhanced joint energy model (MEJEM) tailored specifically for OOD detection in dialects. By integrating a generative model and the energy margin loss, our approach aims to enhance the robustness of dialect identification systems. Furthermore, we explore two OOD scores for OOD dialect detection, and our findings conclusively demonstrate that the energy score outperforms the softmax score. Leveraging Sharpness-Aware Minimization to optimize the training process of the joint model, we enhance model generalization by minimizing both loss and sharpness. Experiments conducted on dialect identification tasks validate the efficacy of Energy-Based Models and provide valuable insights into their performance.
[ 0.017423179, 0.0453409627, 0.0257431194, 0.0620901808, 0.0234664734, 0.0253013093, -0.0396624841, -0.0113015668, -0.0523551628, -0.0184500311, -0.0365652591, -0.0376596786, 0.0193576459, -0.0189093687, 0.0256831497, -0.0151831508, 0.0431858413, -0.010739712, 0.0203573927, -0.0326295272, 0.0112207793, -0.0098563051, -0.0070636198, 0.0113802124, -0.0889427736, 0.0428325757, 0.003502636, 0.0924868435, 0.0532855466, 0.0095447013, 0.0465259738, 0.0262941383, -0.0367153287, 0.0312634408, 0.0000019915, -0.0433154069, -0.0345428512, -0.0246695746, -0.0192187708, 0.0021111765, 0.0313398726, 0.0001389602, -0.0584399141, -0.0130641134, 0.0240770858, -0.012396019, 0.1181958243, -0.0109727299, 0.0125088478, 0.0777270198, 0.0098457756, -0.0358091593, 0.0229447149, -0.0118198218, -0.0315252319, 0.0100404946, 0.013023708, -0.0610730723, -0.1054164618, -0.0815690234, -0.0031922364, 0.0294117481, -0.0211049542, 0.0393879563, -0.0048350575, 0.0043419623, -0.0760893598, 0.0165559221, -0.0088810902, 0.0502932146, -0.0638609305, -0.0174743142, -0.0048515452, -0.0115451785, -0.0513382442, 0.0256263018, 0.0143250208, 0.0049394178, 0.0021742256, 0.0554511733, -0.0319357663, 0.0101661514, 0.0401599221, 0.0046050884, -0.0112311663, 0.0492904522, 0.0368515849, -0.0171184279, 0.0243396014, 0.0265268404, -0.0997306928, 0.015532176, 0.0139687713, 0.0224654488, 0.0256163981, 0.0014929015, -0.0068297833, -0.0288563706, 0.0080769258, -0.0462197922, 0.0424318016, 0.007920715, -0.0438144542, 0.0005917198, -0.0083707348, -0.0054910122, -0.0114454236, 0.0163463634, -0.0216020066, -0.0325663984, -0.0392887779, 0.020674441, -0.0278665628, -0.0000397072, 0.0262421835, -0.0041534593, -0.0151993548, -0.033873383, -0.0416798629, 0.0033095009, 0.0393498763, 0.0096991314, -0.0443269536, -0.0447760783, 0.0477974899, 0.0483302474, 0.0051616267, 0.005677165, -0.0390587635, -0.0746776238, -0.0237476602, 0.0337304249, 0.019734161, -0.0012626672, 0.0166398715, -0.0025483801, 0.0181788728, -0.0296672489, -0.1033855975, -0.0221195016, -0.070551075, -0.0484548733, -0.0130098676, -0.0803494006, -0.0263668895, -0.0298838317, 0.0427879132, -0.0507452562, 0.0174497776, -0.0351723768, 0.0255709924, 0.0202505775, -0.0321210586, 0.0198867433, 0.0515920259, -0.0197268259, 0.0027256226, 0.1124995351, 0.0327746123, -0.0303287655, 0.0120386789, 0.0276745893, -0.02123948, 0.0677172467, 0.0032096361, 0.0413405448, -0.1022602171, -0.0593205653, 0.0228902176, -0.0021687928, 0.0700014904, 0.0522898026, 0.0035568238, -0.0089961151, 0.0740249902, 0.0044887136, 0.0319612361, -0.0144370478, 0.020755792, 0.0101606473, -0.0231118388, 0.0149450274, 0.0352799445, -0.046300292, -0.032106407, -0.0082469089, 0.0169535577, 0.0584616363, -0.0519345291, -0.0014857945, -0.0342018083, -0.0828248113, -0.0263296515, 0.0072403778, 0.0840456188, -0.0117676267, 0.0015167653, 0.06647221, 0.0497293733, -0.0032618102, 0.0000054127, -0.0352309309, -0.0139832413, 0.003308954, 0.0030765785, -0.0068680607, 0.0386558957, 0.0143906614, -0.0262155086, -0.0057420344, 0.0455618203, -0.0050428296, 0.0576823503, -0.0551329851, -0.0094435802, 0.0163461324, 0.0447900891, 0.0011304256, -0.0452709459, 0.0329978168, 0.0294025447, -0.0625815317, 0.0085914955, -0.0015027146, 0.0444959849, -0.013924283, 0.0303504132, 0.0327859037, 0.0131722912, 0.0357412212, -0.0194269102, -0.0069635315, 0.0029763246, -0.0579945706, 0.0425442718, 0.0025055485, -0.0377845988, -0.0469654873, 0.0584204793, -0.0077278963, 0.0019526969, 0.0175182819, -0.0012019095, -0.0581450649, 0.0353635624, -0.06668365, 0.058256112, -0.0827701464, 0.0157914478, -0.0073814015, 0.0118935686, -0.0042623798, 0.1046995595, -0.0022357299, 0.0298116915, 0.0132324258, -0.0350388177, 0.0490997061, -0.0924439728, -0.0096069369, 0.0338674188, -0.0084505826, -0.0054411762, -0.0165722128, -0.027524557, -0.0167388003, -0.0289754719, -0.0658257976, -0.0168249365, -0.0531264655, 0.0221698638, 0.0241093561, -0.0092813885, -0.0316779055, -0.0089893695, 0.0053427499, 0.000485218, -0.0038679044, -0.0011612729, -0.035099335, 0.0251962151, -0.0193902664, -0.0114387944, -0.0178535488, 0.0081445081, 0.0430891439, -0.0559662208, -0.031594649, -0.0328433365, -0.0137581835, -0.0014649973, 0.0375347286, 0.0326684825, -0.0570060723, -0.0488603935, -0.0268340092, 0.0812969729, 0.0524854511, -0.0104808221, -0.0381948501, 0.054228723, -0.0635444224, -0.0122091258, 0.0047663464, -0.0372589529, -0.0090904171, -0.0088705476, -0.0455569439, -0.0274217632, 0.0606728792, 0.0294605214, -0.000863593, -0.0161996651, 0.0007463399, -0.0080253603, 0.0112300012, 0.0192396678, 0.0202836227, -0.0335096568, 0.0169725399, -0.0439831465, -0.0184135865, 0.0649279356, -0.0795688033, -0.01776563, -0.0371838398, -0.0281153545, 0.0062029082, -0.0302471984, -0.0103908191, 0.037701197, 0.0023067843, -0.0460957363, 0.005554067, -0.0241162162, 0.0229815487, -0.0145736998, 0.0400406718, -0.0363802686, -0.0051096133, 0.0069242171, 0.0263201967, -0.0018649711, -0.0117080957, -0.0454440601, 0.0332371518, -0.0065959431, -0.0126061263, 0.04632603, 0.0282253455, -0.0106269531, 0.0515574403, -0.0264219381, 0.0279865973, -0.0041521476, 0.0669629797, 0.039298404, 0.0202297699, -0.0153509863, -0.0373951495, 0.0076545193, -0.02190727, -0.0190194976, 0.0381896086, -0.0115639446, 0.0146582089, 0.031205168, 0.0144095402, -0.0200686362, 0.0354280584, 0.0641638413, -0.024252601, -0.0086555574, 0.0814561024, 0.0084823165, -0.0225119404, 0.0030330154, -0.005509031, 0.053489238, 0.0013765506, -0.0311952047, -0.0054330737, -0.045363266, 0.0362191871, 0.0549892448, -0.0019364507, 0.010460984, 0.0052023763, -0.0802890211, 0.0077482262, 0.050682921, 0.0074406164, -0.0166492146, 0.004348536, 0.0251785405, -0.026403632, -0.0246031694, -0.02394199, 0.014066631, -0.0384884588, 0.0727959126, 0.0014947827, -0.0911171883, -0.0280694198, -0.0713164732, 0.0289476365, 0.0302408356, -0.0326607563, -0.0278007668, -0.0872714669, 0.0326793976, -0.1013592556, -0.0236603245, -0.0240079705, -0.0008821813, 0.0406728834, -0.0272135716, 0.0147131924, -0.0068399454, -0.0207037982, 0.0168426912, -0.0321779884, 0.0563796237, -0.0283093061, 0.0076690726, 0.0415362418, -0.0066920081, 0.0608241297, 0.0301052537, 0.0381700248, -0.0036064943, -0.0228035338, -0.0133110723, 0.0276830681, -0.0079198871, 0.0163690001, 0.0596628413, 0.0369326137, -0.1019840911, -0.0295402817, -0.0308051705, -0.0368963927, 0.0268162675, 0.0280030221, -0.0218847394, -0.0211178008, 0.0153158512, -0.0188562963, 0.0456558801, -0.0236298367, 0.014424691, -0.0653638169, -0.0453176796, 0.0468720309, -0.0178047344, -0.0018241015, 0.006457401, -0.0614181608, -0.075033918, 0.014709793, 0.0189935472, 0.0220472254, 0.0303137191, -0.0404120609, -0.0322046205, -0.0192276258, 0.0192878395, 0.0256742407, 0.0194788612, 0.0504337028, -0.0327492915, -0.01836426, 0.0403643064, -0.026677439, 0.035068728, 0.0123208221, 0.0229793862, -0.0598074608, -0.0716220066, 0.0213422198, -0.0516526327, 0.047104653, -0.0074268649, -0.0482340939, -0.0579815619, 0.0117279543, 0.0481831133, 0.0378358923, -0.0302997641, -0.0278683808, 0.0577069223, 0.0179034099, -0.0011155918, 0.0725535303, -0.0149500631, 0.0252893623, -0.0190328546, 0.0349504463, 0.0530279614, 0.032783553, 0.0156099629, 0.0261577442, -0.0601954237, 0.0526871085, 0.0214706361, 0.0058377907, 0.0117090084, 0.0219150763, 0.0012518809, -0.027840754, -0.0145841986, -0.0173494425, 0.0382881202, 0.0076213148, -0.0610174648, -0.0125071099, 0.0504230298, -0.0253552012, 0.0254298747, -0.0091598853, -0.0081298556, -0.058142744, -0.0036919003, -0.0004554145, -0.0215739887, 0.0434862822, 0.0018494645, -0.001371634, 0.0198175404, 0.0659798533, 0.0124231726, 0.0509003028, -0.0342464522, 0.0003434917, -0.0031198936, 0.0129281078, 0.0332110487, -0.0008199772, -0.0374063104, 0.0351511277, -0.0031788223, -0.0492196679, 0.0078095798, -0.0279674456, 0.0190998726, 0.0284787063, 0.0474916212, 0.0201825853, -0.0581848137, -0.0296928566, 0.004296605, -0.0332565159, 0.0098994793, 0.0189011302, 0.0050229565, -6.692084931e-33, -0.0023682609, 0.0066174041, -0.0488678999, -0.0264057945, -0.0140349101, 0.0073459866, -0.0302478969, 0.0331584178, -0.0359808728, -0.0078766141, 0.0317258015, -0.0130042713, 0.0319922566, 0.0281198211, -0.0319500715, -0.0042593363, 0.0612354837, 0.0044406601, 0.0099669825, -0.0023930105, 0.0310536679, 0.031063173, 0.0927255377, -0.0361634344, -0.0222996194, -0.013145024, -0.0467722565, 0.0401571989, 0.0348201469, 0.0414376296, -0.0363917574, 0.0756834522, 0.0018340971, -0.0250814334, -0.0091704195, 0.0042620208, -0.039374467, -0.0141994758, -0.0485013053, 0.0448271222, 0.0413913913, -0.0496111885, 0.0381908417, 0.0023967999, -0.0184949357, 0.0076292595, 0.0382211693, -0.055755645, -0.006102928, -0.0248347968, -0.0215044413, 0.0569796301, -0.0001628326, -0.0472293645, 0.049701076, -0.0070166877, 0.0053878231, -0.0254160929, -0.0337464362, -0.0027809611, 0.0708206967, 0.0497663133, 0.0480771363, 0.0290647037, -0.0425059237, 0.022284437, 0.0035904017, -0.0435254574, 0.0345095061, -0.0424160808, 0.0657430515, 0.0258818381, 0.0099934852, 0.0052226656, 0.0406608805, -0.02714785, -0.0672096536, 0.0012536276, 0.025499342, -0.0087664174, -0.0190309137, 0.0021494313, -0.0026014894, -0.0362062566, -0.0016835427, -0.0557498001, -0.0205895882, -0.0053567328, 0.0104202889, 0.0055810288, 0.0053791348, 0.0117170811, 0.0371331647, 0.0023698232, 0.0384763069, 0.0253424067, 0.0181208905, 0.0172832943, 0.0025783309, 0.0051475032, -0.04518288, -0.0292884167, 0.0194576439, 0.038695801, -0.0016911938, -0.0129037946, 0.0489945337, 0.0456389189, -0.0284673404, 0.0316909552, -0.0423429757, 0.0141469957, -0.0290705562, 0.0837465152, 0.0401004069, -0.0272876732, 0.0127066802, 0.0806377307, 0.002793327, 0.0489725433, -0.0031089715, 0.0128645534, 0.0154842045, 0.0695109591, 0.0128003564, 0.0214753002, 0.0370010696, -0.0328971781, -0.0152835567, 0.0072724074, 0.0037999251, 0.0278894026, 2.744e-7, 0.0273172222, -0.0231849179, 0.0158582181, -0.0458643138, 0.0288820621, 0.0093116062, -0.0307575949, -0.0523930639, -0.0100799808, -0.0643612444, 0.0370633043, 0.0017739797, 0.0802118629, 0.0549825989, -0.083027944, 0.0177972987, -0.0544452816, 0.0244477559, -0.0253092628, -0.0495412312, 0.012513225, 0.0664198846, -0.0212513264, -0.0279115345, 0.0228357632, -0.0470181927, 0.0411016308, -0.010690351, 0.0499895997, -0.0039774459, 0.0354072116, -0.0346508957, 0.0121253664, 0.0002250591, -0.0219252035, 0.0278936848, 0.0057023275, 0.0378628522, -0.0214782581, -0.0835155919, 0.0181048792, 0.0565452687, -0.0339063667, -0.0373787247, 0.053947892, 0.0021109919, 0.0011116617, 0.0230152365, -0.0346199311, -0.0046358518, 0.0140747391, -0.0339095779, -0.0022542838, -0.0039678691, 0.0192520227, -0.0237851795, -0.0369224511, -0.0329320505, -0.0105376495, 0.0954980627, -0.0604819804, -0.0330512896, 0.0313119367, -0.0185213611, -0.0100523904, -0.0180453341, 0.0050058765, 2.103786831e-34, 0.0167265758, -0.0118064769, -0.0060681659, -0.0037294445, -0.0060635079, 0.0171599463, -0.0657814518, -0.0171005297, -0.0219515394, -0.0249817222, -0.0148741528 ]
http://arxiv.org/abs/2406.18064v1
http://arxiv.org/pdf/2406.18064v1
Evaluating Quality of Answers for Retrieval-Augmented Generation: A Strong LLM Is All You Need
2024-06-26T04:49:41
2024-06-26T04:49:41
Yang Wang, Alberto Garcia Hernandez, Roman Kyslyi, Nicholas Kersting
cs.CL
cs.CL
We present a comprehensive evaluation of answer quality in Retrieval-Augmented Generation (RAG) applications using vRAG-Eval, a novel grading system that is designed to assess correctness, completeness, and honesty. We further map the grading of quality aspects aforementioned into a binary score, indicating an accept or reject decision, mirroring the intuitive "thumbs-up" or "thumbs-down" gesture commonly used in chat applications. This approach suits factual business settings where a clear decision opinion is essential. Our assessment applies vRAG-Eval to two Large Language Models (LLMs), evaluating the quality of answers generated by a vanilla RAG application. We compare these evaluations with human expert judgments and find a substantial alignment between GPT-4's assessments and those of human experts, reaching 83% agreement on accept or reject decisions. This study highlights the potential of LLMs as reliable evaluators in closed-domain, closed-ended settings, particularly when human evaluations require significant resources.
Evaluating Quality of Answers for Retrieval-Augmented Generation: A Strong LLM Is All You Need: We present a comprehensive evaluation of answer quality in Retrieval-Augmented Generation (RAG) applications using vRAG-Eval, a novel grading system that is designed to assess correctness, completeness, and honesty. We further map the grading of quality aspects aforementioned into a binary score, indicating an accept or reject decision, mirroring the intuitive "thumbs-up" or "thumbs-down" gesture commonly used in chat applications. This approach suits factual business settings where a clear decision opinion is essential. Our assessment applies vRAG-Eval to two Large Language Models (LLMs), evaluating the quality of answers generated by a vanilla RAG application. We compare these evaluations with human expert judgments and find a substantial alignment between GPT-4's assessments and those of human experts, reaching 83% agreement on accept or reject decisions. This study highlights the potential of LLMs as reliable evaluators in closed-domain, closed-ended settings, particularly when human evaluations require significant resources.
[ 0.080101423, 0.0225383155, -0.0230173152, 0.0039026241, -0.0393909924, -0.015013014, -0.0297701936, 0.0039200122, 0.0165132638, -0.0503195971, -0.0150626563, -0.0492960811, -0.0019326835, 0.0138948662, 0.0082671568, 0.012953531, 0.0367214419, 0.011159515, 0.0286546163, -0.0065691886, 0.0260156412, 0.0238408893, 0.0125357341, 0.0531695224, -0.0169347636, -0.0013946303, 0.0084880255, 0.0106422426, 0.0130558768, -0.0117810136, 0.0039242492, 0.0531162843, -0.0374013893, 0.0188815985, 0.000002047, -0.0616709813, -0.0000760876, 0.0071372306, -0.0235514119, 0.0361943319, 0.0265146829, 0.073669374, -0.0068831723, 0.048085466, -0.0569527, 0.0076433616, 0.0226835907, 0.0037123121, 0.1157982647, 0.0395275429, 0.013249049, 0.0032769751, 0.0255399998, -0.0124831945, 0.0522343554, -0.0377641916, -0.0198499002, 0.0614932105, 0.0280195903, 0.0115900002, 0.0422191471, 0.0506027229, 0.013322372, -0.0367714018, 0.0158923343, 0.0234840047, -0.0469474718, -0.0181899406, 0.0160109997, 0.0411445089, 0.0541822501, -0.0068519129, -0.02533466, -0.0076314136, 0.0319895446, 0.0321843289, -0.0475018919, 0.0105580324, -0.017618997, -0.0041255332, -0.0361350439, -0.0020321559, 0.0013211106, 0.0070699039, 0.0325593688, 0.063674055, 0.0084488895, -0.0241904333, -0.0267980453, 0.0123781096, -0.0021114342, -0.0318070464, 0.0028094002, 0.0463186651, -0.0227256194, 0.0245531332, -0.0034580203, -0.0067220288, 0.0771481618, -0.0297962334, 0.0324823298, 0.0151889371, 0.007342909, 0.0095071038, -0.0296772122, 0.0557086319, 0.018934248, -0.0165367778, -0.0510811694, -0.0154920341, -0.0458096117, -0.0264473073, 0.0063138395, 0.0529963486, -0.0117627718, 0.0136408964, -0.064782232, -0.001738948, -0.025500644, 0.0032239524, 0.0222301651, 0.0392061174, -0.006823651, 0.0146588758, -0.0495386869, 0.0244334191, -0.03474053, 0.0131940702, 0.0296120178, -0.0940950513, 0.0261945352, 0.002001687, 0.0086549604, 0.0092930952, 0.0162779763, 0.0569881313, 0.0024695941, -0.0064048916, -0.0164636094, -0.0627492294, -0.018530298, -0.0446674302, -0.0357289463, -0.0843688324, 0.0169284977, 0.0287308171, -0.0427333266, 0.0119318217, 0.027019294, 0.0567540862, -0.021013042, -0.0035487323, -0.0240260568, 0.0412427373, 0.0386519581, 0.0286164898, 0.0083582439, 0.0226946864, 0.0145566454, 0.070970647, 0.0502279289, -0.0101673398, 0.0616608821, 0.015739413, -0.0032647068, 0.0066950582, 0.0269339215, 0.0453987047, -0.0168386661, -0.0083798142, -0.0137216151, 0.0778757706, -0.0403963365, -0.0110571319, 0.0562670678, 0.0284802262, 0.0257500559, -0.0016158856, 0.0154949054, 0.0234037228, 0.0223967526, 0.0181022026, -0.0063358643, 0.043443691, -0.0661875904, 0.0034798463, 0.0572352335, 0.0330815129, 0.0128607266, -0.0631169677, -0.0239845421, 0.0245752688, 0.0139040397, 0.0228781123, -0.0115076099, -0.0177155025, -0.0323957838, 0.0190020129, -0.0079709236, -0.0748470053, 0.0213839281, -0.07250572, 0.0493884906, 0.0454412401, 0.0853921771, 0.030341588, -0.0016115818, -0.0067293425, -0.0377233624, 0.0182095282, 0.0536138266, -0.0108616576, -0.032209903, -0.0368371606, -0.0148807876, 0.0148442406, -0.0159106795, 0.0100003779, 0.0035120735, 0.0287298802, 0.0447802395, -0.046282243, -0.0466094576, -0.0021330549, 0.042731341, -0.0284092911, -0.0063869148, -0.0098361885, 0.0157446899, 0.0506402962, -0.0266714487, 0.0617786087, -0.0024928185, 0.0049240468, -0.0315096304, -0.0489249304, -0.006519591, 0.0311397482, 0.0242680963, -0.0414763764, 0.0127641093, 0.0190922134, -0.0059342268, -0.0305390786, -0.0186664108, -0.0654332489, -0.0018806936, -0.0778435469, -0.0058330158, 0.0093426788, -0.0142114274, -0.0608957782, 0.0521702692, -0.0019229624, 0.0537043326, 0.0182365496, -0.0337755233, 0.0427157171, -0.0496755093, -0.0293365661, 0.0369241275, -0.0087041343, 0.048802983, 0.084095493, 0.0305246189, -0.0523677766, -0.0214406811, -0.0994553566, -0.0162239987, -0.0100210514, 0.0432584248, 0.0171206519, -0.0088601168, 0.0008872219, 0.0095615201, -0.0951687321, -0.0113891643, 0.0736636892, 0.0281110443, 0.0233600419, 0.0012616437, 0.002719505, -0.0551618487, -0.0129956994, 0.0150356917, 0.0070274379, 0.0526008345, 0.029481139, -0.0574788377, 0.0109643973, -0.0837808475, -0.0229202472, 0.0281995907, -0.0663651526, 0.0256895944, -0.0010435012, -0.0039931769, -0.0107717104, 0.0347658917, -0.021030575, 0.030242499, -0.0225647129, -0.0833220258, -0.0219863858, -0.0233741179, -0.0390049629, 0.0120201027, -0.0836708024, 0.0402712375, -0.0051511377, 0.0443323217, -0.0049072565, -0.0540574305, 0.0178951696, -0.0136981867, -0.034605328, -0.0303740967, -0.0251958221, -0.0203947891, 0.056959372, 0.0079246182, -0.0599334054, 0.0631699488, -0.05235333, 0.0066377711, -0.0277934521, -0.0039502853, -0.0244005304, 0.0284673218, -0.0289927851, 0.0317021981, -0.0120990053, 0.0092302049, -0.0395411327, -0.0080725839, 0.0324287377, -0.0995154753, 0.0029019632, -0.035058856, -0.0030845762, -0.0029583939, 0.0034191266, -0.0456170179, -0.0459174551, 0.0144974757, 0.0168141834, 0.0124303112, 0.0070578032, -0.0043036686, -0.0031006881, 0.0442285053, -0.0221471805, -0.0370173939, -0.0191317219, 0.0184515696, 0.0547986254, 0.0750874057, 0.0138042578, 0.0527184792, 0.0127993645, -0.0335951485, 0.0816239342, 0.0166507941, 0.0095555056, 0.0338718332, -0.0376592502, 0.0527205803, -0.0225133933, -0.0444196165, 0.0299865305, 0.0224943236, -0.0387694798, -0.0132533843, 0.0672773868, 0.0435909964, -0.0072952583, -0.0229211021, -0.0330035537, 0.0123074157, 0.0096952794, 0.0593742244, -0.0917186812, -0.0019505692, -0.019677341, 0.0575339198, -0.0222583693, -0.0164778512, -0.0032982233, -0.1284004301, 0.056397181, 0.0011314714, 0.0050383084, -0.0147147551, -0.0348264314, -0.0006501608, 0.0292774849, -0.0049634599, 0.0406802744, 0.000588842, 0.02614001, 0.0117174406, -0.0017001454, 0.0022702904, 0.0449780263, -0.0180942174, 0.0425911061, 0.0833732858, 0.025571214, 0.0071191117, -0.1025435701, -0.0056366432, -0.0475001372, 0.0270008463, 0.0081789037, 0.0168037098, 0.0021422186, 0.0245643258, -0.0467001498, -0.0300593022, -0.0196061675, 0.0035950597, 0.0032860965, 0.0599174611, -0.0208602548, -0.0019293809, -0.0609796457, 0.0071781026, 0.00788486, 0.0771751478, 0.0321737602, -0.0273464099, -0.035257604, -0.0193473194, 0.0344588533, -0.0316672176, -0.0469804481, 0.0118357819, 0.0030734341, -0.0124178892, -0.0997159332, -0.0491009876, -0.0100967223, 0.0158765912, -0.0246237796, -0.0136438608, -0.0410831831, 0.0172883775, -0.0087549547, -0.0107644247, -0.0299575813, 0.0288952608, -0.0557137989, 0.0152948713, 0.0105366753, -0.0521400832, -0.0130899986, 0.0023034588, -0.0138141057, -0.0370415077, -0.0042514768, 0.0452987291, 0.0299015567, -0.0223774649, -0.01294633, -0.0323745571, 0.0271835364, 0.0303229392, -0.055448778, 0.0247161724, 0.0021684659, 0.0021492927, -0.0103413388, -0.0035626779, -0.0831161514, 0.0324905403, -0.0077319364, 0.0501988083, 0.0340778641, -0.078113243, 0.0228823815, 0.0825393051, -0.0051726825, -0.0028411313, 0.0098243877, 0.0017122612, -0.0032929005, -0.0169405621, -0.0120519213, -0.0389468633, -0.0148539068, -0.0158699881, -0.0272570793, 0.0117999893, 0.0006170304, -0.0552203543, 0.0667589977, -0.0230787266, 0.0129909059, 0.0030224361, -0.001228264, -0.0163238626, -0.0134565439, -0.0469144173, -0.0265058614, -0.0099313008, 0.0497171432, -0.0058656177, 0.0692462176, -0.0178404655, -0.0191050824, 0.0677374974, 0.0491554141, -0.0621323511, -0.0535352677, -0.0351097696, 0.014257892, 0.0126279872, 0.0405584648, 0.0689616203, -0.0140258484, 0.0114924423, -0.0204388611, -0.0054289941, 0.0344233662, -0.0286859684, 0.0153947175, 0.0173187647, -0.025834756, -0.0016815744, 0.0426236279, -0.0446571, -0.0232215226, -0.0712162256, -0.0420158803, -0.0204715133, -0.0516724959, -0.0125722392, 0.0080373855, 0.0005337285, -0.0309930854, -0.0932205319, -0.0371769741, -0.037889652, 0.0654523447, 0.026599396, -0.0156326443, 0.0243741702, 0.0886849165, 0.0194658469, -0.0076726642, -0.0761258826, 0.0112740817, -0.0681924, 0.0257767923, -0.0516731367, -6.550375414e-33, -0.012339347, -0.0516976044, 0.0107816514, -0.0402617492, -0.0093393419, -0.0032824946, -0.0153672369, 0.0174032785, -0.022779271, -0.0262531266, -0.015709646, -0.0028474466, 0.0058430303, -0.0088298097, 0.0033077474, -0.0044438839, 0.036025513, -0.0012275699, -0.0607013442, 0.0406968035, 0.0326715, -0.0324361958, 0.0421998277, -0.0505501814, 0.0522786528, -0.0285749566, -0.0253983513, 0.0058033476, -0.0127041526, 0.0677857324, 0.004168232, 0.0520818755, 0.0249361955, -0.039272245, 0.0160634313, 0.0041319067, -0.0288699567, -0.0317856595, 0.0423683301, -0.0261067562, -0.0483901761, -0.0033713703, 0.0378936231, -0.0235468112, -0.0492288172, 0.0274794046, 0.0245113187, -0.0138153154, 0.0089759314, -0.0581731722, -0.0171101317, 0.0244713034, 0.0200243611, -0.0094458647, -0.038180206, 0.0854370147, -0.0018548153, 0.0029580991, -0.0446796007, 0.0251841601, 0.0030812414, 0.0314665809, 0.0831664801, -0.038242314, -0.0203485992, 0.0622030385, 0.0173182394, -0.0161295198, -0.010520637, -0.0513097681, 0.0057227644, 0.020154722, 0.0608016327, 0.0720529631, 0.0325909741, -0.0236632116, -0.0082884636, 0.0908129066, 0.0249657705, 0.0530787669, 0.0100461598, -0.0203628372, -0.052703727, 0.0027649791, -0.0359331444, 0.0226691253, -0.0662191659, -0.0759916753, 0.0444136076, -0.0020765872, -0.0337423645, -0.0409710072, 0.0412923284, -0.0266540553, -0.0122679854, 0.0307327099, -0.0359782465, 0.0080459686, -0.0109671233, -0.0052691414, -0.0029530937, 0.0526000075, 0.0250335392, 0.0074341367, -0.0057049477, 0.0099668577, -0.0175972767, -0.0064180312, -0.0573427714, 0.0198899768, -0.024037607, 0.0434655659, -0.0190921538, -0.0545737855, 0.0139117585, -0.0106775304, -0.0162547417, 0.042658221, -0.0091834906, 0.0260928515, -0.050262101, -0.0460827947, 0.0285218004, 0.0058282004, -0.041871164, 0.0279409438, -0.0010064533, 0.0223837681, -0.0137553327, 0.0080989832, -0.0167688224, 0.0043077017, 2.822e-7, 0.0393793918, 0.0145822112, 0.0095231161, 0.0343921892, -0.0078935968, 0.0337841883, 0.0231822655, 0.0105197197, 0.0468943119, -0.0078806439, -0.0000122675, -0.0214587878, 0.0370340236, 0.0661009252, -0.0381956808, 0.0063486374, -0.0655059367, -0.0051349211, -0.0156926587, 0.0009556094, 0.1092028916, 0.0279472228, -0.020474067, -0.020979492, -0.0306308698, -0.0104526198, 0.0377607606, -0.0029845207, -0.0138685089, 0.032020025, 0.0977102667, -0.019297421, 0.0465907268, -0.0146842469, -0.0105512543, -0.0203027986, 0.0339253955, 0.0326365605, -0.0009289848, -0.0561679192, 0.0001739332, -0.0432963148, 0.0401408821, -0.0673340484, 0.0679983795, 0.0041704588, -0.0087806266, -0.0350162983, -0.0708052889, -0.0286680441, 0.0155266598, -0.015596183, -0.0447310656, 0.0334478877, -0.0228208601, 0.0157148186, -0.0066495696, -0.0160237439, -0.032198485, 0.0095349234, -0.0127619933, -0.0726346523, -0.0189864561, 0.0043476787, -0.0453003272, -0.055879619, 0.0103094857, 2.59536709e-34, 0.0241011344, -0.0123682069, 0.0290189441, 0.0043482636, 0.0437488705, 0.0463885963, 0.0015750817, 0.0149383927, 0.0216115545, -0.0474864431, -0.0244314 ]
http://arxiv.org/abs/2406.18060v1
http://arxiv.org/pdf/2406.18060v1
AdaZeta: Adaptive Zeroth-Order Tensor-Train Adaption for Memory-Efficient Large Language Models Fine-Tuning
2024-06-26T04:33:13
2024-06-26T04:33:13
Yifan Yang, Kai Zhen, Ershad Banijamal, Athanasios Mouchtaris, Zheng Zhang
cs.CL
cs.CL, cs.AI, cs.LG
Fine-tuning large language models (LLMs) has achieved remarkable performance across various natural language processing tasks, yet it demands more and more memory as model sizes keep growing. To address this issue, the recently proposed Memory-efficient Zeroth-order (MeZO) methods attempt to fine-tune LLMs using only forward passes, thereby avoiding the need for a backpropagation graph. However, significant performance drops and a high risk of divergence have limited their widespread adoption. In this paper, we propose the Adaptive Zeroth-order Tensor-Train Adaption (AdaZeta) framework, specifically designed to improve the performance and convergence of the ZO methods. To enhance dimension-dependent ZO estimation accuracy, we introduce a fast-forward, low-parameter tensorized adapter. To tackle the frequently observed divergence issue in large-scale ZO fine-tuning tasks, we propose an adaptive query number schedule that guarantees convergence. Detailed theoretical analysis and extensive experimental results on Roberta-Large and Llama-2-7B models substantiate the efficacy of our AdaZeta framework in terms of accuracy, memory efficiency, and convergence speed.
AdaZeta: Adaptive Zeroth-Order Tensor-Train Adaption for Memory-Efficient Large Language Models Fine-Tuning: Fine-tuning large language models (LLMs) has achieved remarkable performance across various natural language processing tasks, yet it demands more and more memory as model sizes keep growing. To address this issue, the recently proposed Memory-efficient Zeroth-order (MeZO) methods attempt to fine-tune LLMs using only forward passes, thereby avoiding the need for a backpropagation graph. However, significant performance drops and a high risk of divergence have limited their widespread adoption. In this paper, we propose the Adaptive Zeroth-order Tensor-Train Adaption (AdaZeta) framework, specifically designed to improve the performance and convergence of the ZO methods. To enhance dimension-dependent ZO estimation accuracy, we introduce a fast-forward, low-parameter tensorized adapter. To tackle the frequently observed divergence issue in large-scale ZO fine-tuning tasks, we propose an adaptive query number schedule that guarantees convergence. Detailed theoretical analysis and extensive experimental results on Roberta-Large and Llama-2-7B models substantiate the efficacy of our AdaZeta framework in terms of accuracy, memory efficiency, and convergence speed.
[ 0.019506108, 0.0597144701, 0.0002742128, 0.0176635012, 0.0069130743, 0.0337005779, -0.0056393566, 0.0120048905, -0.0040329592, -0.0290170498, -0.0462414548, -0.0738393366, 0.009258206, -0.0064684879, 0.0052194893, -0.0025142874, 0.0070982208, -0.049424585, 0.0111896442, -0.0109018292, -0.0188236963, 0.0258570593, -0.0143819656, -0.0082888985, 0.0359697528, -0.0000043418, -0.0088785309, -0.0005795437, -0.0055763004, -0.0163802337, -0.0102048647, 0.0633997172, 0.0195519459, 0.0597088337, 0.0000019107, -0.0120007396, -0.0115171243, -0.0058785318, 0.0119503103, 0.0058574276, 0.0450428016, -0.0554008484, -0.0242409278, 0.019052349, 0.0051462492, -0.0277875531, 0.0561840534, 0.0401462652, 0.0387814231, 0.0939652771, -0.0155095924, -0.0168570336, 0.048977159, -0.0396881476, -0.0200422835, -0.0824259073, 0.0504127964, -0.0530313142, -0.0288587734, -0.0022515245, -0.0038546836, 0.0271744262, -0.0342535935, 0.0189079978, -0.009076071, 0.0316375606, -0.0162778925, -0.0115467729, -0.0166379847, 0.0744984448, -0.0254169255, 0.014169991, -0.0037563825, 0.0009281151, -0.0403673239, -0.0053583249, -0.0427065715, 0.0092567196, 0.0494790338, 0.012829328, 0.0070170895, -0.0607891753, 0.0258503389, 0.0341646634, -0.0862762406, 0.0451839045, 0.0471913368, -0.0050429478, 0.023943197, -0.0092847412, -0.0309977345, 0.0033943185, -0.0061034774, 0.0079570329, -0.0181358065, -0.021426389, -0.0362461843, -0.118784219, -0.0097353859, -0.048904974, 0.1127542481, 0.0510424264, -0.0078466069, 0.0526877008, -0.057982035, 0.0282007623, -0.0177439358, -0.0328990854, -0.0417922251, -0.0271797366, -0.0079247933, 0.0238504279, -0.0553181954, 0.0081046289, -0.0458900928, -0.0375297256, -0.0504278354, 0.0159783605, -0.0047386829, -0.0472309254, 0.0061269049, 0.0086716907, -0.0027259279, -0.0022923511, -0.0190913789, 0.0764708221, 0.009008673, 0.0083108135, -0.0857789442, -0.019635668, -0.0265370272, 0.0321456045, 0.0387512334, -0.0349525996, 0.0368125103, 0.025708694, -0.0152990865, -0.0466874465, -0.0206702948, 0.0379555561, 0.0136929983, -0.0823110566, 0.0193083119, -0.0308981687, 0.0438622944, -0.0098453285, 0.0310503691, -0.0294894632, -0.0056213913, -0.0005009004, -0.1040291414, 0.050103616, 0.0100158053, 0.0117602842, -0.0256111752, -0.0093896994, 0.0885332376, 0.0651171431, 0.0361809097, -0.0124982931, 0.0075203637, -0.0159392748, -0.0601089187, 0.0429540537, 0.0341833085, 0.0208921321, -0.0440550782, -0.0016633524, 0.0086389109, 0.0518510379, 0.0351103283, 0.0376284122, -0.047222808, -0.0186045095, 0.0339479037, 0.0398946106, 0.0659065694, 0.0102737341, -0.0062572919, -0.0386291556, -0.0271141157, -0.0479260944, -0.0575947985, 0.0429903828, -0.0405567512, -0.0054568592, 0.0654703379, 0.0498451628, -0.0320279486, -0.0048704916, -0.0015618507, -0.0041928361, -0.000921146, 0.0007428969, 0.0706197098, -0.0132087152, -0.0283453744, 0.0318382457, 0.0304613318, -0.0320401862, -0.0176512171, -0.0445525199, 0.0132749472, 0.0687265769, 0.0233448092, -0.036034219, -0.0201079082, 0.0166712478, 0.0370308422, 0.00879806, 0.0083356816, 0.0245597716, 0.0423063487, -0.0595215745, -0.0084790699, 0.0266720541, 0.0088566598, -0.0031201863, -0.003510959, 0.052311942, 0.0330454297, -0.0137498323, -0.0230364092, 0.0308159124, -0.0312153995, -0.0442011803, 0.0359988324, 0.0089536095, 0.0026530158, 0.044044625, -0.0008806035, -0.00727128, 0.0374695994, -0.0342058614, 0.0018975287, -0.0768453553, -0.0262023862, 0.0119640585, 0.0468220077, 0.0181722436, 0.0421353653, 0.0125925103, -0.0532065481, -0.0549311787, 0.0955339596, 0.0090059079, 0.0221031234, -0.0439235009, 0.0093671801, 0.0148230875, 0.0634550974, -0.0175503641, -0.0022459556, -0.0112281451, 0.0064047072, 0.0231005456, -0.0295858029, -0.0064151688, -0.0061531425, -0.0659122765, 0.0573036298, 0.0422467254, 0.0310117584, -0.0002226063, 0.007886067, -0.0037614417, -0.0299116652, -0.1024891287, 0.0247857105, -0.0463754199, 0.0065087541, 0.0308152661, 0.0116280001, -0.0011784276, -0.0074855089, 0.010604335, 0.0420239307, -0.0296792425, 0.0274140015, -0.0055121863, -0.0075766877, -0.0138496226, -0.0034964662, -0.0455764607, 0.0043168026, 0.0388374999, -0.0544584766, 0.0060027549, -0.0308335386, -0.0383162163, -0.0441133454, 0.0627107918, 0.0028379729, -0.0088397283, 0.0061520301, -0.033516854, 0.0470321365, 0.0554990098, 0.0058493637, 0.0245571379, 0.0573731624, 0.0135872243, 0.0040772995, -0.0090085715, -0.0463933349, -0.0626908913, -0.0245499704, -0.0254334584, -0.0273933038, 0.0637887493, -0.0290469658, -0.0151124801, -0.0426231809, -0.0237717684, 0.0057895114, 0.012295776, -0.0048264479, 0.0307388492, -0.009219043, -0.0034239597, 0.0181685071, 0.0075509697, 0.0352893621, -0.0337636806, -0.0088129872, -0.0352980494, -0.0242832508, -0.0431490988, 0.0019025378, 0.0170184206, 0.0415507071, 0.008365849, 0.0161036383, -0.0127245951, -0.0293822382, 0.0247222092, -0.0222916752, 0.0356317125, -0.0095966291, 0.0157089084, -0.0309955329, -0.0022747996, 0.0103957709, -0.0221914016, -0.0163617712, 0.04803418, -0.0207167435, -0.053352803, -0.0283507593, 0.0368867628, 0.0039604683, 0.0630553216, -0.0334382653, 0.0057580909, -0.0147744799, 0.0428488441, 0.075754188, 0.0288399216, 0.0164516885, 0.0064782156, -0.0072191004, 0.0529488027, 0.0729270875, -0.0089696413, -0.0444270372, -0.0127641885, 0.0503391884, -0.0193480235, 0.0017831621, 0.00038086, -0.011305932, -0.0358651392, 0.0177352708, 0.1093422845, 0.0047378591, -0.0524617247, -0.0067844945, 0.0363284163, 0.0503362194, 0.0253131203, 0.0335214175, -0.0612914935, -0.0296461117, -0.0137598692, 0.057138972, 0.0052414741, -0.0418712981, -0.0112089841, -0.0939037204, 0.0074264877, 0.0207391903, -0.0166169014, -0.0074029467, -0.061647296, 0.034672115, -0.0268078372, 0.0434920639, 0.0254802443, -0.004471275, -0.0399055518, 0.0221668966, 0.0697040781, -0.0213226806, -0.0108336043, -0.0222063735, 0.0566204228, -0.0219604298, 0.0299055818, -0.0040160287, 0.0063329106, 0.0429961868, -0.0210443158, -0.0008354694, -0.0178483538, 0.0207938198, -0.0215261783, -0.0001851434, 0.00417588, 0.0412093401, -0.0328870825, 0.0104433186, -0.0408223011, 0.076460585, 0.0193192735, -0.0293018986, -0.0253720656, -0.0185389724, -0.0311065428, 0.0549815185, 0.0066617201, -0.0535533167, -0.0381211489, 0.0410865583, -0.0015601697, -0.013844477, 0.0113203228, -0.0387491994, -0.0205446258, -0.0423068851, -0.021169303, 0.0556560382, 0.0112041356, 0.0384324007, 0.0029474273, 0.0149728721, 0.0177699067, -0.0305884797, -0.0173420925, 0.0622544996, -0.0343313292, 0.0158667509, -0.0658251941, 0.0017699436, 0.0380438305, 0.001177705, -0.0000878089, 0.0137184495, -0.0836575106, -0.0782092959, -0.0074936263, -0.0011292755, 0.008650152, 0.0199445225, 0.0089661647, -0.0825759098, 0.0577106476, 0.027312519, -0.0548951514, 0.0574822128, 0.0227390565, -0.0373283289, -0.0088231089, 0.0438301712, -0.0187095068, -0.0002620241, 0.0306544453, 0.0347074755, -0.0009650656, -0.0502322838, -0.0260451771, 0.0224898718, 0.0525246821, -0.0010723833, 0.0295430869, -0.0401307717, 0.0475712121, 0.0072132437, 0.016729068, -0.0298788808, -0.0066890204, 0.0394884832, 0.0076164966, -0.0420436971, -0.0164639447, 0.0199637394, 0.0054510571, -0.0403420702, -0.0037137892, 0.0077717477, -0.0262708422, 0.0571291968, 0.0038770691, -0.0056459694, 0.0272451378, 0.0029234339, 0.061970789, -0.0533568487, 0.0406720378, 0.0098643377, -0.0085964054, -0.0809813812, 0.0540756248, -0.0160376113, -0.0845510215, -0.0153490556, 0.0290913247, 0.0481132902, -0.0116121192, 0.0600123145, -0.0081839571, -0.0194474068, -0.0099712266, 0.0130199753, -0.0261713453, 0.0269447826, 0.0122714257, 0.0001611948, 0.0081905462, 0.0701841041, 0.0489228591, 0.0480894968, 0.0278024469, -0.014639535, -0.001612911, -0.0053190519, -0.0237170309, 0.0107983928, 0.0638979226, -0.0449436083, 0.0188182238, -0.0093443124, -0.0641473606, -0.0209984239, -0.0740678757, -0.0376956016, -0.0515708365, 0.0206873138, 0.0204038881, -0.0124833779, 0.0296371225, -0.0106828045, -0.0164183341, -0.0076116864, -0.0327266864, 0.0239231288, -6.566315117e-33, -0.0318958163, 0.0358268395, 0.0012705507, -0.0412256345, -0.0272593126, -0.0208310131, -0.0152338622, 0.0040819086, -0.0232022107, -0.0225422792, -0.016994819, -0.0131015228, -0.0049602934, 0.0095504476, -0.0145725617, 0.0472215973, 0.0450145379, -0.0468170717, -0.0277180634, 0.0308533739, 0.0386937968, 0.0307250097, 0.0693378225, -0.049226515, -0.0299145821, 0.0088489829, -0.0025224285, 0.032947924, 0.0252044778, 0.0176978167, -0.0203838609, 0.0372142531, -0.0013586831, -0.050244242, 0.0111115547, -0.0115555432, -0.0982439294, -0.0135622965, 0.040403679, -0.0108185709, 0.0075058057, 0.0069054761, 0.0174522791, 0.0389381796, -0.010773032, -0.0294146203, 0.0015656486, -0.0024800515, -0.0587419532, 0.0134713817, -0.0429939665, 0.0288207978, -0.0099684028, -0.0029291327, 0.0687131435, -0.0064270762, -0.0180684607, 0.0424338616, -0.0543148182, 0.0030675954, 0.0183376018, 0.0337104164, 0.0417055488, 0.0345835648, -0.0016669456, 0.0171296764, 0.0422042198, -0.0068023889, 0.0097645093, -0.0119294999, 0.0438853092, 0.0090812938, 0.0196194164, -0.0107662007, 0.0090925694, -0.0414228216, -0.0685475916, -0.0535727181, 0.0150744207, 0.0349948481, 0.06137174, 0.0142470738, -0.0344030671, -0.030513281, -0.0446682833, -0.0211620275, -0.059782777, 0.0139107993, -0.0116255339, 0.0334992595, -0.0606897101, -0.0584717765, 0.0310272258, -0.0631948486, 0.0358925574, 0.036648307, -0.0242121834, -0.0112266913, 0.0087629668, 0.0305635259, -0.093615599, -0.0498993881, 0.0020881684, -0.0540634207, -0.0064090998, -0.0129289124, 0.0311548412, 0.0400574133, -0.0816743821, -0.0318239741, -0.0393276997, 0.0067531094, 0.0173366461, 0.0959298387, -0.0203639902, 0.0029457523, 0.006792746, 0.0651928112, 0.0417711921, 0.0306671802, 0.0133576831, -0.0345699154, 0.0538701117, -0.015757395, -0.0355769247, 0.0087259682, 0.0132732568, -0.0108655877, 0.0522318371, -0.0541180708, -0.0410153419, 0.0364295579, 2.756e-7, 0.011613695, 0.0297379345, -0.0348177217, -0.001003294, 0.0316463336, 0.0386890769, -0.0061248979, 0.0215292573, 0.0483477488, 0.0514731184, 0.0357525125, 0.0269421302, 0.007247203, -0.0259492751, -0.0555904806, -0.0214686263, -0.0287431236, 0.0296667218, -0.0547003411, -0.0265626907, -0.027497787, 0.092299968, 0.025301937, -0.0171098448, 0.0135128302, -0.0617757626, 0.0429397225, -0.0059561045, 0.0245507136, 0.0248915907, 0.0252874587, 0.0639017373, -0.00826653, -0.0052036364, 0.0014376098, 0.0750389099, -0.012281986, 0.0645070076, -0.0404975824, -0.0345485099, 0.0514509976, -0.0115981465, 0.0044233766, -0.0579170994, 0.0781484246, -0.0055903047, -0.0144540528, -0.064523913, 0.0003174699, -0.0254983027, 0.0382602736, -0.0288833603, 0.0328253321, 0.0290873032, 0.0564721487, -0.0250737835, -0.0266339369, -0.0647219121, 0.0057095666, 0.0212450083, -0.0134267686, -0.0576560348, -0.0314663313, -0.0238940809, 0.0165839382, -0.0241181105, -0.0032389062, 2.125664572e-34, 0.0464235246, -0.0203996673, 0.002073688, 0.0204154942, 0.0257545765, -0.0323357433, 0.0043684295, 0.0100257564, -0.0055877734, -0.1345548332, -0.0409772247 ]
http://arxiv.org/abs/2406.18616v1
http://arxiv.org/pdf/2406.18616v1
Towards Large Language Model Aided Program Refinement
2024-06-26T04:29:27
2024-06-26T04:29:27
Yufan Cai, Zhe Hou, Xiaokun Luan, David Miguel Sanan Baena, Yun Lin, Jun Sun, Jin Song Dong
cs.SE
cs.SE, cs.AI, cs.CL, K.6.3
Program refinement involves correctness-preserving transformations from formal high-level specification statements into executable programs. Traditional verification tool support for program refinement is highly interactive and lacks automation. On the other hand, the emergence of large language models (LLMs) enables automatic code generations from informal natural language specifications. However, code generated by LLMs is often unreliable. Moreover, the opaque procedure from specification to code provided by LLM is an uncontrolled black box. We propose LLM4PR, a tool that combines formal program refinement techniques with informal LLM-based methods to (1) transform the specification to preconditions and postconditions, (2) automatically build prompts based on refinement calculus, (3) interact with LLM to generate code, and finally, (4) verify that the generated code satisfies the conditions of refinement calculus, thus guaranteeing the correctness of the code. We have implemented our tool using GPT4, Coq, and Coqhammer, and evaluated it on the HumanEval and EvalPlus datasets.
Towards Large Language Model Aided Program Refinement: Program refinement involves correctness-preserving transformations from formal high-level specification statements into executable programs. Traditional verification tool support for program refinement is highly interactive and lacks automation. On the other hand, the emergence of large language models (LLMs) enables automatic code generations from informal natural language specifications. However, code generated by LLMs is often unreliable. Moreover, the opaque procedure from specification to code provided by LLM is an uncontrolled black box. We propose LLM4PR, a tool that combines formal program refinement techniques with informal LLM-based methods to (1) transform the specification to preconditions and postconditions, (2) automatically build prompts based on refinement calculus, (3) interact with LLM to generate code, and finally, (4) verify that the generated code satisfies the conditions of refinement calculus, thus guaranteeing the correctness of the code. We have implemented our tool using GPT4, Coq, and Coqhammer, and evaluated it on the HumanEval and EvalPlus datasets.
[ -0.0248057991, 0.0716641992, -0.0125206886, 0.0364422314, -0.0101864832, -0.0086787567, -0.0579825863, -0.0277496632, -0.0029434136, -0.0086768679, 0.0112043293, 0.0083627636, 0.0236807056, 0.0442799814, 0.0067828102, -0.0107714869, 0.0454992391, -0.043644052, 0.0047477908, 0.0242253207, -0.0097716944, 0.0214641578, 0.0417940877, 0.0354853608, 0.0239010677, -0.0322591625, -0.004501686, 0.0426375717, -0.0310694259, -0.0216246787, 0.0093598254, 0.0589363724, 0.0073005646, -0.003960595, 0.0000024411, -0.0612398386, -0.0353833437, 0.0569160245, 0.007897147, 0.071089536, -0.0038023842, -0.0017057781, -0.0020493416, 0.0241412148, -0.0486901179, -0.0438395068, 0.0196289923, -0.0263551343, 0.1168448254, 0.0593328252, 0.0116466656, -0.0629876778, 0.0079279197, 0.0349930823, -0.0124008162, -0.0281984787, 0.0289722811, 0.0176478047, 0.0011895059, 0.0044421977, 0.0299453773, 0.0015794077, 0.0435304828, -0.0334495269, -0.0256993063, -0.0101187052, -0.0515081398, -0.0224399101, -0.0095524499, 0.019866446, -0.0034328473, -0.0569675714, 0.0107259341, 0.0189165808, -0.0199351422, -0.0499168634, -0.0298868325, 0.0277474448, -0.0067923316, 0.0238209683, 0.0069299899, -0.0528766029, 0.0300528705, -0.0132108843, -0.0247354973, 0.0895710364, 0.0233101584, -0.0250388719, -0.0349449106, -0.0438585877, -0.0204122812, -0.0420579389, 0.0246717371, 0.0416406989, -0.0325276256, -0.0143013615, -0.0267444011, 0.0017460785, 0.0391326621, -0.00637687, 0.0287806168, 0.0384908579, 0.0030351048, 0.0140912067, -0.0320348665, 0.0556891002, 0.0017091867, -0.0258270912, -0.0258668736, 0.024053216, -0.0164522268, -0.012155734, -0.0785855874, 0.0060981037, 0.0076318509, 0.0865417719, -0.0350531451, 0.0514016189, -0.0258642584, 0.0911100954, -0.0140690776, -0.0536047295, 0.0143139157, -0.001236948, 0.0263012815, -0.0240239408, -0.0061166212, -0.0456138439, -0.036773473, -0.0574789792, 0.0114748282, 0.0179472249, -0.0621860474, -0.006277448, 0.0087205004, 0.0466224067, 0.0042430125, 0.0461079143, -0.0483568422, -0.0562030897, 0.0333976522, -0.0255724676, -0.0245337673, -0.1153130308, -0.0211452916, 0.0003700251, -0.0102173304, -0.0307727475, -0.0203404874, -0.0232720505, 0.0201492738, 0.0457515828, -0.0685474724, 0.0287090875, 0.0204233006, 0.0255262032, 0.0272203758, 0.0217833295, 0.0381556153, -0.0372112021, 0.0185698513, -0.0525181182, 0.023564823, 0.0520256385, -0.0353112333, 0.0112613924, -0.0291733481, 0.0306688696, 0.0245013088, -0.0157722458, -0.0293325763, 0.0399538241, -0.0096743535, 0.0141960159, -0.0186373331, 0.0202751011, 0.0021302772, 0.0219845586, 0.0724190697, 0.056846343, 0.0050519919, -0.0070604449, -0.0368487164, 0.0007996103, -0.0175634101, 0.0128839063, -0.0147106424, 0.0483211577, 0.0011367593, -0.0785406604, -0.0019060263, 0.0323406011, 0.02645972, -0.0287699997, 0.0170850772, 0.0165675674, -0.0232689623, 0.030105941, 0.009972333, -0.0703293979, -0.0302941315, -0.029282093, 0.0536035635, 0.0555739589, 0.0401430689, -0.0472852997, -0.0443861336, -0.0011803814, -0.0148197813, 0.0015332878, 0.0435735025, -0.0189580712, -0.0250671841, -0.0227956399, 0.0023373864, 0.0104815578, 0.0385606363, 0.0241331104, 0.0242976267, -0.028177442, 0.034285605, 0.0096263075, -0.0643708408, 0.0385560505, 0.0241113082, -0.0361843593, 0.019506041, 0.0021303291, 0.0222466774, -0.016254494, -0.0410531797, 0.0260510929, 0.0761161223, 0.0346727408, 0.0133690499, -0.0563730486, -0.0025493165, -0.0058608721, 0.0460426919, -0.0102154445, -0.0520991683, 0.012916985, -0.064363271, -0.0006742051, 0.0439706594, 0.0265896227, 0.0128672086, -0.020728657, 0.059754815, -0.0537568443, -0.0260704234, -0.0231594499, -0.0059466967, 0.0541669279, -0.0179750659, 0.0066340407, 0.0111289453, 0.000669149, 0.0273972135, -0.0384430289, -0.0202199053, -0.0094622811, 0.0347460546, 0.0016412825, 0.0989617184, 0.0036831943, -0.0009552557, -0.1028385684, 0.0606659055, -0.0069158385, 0.0133083733, 0.0196526647, 0.0281814616, -0.0153050851, 0.0071809608, 0.0147778047, -0.0103022028, 0.0084043574, 0.0099175842, 0.0381076485, -0.0146309193, -0.0917469412, 0.0129938331, -0.0055086268, -0.0339741148, 0.0258897264, -0.0008613096, 0.016716281, -0.0877332911, 0.0406196788, -0.0071519115, 0.0253369156, 0.0008027504, -0.0107601667, 0.0264133047, -0.0281703454, 0.0085653206, 0.0132890809, -0.0136969294, -0.0174680855, 0.0634103715, 0.0098592341, -0.0358458571, 0.0225746725, 0.0390643515, -0.0498917513, 0.0051249168, -0.0183999874, 0.0366589725, 0.060942784, -0.0289870501, -0.0166635141, -0.0193309449, 0.0187062714, 0.0125865415, 0.0302486513, -0.0164541695, -0.0585622564, -0.0163060874, 0.0281729642, -0.003481379, -0.0724255964, 0.0771720186, -0.0693133399, -0.0375513434, -0.0060302909, -0.0424356274, -0.0284550041, 0.0480499007, -0.01678502, -0.0046714456, 0.0094732996, 0.0052777152, 0.0130005451, -0.0197586361, 0.0170578621, -0.0263086297, 0.111837171, -0.0691509172, 0.0087356893, -0.015900271, -0.0224518403, -0.0311897304, -0.0145199783, -0.0045778938, 0.0127776619, 0.0140648251, -0.0077168355, 0.0591896102, 0.0117711667, -0.0279874839, 0.0457737297, 0.0155709786, -0.0393519364, 0.0089499978, -0.0562978946, -0.0271794647, -0.0337256379, -0.0196958669, 0.0062073539, 0.0099657606, 0.0256693605, -0.0126752257, 0.0032560325, -0.0848962814, -0.020357715, 0.0338124707, -0.0139105115, -0.0078590401, 0.0552864783, 0.0287570283, -0.0888318792, -0.02026687, 0.0711589828, 0.0179626383, -0.0283916164, 0.0160679873, -0.0078690611, 0.0583063401, 0.0483132564, 0.066635929, -0.0914749876, -0.0259160437, -0.003102741, 0.0286626685, 0.013698048, 0.0121200783, -0.0439570583, -0.0932272971, 0.0707578063, 0.0293609053, -0.0372016802, 0.0193762127, -0.0246900227, -0.0283522159, 0.041520834, -0.0199960433, 0.0718434751, 0.0156256892, -0.0073417625, -0.0204000827, -0.0356387906, -0.0048011704, 0.0446640477, -0.0253657866, 0.0187742747, 0.055490911, -0.0063361772, 0.0003264633, -0.0096790921, 0.0890223533, 0.0178931355, -0.005894226, -0.0631209314, -0.0440573134, 0.0127731031, -0.002610588, 0.0238716155, -0.0035975301, -0.0289236624, -0.019669354, 0.0549504422, 0.0704097301, -0.0144153684, -0.0436562821, 0.0121649755, -0.0174751878, -0.0094746575, 0.0365838073, 0.0122260498, 0.0135855963, -0.0291560832, -0.0500857532, -0.0685797632, -0.0173675213, -0.0095762257, -0.0471812561, -0.0311067458, 0.0016020049, -0.0110250209, 0.0703908652, 0.039408952, 0.0188964531, 0.0274693519, -0.0357831307, 0.047631301, -0.0334037505, -0.0744029284, 0.0436736718, -0.049254518, 0.0106880842, -0.020561995, -0.0053053675, 0.0033056, -0.0040592183, -0.0183025636, 0.0607824549, -0.0345694982, -0.0508266762, -0.040695224, 0.0794584081, 0.0226207301, 0.0273452085, -0.0541145653, -0.0385473184, -0.0487209819, 0.0271365587, -0.0426994488, 0.0040476643, -0.0018745363, -0.0091632027, 0.0105480421, -0.0253687501, -0.0509215035, -0.0048174947, -0.0002391778, -0.0329321064, 0.0121096671, -0.0266820807, 0.0445477478, 0.0888344571, 0.029339226, -0.0077097039, -0.0095457826, 0.0005187761, -0.0103996601, -0.0047528245, -0.0626952648, -0.0618849732, 0.0606875382, -0.0017028019, -0.0421839356, 0.0189465974, -0.0261335298, -0.0611741133, 0.0453696847, -0.0108670676, 0.0321053304, -0.0153104123, 0.0022211333, 0.0931808352, -0.0258514695, -0.0019647186, 0.0283398125, -0.0016601271, 0.0304906983, -0.0253496878, 0.0083599947, 0.0166352112, 0.0667760372, -0.0266155023, 0.0211670082, 0.0150341773, -0.0578515455, -0.0070915665, 0.0455923826, 0.0177571662, 0.0585829243, 0.0316635333, -0.0376460478, 0.0037986466, -0.0400411375, 0.0401363745, -0.0018897059, -0.0659506917, 0.01940725, 0.0451861769, -0.0236665495, -0.0520365871, 0.0542096309, 0.0066149193, -0.0105082439, 0.0044070436, -0.0264215302, -0.0382211469, -0.0298443884, 0.0273202974, -0.0312744528, -0.0593189634, -0.0590955056, -0.0327621512, -0.053760279, -0.0120535744, 0.0360445231, 0.0042152349, 0.0144318799, 0.0125558088, 0.0746431127, 0.0109928511, -0.0686282292, -0.0181175787, 0.0034298883, 0.0098231537, -0.0506816581, -0.009028431, -6.494262456e-33, -0.0152116278, -0.0031978379, 0.0204087757, 0.0010499026, -0.0071779876, -0.0305469874, -0.0554190986, -0.0178529732, 0.0617177784, -0.0024350958, 0.0313883573, -0.0023990234, -0.0025044007, -0.0427162647, -0.00501457, -0.0034376294, 0.0376043916, 0.022737287, -0.0842747092, -0.0171947535, -0.0176427998, -0.0211171303, 0.0226300862, -0.0359737054, 0.0094183497, -0.025115246, -0.0105189011, 0.060496565, -0.0270798933, 0.0309726801, 0.0332075544, 0.0042979722, 0.0005068837, -0.0289411061, -0.011918759, 0.0006034614, -0.0349583961, -0.0133510036, 0.0654083863, -0.0100169368, -0.0197069645, -0.0206446089, 0.0814909637, 0.010411201, -0.0587987453, 0.0596376024, -0.0188304838, 0.0149545204, 0.0476610251, -0.0234616715, -0.066721648, 0.0266156308, -0.0519510731, -0.0154662421, -0.0327371061, 0.0059061823, -0.0117812026, 0.0097165005, 0.0335355327, -0.001023515, 0.0656175464, 0.0198652633, 0.0852798596, -0.0151843317, 0.0038275856, -0.007621666, -0.0086615393, 0.0435189083, -0.0395952836, -0.0018584942, 0.0146765737, 0.0144618778, -0.0031230308, 0.0163797699, 0.0096168555, -0.0158287939, -0.0339186117, 0.0414423235, -0.0384561196, 0.0483501591, 0.0081306314, 0.0012222059, -0.0041162795, -0.0327621698, -0.0529228151, 0.0454485305, -0.0443277434, 0.0043439656, 0.0282793976, -0.0086604785, -0.0398280583, 0.0191858951, -0.027159391, -0.0241120141, -0.0101324553, 0.0117411641, -0.048607856, -0.0347538516, -0.0153923389, 0.0292793065, 0.0835233778, -0.0099922828, 0.0193422679, -0.0527782924, 0.0658992603, 0.0162617303, -0.0005727149, 0.0259943381, -0.0481804311, 0.030049311, -0.0421032272, -0.0024701189, -0.005229753, 0.0214978158, 0.013133605, -0.0020781846, 0.0204540696, 0.047921475, 0.0084067332, -0.0366251692, -0.051387988, 0.0388988964, 0.0566974021, 0.0299403947, -0.0055374857, -0.0077495263, 0.0066238102, 0.0218842905, 0.0267812461, -0.0007953727, -0.0313333273, 0.0256389212, 3.226e-7, 0.0192395188, 0.0405345671, 0.0065208687, 0.0367976651, -0.0226695407, 0.02639376, -0.0254585464, -0.0284583997, 0.0172222257, -0.0097958455, -0.0191605333, 0.0417869203, 0.041991584, 0.0219709296, -0.041243948, 0.0147275608, -0.010848904, -0.0056017712, -0.0735088661, -0.0099683627, 0.0556927659, 0.0393712372, 0.0222131945, 0.0030705994, 0.0035634162, -0.0637279525, -0.0249761976, -0.0263675991, -0.0101620117, 0.0623152293, -0.0553502999, 0.000183261, 0.0256452635, 0.0140468618, 0.0075701536, -0.0249592364, -0.0330076478, 0.1250246465, -0.006293579, 0.107702896, 0.0318004079, -0.0399902426, 0.0024721175, -0.0603707954, 0.0391806252, -0.0047528688, 0.0055247489, -0.0374124609, -0.0179560874, -0.0475354083, -0.0128927846, -0.0434207059, 0.0348242335, 0.0140590817, 0.0036101597, -0.0000525221, -0.0306120571, 0.0077192183, -0.0756812096, 0.0243851934, -0.0182346646, -0.0499487408, -0.0169246923, 0.0436606519, 0.0088086613, -0.0362548605, 0.0674814433, 3.538066264e-34, 0.0700887591, -0.0186412353, 0.0286350232, 0.004766332, 0.0393707678, -0.0005967466, -0.0583003499, 0.037009839, -0.0042010974, -0.0118370857, -0.0132861054 ]
http://arxiv.org/abs/2406.18049v1
http://arxiv.org/pdf/2406.18049v1
Improving Entity Recognition Using Ensembles of Deep Learning and Fine-tuned Large Language Models: A Case Study on Adverse Event Extraction from Multiple Sources
2024-06-26T03:56:21
2024-06-26T03:56:21
Yiming Li, Deepthi Viswaroopan, William He, Jianfu Li, Xu Zuo, Hua Xu, Cui Tao
cs.CL
cs.CL, cs.AI
Adverse event (AE) extraction following COVID-19 vaccines from text data is crucial for monitoring and analyzing the safety profiles of immunizations. Traditional deep learning models are adept at learning intricate feature representations and dependencies in sequential data, but often require extensive labeled data. In contrast, large language models (LLMs) excel in understanding contextual information, but exhibit unstable performance on named entity recognition tasks, possibly due to their broad but unspecific training. This study aims to evaluate the effectiveness of LLMs and traditional deep learning models in AE extraction, and to assess the impact of ensembling these models on performance. In this study, we utilized reports and posts from the VAERS (n=621), Twitter (n=9,133), and Reddit (n=131) as our corpora. Our goal was to extract three types of entities: "vaccine", "shot", and "ae". We explored and fine-tuned (except GPT-4) multiple LLMs, including GPT-2, GPT-3.5, GPT-4, and Llama-2, as well as traditional deep learning models like RNN and BioBERT. To enhance performance, we created ensembles of the three models with the best performance. For evaluation, we used strict and relaxed F1 scores to evaluate the performance for each entity type, and micro-average F1 was used to assess the overall performance. The ensemble model achieved the highest performance in "vaccine", "shot", and "ae" with strict F1-scores of 0.878, 0.930, and 0.925, respectively, along with a micro-average score of 0.903. In conclusion, this study demonstrates the effectiveness and robustness of ensembling fine-tuned traditional deep learning models and LLMs, for extracting AE-related information. This study contributes to the advancement of biomedical natural language processing, providing valuable insights into improving AE extraction from text data for pharmacovigilance and public health surveillance.
Improving Entity Recognition Using Ensembles of Deep Learning and Fine-tuned Large Language Models: A Case Study on Adverse Event Extraction from Multiple Sources: Adverse event (AE) extraction following COVID-19 vaccines from text data is crucial for monitoring and analyzing the safety profiles of immunizations. Traditional deep learning models are adept at learning intricate feature representations and dependencies in sequential data, but often require extensive labeled data. In contrast, large language models (LLMs) excel in understanding contextual information, but exhibit unstable performance on named entity recognition tasks, possibly due to their broad but unspecific training. This study aims to evaluate the effectiveness of LLMs and traditional deep learning models in AE extraction, and to assess the impact of ensembling these models on performance. In this study, we utilized reports and posts from the VAERS (n=621), Twitter (n=9,133), and Reddit (n=131) as our corpora. Our goal was to extract three types of entities: "vaccine", "shot", and "ae". We explored and fine-tuned (except GPT-4) multiple LLMs, including GPT-2, GPT-3.5, GPT-4, and Llama-2, as well as traditional deep learning models like RNN and BioBERT. To enhance performance, we created ensembles of the three models with the best performance. For evaluation, we used strict and relaxed F1 scores to evaluate the performance for each entity type, and micro-average F1 was used to assess the overall performance. The ensemble model achieved the highest performance in "vaccine", "shot", and "ae" with strict F1-scores of 0.878, 0.930, and 0.925, respectively, along with a micro-average score of 0.903. In conclusion, this study demonstrates the effectiveness and robustness of ensembling fine-tuned traditional deep learning models and LLMs, for extracting AE-related information. This study contributes to the advancement of biomedical natural language processing, providing valuable insights into improving AE extraction from text data for pharmacovigilance and public health surveillance.
[ 0.0167277679, 0.0680321604, 0.0093114125, -0.0060856845, -0.0041493392, 0.0456447043, -0.0029764397, -0.0148303863, 0.0313550085, -0.0299989823, -0.0078206528, -0.0557020605, -0.0151984459, 0.0360514335, 0.0026413647, -0.0405094363, 0.0404658616, 0.0182334632, 0.0062849023, 0.0214872882, -0.0120341498, 0.0283359028, -0.0179067478, 0.0396260656, 0.0449215211, 0.0140432389, 0.0069814003, -0.0025101637, 0.073116295, -0.0871593058, 0.0152030168, 0.0065963045, -0.0152396299, 0.0841341242, 0.0000019216, -0.0302725285, -0.0346714035, 0.0112869302, -0.0369954444, 0.0013844291, 0.011239877, -0.0360412896, -0.0083667617, 0.0065840641, -0.0140267462, -0.095518209, 0.0390742421, 0.0569481589, 0.0104905795, 0.0753082484, -0.0097915232, -0.0494100004, 0.0692309812, -0.0056918818, -0.006110793, -0.0507327206, 0.0660257712, -0.0435777754, 0.0141914794, 0.0150009459, 0.0405864269, 0.0394693539, -0.0025266025, 0.0199821945, 0.0170595776, 0.0560148358, -0.0427180827, -0.0050240853, -0.0041626128, 0.1070941687, -0.0143688256, 0.0017131992, 0.0223293323, -0.0171506107, -0.0335399024, -0.0323788561, -0.0262332018, 0.0027518503, -0.0362299308, 0.0328950882, 0.0032816348, 0.0165341813, 0.0662275255, 0.0183622148, -0.0228040852, 0.0163613781, 0.0019061932, -0.0343740061, 0.0189595334, 0.0078194775, -0.0194323789, -0.0388853848, 0.0742049813, 0.0290453602, -0.0148834344, 0.0371672586, -0.0085058054, -0.0565054603, 0.0365833305, -0.058581166, -0.014173802, -0.0091552949, -0.0104998145, 0.0299370158, -0.0452024341, -0.0082890876, -0.0158528909, 0.022207899, 0.0370218232, -0.0023117794, -0.0824910179, -0.000465663, -0.0064121815, 0.0660369545, -0.0120925941, 0.0175866727, -0.0697423071, -0.0263069011, 0.0050365967, -0.0015808302, -0.0057004672, 0.0231462289, -0.02560056, 0.0073270155, 0.0227528568, 0.0424873047, -0.0007292277, 0.0017180071, 0.0084352223, -0.0610412695, 0.0131442696, -0.0451599546, 0.0340435095, -0.0228382181, 0.0252193529, 0.0756483078, 0.0345837846, -0.0844028294, -0.0108562326, 0.0010522163, -0.0361954011, -0.0318720005, -0.02482705, -0.0396514237, 0.0612947755, -0.0014037323, -0.0099185081, -0.0446186475, 0.0094205542, -0.0288296062, -0.0783779547, 0.0667634606, -0.0461808741, -0.0044373581, 0.0497139767, 0.0236496199, 0.0577886291, 0.0267130956, 0.0510348417, 0.0210361835, 0.0093173953, -0.0068051405, 0.0192258861, 0.0168341529, 0.0047659269, -0.040292047, 0.0161251072, -0.0100865234, -0.0187274665, 0.0384624526, 0.0008338143, 0.0407965109, -0.0016528735, -0.0045867153, 0.0554370657, 0.0212113652, -0.0013109783, -0.0206331443, 0.0062672347, 0.0053610513, -0.0019004694, -0.0290997587, -0.0192810819, -0.0148935439, -0.0009433135, -0.0091193775, 0.0260207597, 0.0484232381, -0.022661997, 0.0020062537, -0.0148029895, -0.0043127709, -0.0317874588, 0.0102084838, 0.0530957691, -0.054541789, -0.0821396634, 0.0395797715, 0.0538262799, -0.0209191665, 0.0125668487, -0.0257313643, 0.0020871363, 0.0746954009, 0.0196390375, -0.0227215551, -0.0564974546, -0.0413444154, -0.0181396883, 0.0040479549, 0.0152208079, 0.0127500463, 0.0506008677, -0.0272421576, 0.0077597378, 0.0313462205, 0.0293465741, 0.0339816175, -0.0027150149, 0.0430573002, 0.0009144843, 0.0369168036, -0.0308790766, 0.0480974168, -0.0084452964, -0.069511354, 0.0074671702, 0.0174801387, 0.0500508286, 0.0324343629, -0.0000224449, -0.0022478132, 0.0393018201, -0.0645588711, -0.0385564305, -0.0276759304, -0.0284078158, -0.0310110431, 0.0015903311, -0.0080327643, -0.0206841063, -0.004444927, -0.0592856966, -0.01491661, -0.0279846992, -0.0013631415, -0.0137822488, -0.0041842475, 0.0155004784, 0.0002742353, 0.0296480246, -0.0322110541, 0.0456970111, -0.0059863944, 0.0017195478, -0.023779761, -0.050043121, -0.0053896084, -0.0479566418, -0.016232634, 0.0414350368, -0.0067769596, -0.0247480385, -0.0161247049, -0.0262464508, -0.046420563, -0.00274742, -0.0927349404, 0.0447818115, 0.0076666237, 0.0482608378, 0.01906606, 0.0395336561, -0.0236956887, -0.0069887647, -0.001046181, -0.060475301, 0.0046086102, 0.0229441412, 0.0435420796, 0.0332632288, -0.0242622867, -0.0322859362, 0.0012159116, -0.0692731589, 0.032134831, 0.0016336435, -0.0846498981, -0.0228205677, -0.0038049233, 0.007648549, -0.0036126492, 0.0059646508, -0.0338424966, -0.0102855135, 0.0257197246, -0.0014591037, 0.063664861, -0.0074769165, 0.0522266701, 0.0050879358, -0.0106540872, -0.0202714596, 0.0373416208, -0.0316837803, 0.0043731253, 0.0158217307, -0.0268670823, 0.0050510522, 0.0762964785, -0.0057240534, 0.033790499, -0.025039902, 0.0099125188, -0.0237194523, -0.0128338635, -0.016656464, -0.0410419665, -0.0387381315, -0.0155555699, -0.0228611957, -0.0200606342, 0.0541160814, 0.0089528235, 0.0017277518, -0.0423596837, -0.0098980712, -0.0423238538, 0.0125217112, 0.0209805109, -0.0060760658, 0.0130989831, -0.0519323871, 0.0337664187, 0.0130389854, -0.0211535152, -0.0530548804, 0.014809099, 0.0099515859, 0.0277005415, -0.0164502393, -0.0224276222, -0.0469902046, -0.0513452962, -0.0368728116, 0.0168104451, -0.0007699206, 0.01594254, -0.0596233085, 0.0472853333, -0.0036711574, 0.0304049123, 0.0030120963, -0.0007869983, -0.002385803, 0.0404379107, 0.0247858856, -0.0265160222, 0.0465288572, -0.0252237506, 0.0245938469, 0.033001326, 0.0283077788, 0.0291625988, -0.03858402, -0.0103260307, -0.0171322897, -0.0666233078, -0.0021970724, -0.006366367, 0.0743574649, -0.0755029693, 0.0198513716, 0.0936444253, -0.0002542672, -0.0358857103, 0.0170174632, 0.0606248491, -0.0210724827, -0.0350195654, 0.01436006, -0.0418851934, -0.0369721986, 0.0189372618, 0.001234366, -0.0318729542, -0.0893756375, 0.028673999, -0.0556401834, 0.0125230849, 0.001671569, -0.0590129681, -0.0361506566, -0.0679071099, 0.022737734, -0.0249433704, 0.0265898351, -0.032326743, -0.0196766984, -0.0038346453, 0.0220377911, -0.0493570678, -0.1025485024, 0.0272806045, -0.0530323386, 0.0187778324, 0.0556300431, 0.0238501094, -0.0257829148, 0.0113323545, -0.0249256063, -0.0247283876, -0.0474439822, 0.0107961046, 0.0839832276, -0.0154619217, -0.0210850984, -0.0338026285, 0.0200013444, -0.0480748229, -0.0174656082, -0.0076567498, 0.066874221, -0.0102570439, -0.0485804938, -0.0239571128, -0.0254587792, 0.016068561, 0.0707072318, 0.0208574701, -0.017675776, -0.0108156558, 0.0489615686, 0.0217617545, -0.0903092548, -0.0092531499, 0.032981433, -0.0175773557, -0.0760067254, -0.038339857, -0.0491558872, 0.0478199311, -0.0042713801, -0.0114875752, 0.0209796857, 0.0140272724, 0.0289190747, -0.0256929621, 0.059516564, 0.0022010715, -0.011495118, -0.0349467248, -0.0004941217, 0.0137203569, 0.0593962148, -0.0338067263, 0.0523160249, -0.0506555364, -0.0514047854, 0.0180004966, -0.027464455, 0.0132099846, -0.0075339666, 0.0202178955, -0.0291339122, 0.0573438294, 0.072176531, -0.023368381, -0.0551745631, -0.0303800441, -0.0290590543, -0.0158101227, 0.0126972497, -0.0181460306, 0.0400741845, 0.0168383941, 0.0220255386, -0.0136341052, -0.0194917656, 0.0045677251, 0.0172948036, 0.0356314182, -0.0111096902, -0.0703852773, -0.0590263493, 0.0132790217, 0.0470301695, 0.0403389893, -0.0018553649, 0.0001883159, 0.0303486288, -0.021468211, -0.0514498912, 0.0390224233, 0.0027481138, 0.0081590991, -0.0573946014, 0.0500258282, 0.0047674198, 0.0469008386, 0.0474077202, -0.0224231854, -0.0185363311, 0.0288196281, 0.0345989317, 0.0092804916, -0.0549509786, -0.0332951583, -0.0073830248, 0.0212741625, -0.0056561027, 0.0394974463, -0.053839922, -0.0071761874, -0.0407379717, 0.0442936383, 0.0747252405, -0.0231001042, 0.0346142016, -0.0075152363, -0.0113571836, -0.0317393281, 0.0194844808, -0.02256562, 0.0419135317, 0.0335259922, 0.0291751381, -0.0437423438, -0.0147112459, 0.0589627773, 0.0219828319, 0.0833413377, 0.0506900065, -0.0044460515, -0.0377034806, 0.0085331937, 0.0241388995, 0.0074205068, -0.0567948855, 0.0077864449, -0.0255432632, -0.0179012176, 0.0607859641, -0.0672774017, -0.0144241974, 0.0120942527, 0.0112805134, 0.0287192129, 0.026081996, 0.0095483353, -0.0074828644, 0.0274397321, -0.005774898, -0.0381451286, -0.0125865135, -6.145632505e-33, -0.0009422938, -0.0153288413, 0.0026267353, -0.037829306, -0.0115873609, 0.023693094, -0.0391699336, -0.0271114595, -0.0244026072, 0.0005302965, -0.0066023306, -0.0147416685, 0.0201316141, 0.0172182303, -0.0007897794, -0.0301508997, 0.0469826981, 0.0280603394, -0.0013249849, 0.0477038436, 0.0218667556, 0.0948355719, 0.0745838508, -0.0853582174, 0.0152570866, 0.0065600504, 0.0269816294, -0.0268744864, 0.0411539078, 0.0855278894, -0.0439310446, 0.025541123, 0.0428149365, -0.0308562331, 0.0101275947, 0.0431471355, -0.0904216319, 0.0054757395, -0.0358906649, -0.0032201554, -0.0189209934, -0.0728844181, 0.0912618637, 0.0221721549, 0.0085433591, -0.0131067056, 0.0250608511, -0.0143531272, -0.018903736, -0.0561911017, -0.0639107153, 0.0520121269, -0.0014673733, 0.0245872103, 0.0311236046, -0.0182269439, -0.0113674849, 0.1093434095, -0.0211318526, -0.0492019244, 0.0164752565, 0.0069256811, 0.0226652343, 0.006583666, -0.0157047119, 0.0619897619, 0.0123902569, -0.0154805258, -0.0657944381, -0.0159207806, 0.0075133294, 0.0448878407, 0.0037008785, 0.0394185334, 0.0482524186, -0.0123866685, -0.1378639489, 0.0245302673, -0.0253028013, 0.0262706447, 0.0031455124, -0.0025517256, -0.0510274395, 0.0444863588, -0.012860192, 0.0029930666, -0.0203598477, 0.0051149819, 0.0096568391, -0.0198070947, -0.0295190234, 0.0589708425, 0.001710795, -0.0196601637, -0.0613378398, 0.0015250572, 0.0587404668, -0.0023772866, 0.0019057406, 0.0228474662, -0.0646854192, -0.0324210078, -0.008309667, -0.0021840024, 0.0307631902, -0.0108098993, 0.0503106639, -0.0345271528, -0.0502660088, -0.0091465404, -0.0029467596, 0.0407980047, -0.042741999, 0.0842031613, 0.0347051471, 0.007172415, 0.0204709023, 0.0362235047, 0.0168477148, 0.0164925922, 0.0146289095, -0.0317819826, 0.0086867316, 0.0359853469, -0.0016228495, -0.0135318413, 0.0152285853, 0.0243581552, 0.0053197327, -0.0661865026, 0.0037666075, -0.0225572735, 2.743e-7, 0.0862659439, 0.0547135845, 0.0213068146, -0.0477283373, 0.0072214645, 0.0193466749, 0.005565925, 0.0199885741, -0.0263153762, -0.024129767, 0.0066169854, 0.0153592089, 0.0360769406, 0.0290598217, -0.0882708579, -0.0093045691, -0.0784115717, 0.0259036142, -0.0592498183, 0.0009812799, 0.0405722968, 0.0983965546, 0.002121978, -0.036487449, -0.0102363452, -0.0008090671, 0.0294091422, -0.0102010099, -0.0033034466, 0.0566144586, -0.0109983739, 0.0558892973, -0.0257772133, 0.0488271751, -0.0102852695, 0.006246828, -0.0166490134, 0.0636480227, -0.0309340805, 0.0062170308, 0.0036990645, -0.0335066877, 0.0087562213, -0.0177171901, 0.0794258192, -0.0613482967, -0.0453655683, 0.0083212946, -0.0485677049, 0.0077526835, 0.0368666947, -0.0028290981, 0.0067780395, 0.054062549, -0.0127338264, -0.0397386327, -0.0383611061, -0.0837728083, 0.0314896144, 0.0407847464, -0.033958748, -0.0266608875, -0.0066543366, -0.0127288662, 0.0281171091, -0.0385712124, 0.0218444318, 2.461692489e-34, 0.0012039109, 0.0092457328, -0.0199026726, 0.0351406299, -0.0163570587, -0.0122267054, -0.0135913277, 0.0047290092, -0.0349263437, -0.0416546278, 0.0254102666 ]
http://arxiv.org/abs/2406.18045v2
http://arxiv.org/pdf/2406.18045v2
PharmaGPT: Domain-Specific Large Language Models for Bio-Pharmaceutical and Chemistry
2024-06-26T03:43:09
2024-07-03T12:56:40
Linqing Chen, Weilei Wang, Zilong Bai, Peng Xu, Yan Fang, Jie Fang, Wentao Wu, Lizhi Zhou, Ruiji Zhang, Yubin Xia, Chaobo Xu, Ran Hu, Licong Xu, Qijun Cai, Haoran Hua, Jing Sun, Jin Liu, Tian Qiu, Haowen Liu, Meng Hu, Xiuwen Li, Fei Gao, Yufu Wang, Lin Tie, Chaochao Wang, Jianping Lu, Cheng Sun, Yixin Wang, Shengjie Yang, Yuancheng Li, Lu Jin, Lisha Zhang, Fu Bian, Zhongkai Ye, Lidong Pei, Changyang Tu
cs.CL
cs.CL, cs.AI
Large language models (LLMs) have revolutionized Natural Language Processing (NLP) by by minimizing the need for complex feature engineering. However, the application of LLMs in specialized domains like biopharmaceuticals and chemistry remains largely unexplored. These fields are characterized by intricate terminologies, specialized knowledge, and a high demand for precision areas where general purpose LLMs often fall short. In this study, we introduce PharmGPT, a suite of multilingual LLMs with 13 billion and 70 billion parameters, specifically trained on a comprehensive corpus of hundreds of billions of tokens tailored to the Bio-Pharmaceutical and Chemical sectors. Our evaluation shows that PharmGPT matches or surpasses existing general models on key benchmarks, such as NAPLEX, demonstrating its exceptional capability in domain-specific tasks. This advancement establishes a new benchmark for LLMs in the Bio-Pharmaceutical and Chemical fields, addressing the existing gap in specialized language modeling. Furthermore, this suggests a promising path for enhanced research and development in these specialized areas, paving the way for more precise and effective applications of NLP in specialized domains.
PharmaGPT: Domain-Specific Large Language Models for Bio-Pharmaceutical and Chemistry: Large language models (LLMs) have revolutionized Natural Language Processing (NLP) by by minimizing the need for complex feature engineering. However, the application of LLMs in specialized domains like biopharmaceuticals and chemistry remains largely unexplored. These fields are characterized by intricate terminologies, specialized knowledge, and a high demand for precision areas where general purpose LLMs often fall short. In this study, we introduce PharmGPT, a suite of multilingual LLMs with 13 billion and 70 billion parameters, specifically trained on a comprehensive corpus of hundreds of billions of tokens tailored to the Bio-Pharmaceutical and Chemical sectors. Our evaluation shows that PharmGPT matches or surpasses existing general models on key benchmarks, such as NAPLEX, demonstrating its exceptional capability in domain-specific tasks. This advancement establishes a new benchmark for LLMs in the Bio-Pharmaceutical and Chemical fields, addressing the existing gap in specialized language modeling. Furthermore, this suggests a promising path for enhanced research and development in these specialized areas, paving the way for more precise and effective applications of NLP in specialized domains.
[ 0.0533000454, 0.0637567416, 0.0148267969, 0.0173885655, -0.0077602072, 0.0143560246, 0.0185505208, 0.0100845313, 0.0359651819, -0.0522797145, -0.0296317153, -0.0319062658, -0.025775969, 0.0533350259, 0.0476140082, 0.0209055021, 0.0231551155, -0.0203388855, 0.0209808182, -0.0010606097, 0.0080649378, 0.022892002, 0.0100679304, 0.0404991619, 0.0198046695, -0.0263517983, 0.0570245758, -0.0271059442, 0.0055126133, -0.0259890836, 0.008082157, 0.0266407505, -0.0033656, 0.0478293821, 0.0000016884, -0.0641190559, -0.0319934338, 0.0379781872, -0.0166086033, 0.030242078, 0.0731461719, -0.0297371335, -0.0039306842, 0.0096971532, -0.0180448554, -0.0019514775, 0.0482781529, 0.052195549, 0.0110171027, 0.0583287328, -0.0141464202, -0.0499699153, 0.0323910266, -0.0447415933, -0.0110764001, -0.0408037789, 0.0615242831, -0.0279610455, 0.050107047, 0.0020748624, -0.006022301, 0.035007637, -0.0467858836, -0.0024677957, 0.0057060048, 0.0230910555, -0.0344184525, -0.0488483906, 0.0015861008, 0.0751559511, -0.054803919, -0.0256157052, -0.0144041218, 0.0045687929, -0.0331019387, 0.0009562045, -0.0245612189, 0.0099777579, -0.002303557, 0.0440900922, 0.0494965017, -0.0059053656, 0.0292244032, 0.0209870208, -0.0211067535, 0.0516068451, 0.0166514106, 0.0074004014, 0.016788289, -0.0222048853, -0.0626330376, -0.0122754648, 0.0404879339, -0.0124537945, -0.0146423196, 0.0248098802, 0.0233371109, -0.0301836841, 0.082270585, -0.0817145407, -0.0151214795, -0.0113600604, -0.0183427073, 0.0687438771, -0.0420821346, 0.0158924405, -0.0345618427, 0.0014232191, -0.037285842, -0.0309645664, -0.0588158034, -0.0042270478, -0.0248658638, 0.0216904655, -0.0234490521, -0.0339196771, -0.0687198192, -0.0104511939, -0.0110798087, -0.0112573393, 0.0184313916, 0.0488696657, 0.0099133458, -0.0228324682, 0.0079976618, 0.0489706285, 0.0113578811, 0.0053266115, -0.0498282462, -0.0630422086, -0.001360196, 0.0043739369, -0.0051742201, -0.0300560873, 0.0340783782, 0.1057685539, 0.017756382, -0.0454989336, -0.0093470532, -0.0327548198, -0.0164567474, -0.0526206754, -0.0305491369, -0.0635127947, 0.0681107938, 0.0030693007, 0.0286685079, -0.0616470687, 0.0111346925, -0.0077983895, -0.0590733476, 0.0888145566, 0.0175220277, -0.0348806381, 0.0084258197, -0.0104803722, 0.0075082933, 0.0143898614, 0.0076414426, -0.0088811349, 0.0268323217, -0.0094442442, 0.007990283, 0.0197206251, -0.0212962534, -0.02857605, 0.0040332489, 0.0144366091, 0.0307147205, 0.0528669395, -0.0042771786, 0.0315591171, 0.0005448615, -0.0289375596, 0.0482061394, 0.0813108757, 0.0341145247, 0.0224778242, 0.0218298901, 0.0027116742, 0.0021851663, -0.0811948925, -0.0200406816, -0.0322995409, -0.0087959701, 0.0207785424, 0.0321726799, 0.0376915485, -0.0005607384, 0.000664276, -0.0209363736, -0.0223428067, -0.0166852269, -0.023247242, 0.0615910888, -0.0392524935, -0.0648544058, 0.0692444369, 0.0563609861, -0.0231791995, 0.0062800543, -0.0257623252, 0.0084199728, 0.0715224668, 0.0465174019, -0.0504111052, -0.0206424799, -0.0156718977, -0.022689553, 0.0285737775, 0.0161344949, -0.0079892436, 0.0261641294, -0.0267710574, 0.0186605118, -0.0207699947, 0.0201934483, 0.0097320704, -0.01353506, 0.0374616832, -0.0153814787, -0.0231407434, -0.0212000236, 0.0206353292, -0.0449650586, -0.0374144316, 0.0215503331, -0.0216422454, 0.0159647148, 0.0442698896, -0.0108467825, 0.0316258147, 0.0433780886, -0.0887861028, -0.0102414927, -0.0223650262, 0.0431223586, -0.0126173645, 0.0338470004, -0.0132355848, 0.0173342004, -0.0236427058, -0.0427653641, -0.0223322455, 0.0235058293, 0.0041093328, 0.0482657738, -0.0201589689, 0.0253769662, -0.0209810473, 0.0602964014, -0.0442674048, 0.0038090684, 0.0182445515, 0.0122153247, 0.0135497237, -0.0262040496, 0.008501173, -0.0155838691, -0.0064373803, 0.0378714688, 0.0165940933, 0.0193672124, 0.0143030081, -0.0195062496, -0.0157932062, -0.0056899958, -0.0550262667, 0.0100023607, 0.0069687599, 0.0254476406, 0.0045094774, 0.0055892933, 0.0007192935, -0.007169723, -0.0216001868, -0.0462935828, -0.0180414375, 0.0100671463, 0.0074997037, 0.0158791952, -0.0308590252, -0.0428575054, -0.0187558942, -0.0224899612, 0.0675531104, -0.0210154317, -0.04748559, -0.0239090491, -0.0141334692, -0.0099984817, 0.0230809171, 0.0085162995, -0.0009836826, -0.0271513201, -0.0080890637, 0.003656263, 0.0178687442, -0.0137777328, -0.0017575601, 0.0491002202, 0.0273959264, -0.0093764877, 0.0002620658, -0.040155407, -0.0131097278, -0.0103146462, -0.0106229447, 0.0038854366, 0.0782349333, -0.0167586375, -0.0119927032, -0.0301328003, 0.0052825576, -0.0286125522, -0.0089866947, -0.0282182489, 0.0214395914, 0.0054129749, -0.0013375523, 0.0086500701, -0.0473196357, 0.068661347, -0.0479100309, 0.0224273894, -0.0767860115, -0.017795315, -0.0337710865, -0.0038730453, 0.020886723, 0.0238316152, 0.0213440545, -0.020361172, 0.0413615778, 0.0189129617, -0.0222966224, -0.027853271, 0.0777961388, -0.0386676192, 0.0418715626, -0.0010142785, -0.0267331433, 0.0006181483, -0.0211820938, -0.0410625525, 0.0339380018, 0.057312157, -0.0361973159, -0.027620906, 0.0482665412, -0.0216703247, 0.0229817089, 0.0166726839, 0.0075056064, -0.0116898082, -0.0034551439, 0.0639591515, -0.0282330718, 0.0462882929, 0.0062460769, 0.0618763864, 0.0544636808, 0.0156072313, -0.0147220129, -0.0012234937, 0.0089197932, 0.0175111014, -0.0578568913, -0.0224218443, -0.015366381, 0.0177937225, -0.0332039669, 0.0043543088, 0.1262848377, 0.0289800987, -0.0403907597, 0.0425003543, 0.0187501833, -0.0386038013, -0.0036675697, 0.0402914807, -0.0730055571, -0.0420329757, -0.0284925755, -0.0199116059, -0.0048113093, -0.0383296795, -0.0229694862, -0.0652563199, 0.0339675061, 0.0228844732, -0.0331008397, -0.014101347, -0.0708904043, -0.0344868824, -0.0594579764, 0.0381394699, -0.0074335178, -0.0201027784, -0.0308175236, 0.0088239489, -0.0408042073, -0.0840475112, 0.0476350151, -0.0367333069, 0.0112172514, 0.03474712, -0.0029500297, 0.0019893977, 0.0137789492, 0.0214467198, -0.0144316088, 0.0024809013, -0.0069281631, 0.0315231159, 0.0129687563, -0.0316008441, -0.0164409075, 0.029742144, -0.0914975628, -0.0121208942, -0.0553041659, 0.1015735716, -0.0000657718, -0.0017945973, -0.0389244482, -0.0205199756, -0.0212829039, 0.0490186512, -0.0175349507, -0.0452001058, -0.0643102601, -0.0135004893, -0.019581832, -0.0183586162, -0.0303251967, 0.0072918753, -0.0410313383, -0.0225269962, -0.0618012547, -0.014805112, 0.073765263, -0.0214280989, -0.0043321364, 0.0624526516, 0.0305073019, -0.035710793, -0.0285565965, 0.0630750135, 0.0240730271, 0.0355759189, -0.0705747083, -0.0264960434, 0.0766017511, 0.0162791144, -0.0291610472, 0.0415863357, -0.0774333552, -0.0993722081, -0.0258422308, 0.0033741663, 0.0160403997, 0.0166400466, -0.038591627, -0.0603182539, 0.0685464814, 0.0328080207, -0.0166225266, -0.0377033539, -0.0207141079, -0.002279385, -0.0151822437, 0.0216133669, -0.0289768279, 0.0163435079, -0.0054084035, 0.0550414361, 0.0119225951, -0.018229194, 0.0443584323, 0.0086478125, 0.045841638, -0.0191277713, 0.0137104457, -0.0480124168, 0.0030953309, -0.0071653323, 0.0228172354, -0.0208356362, -0.0072160373, 0.0582541227, -0.0424353108, -0.03066843, 0.0094065312, -0.0425191782, -0.0016105596, 0.0052316207, 0.0129381744, -0.021366762, -0.0254651569, 0.0756838024, -0.0160326827, 0.0002318453, 0.0181194637, -0.0489425734, 0.0014219944, -0.0549234748, 0.0358997434, -0.0307097211, 0.0300592426, -0.0514362529, 0.0465489626, -0.0068813493, -0.0647783652, -0.0387364551, 0.0467041321, 0.0760907009, -0.051777415, 0.074428156, -0.0200116001, -0.0161803793, -0.0585641786, 0.0324077122, -0.0097971251, 0.0309035759, 0.0336023718, 0.0293846559, -0.0379643291, -0.0182116199, 0.1054576337, 0.0538699999, 0.032972578, 0.0116191097, -0.0403406471, -0.0044668838, -0.032507658, -0.0011688436, -0.0076334262, -0.0718080401, 0.0363933407, -0.0427227318, -0.0520937778, 0.0118156616, -0.0231052972, -0.0146836555, 0.0053197038, 0.00238044, 0.0682124421, 0.0264448263, -0.0366717242, -0.0702323839, 0.004785121, 0.0222895946, -0.0467727855, 0.0008445014, -5.658909885e-33, 0.0580079779, -0.0062807784, -0.0003662176, -0.0228711963, 0.0230313204, 0.0101466132, -0.0166470427, -0.0092562484, 0.0050703399, 0.0469155163, 0.0074856197, -0.0146041866, -0.00302443, 0.0136989886, 0.0180845372, 0.0120427841, 0.0248025171, 0.0338709019, 0.0311363675, 0.0075409408, 0.0410988778, 0.0615555793, 0.0693324953, -0.02535519, 0.0185165554, 0.0188142117, 0.0330930836, -0.0070906687, -0.0156798773, 0.0678974614, -0.069851853, 0.0755791217, 0.0176097024, -0.0577572361, -0.0160360578, -0.0199283939, -0.0438903309, 0.0155843953, 0.0231538862, -0.0069873775, 0.0065649012, -0.030584285, 0.0518063195, 0.0031774912, -0.0166761521, 0.0165724494, 0.0233676229, -0.0289774071, -0.0064027705, -0.0307022855, -0.0804363117, 0.0353000872, 0.0085789924, 0.0219371524, 0.0350347906, 0.0132671809, -0.0220961664, 0.0766152069, -0.0248834733, 0.0064688236, 0.0065307701, 0.0410146266, 0.0615042523, -0.0091140494, 0.0000137273, 0.0533089675, -0.0071695526, 0.0272803232, 0.0296964534, 0.006423743, -0.0184455495, 0.057162676, 0.0130063556, 0.0193269644, 0.0381592289, -0.0001300349, -0.0874309838, 0.0060609397, -0.0434818529, 0.0383544005, 0.0096327253, 0.0188388601, 0.0131035848, -0.0122426804, -0.0187558159, 0.0626569539, -0.0323601514, -0.0045266002, 0.0027963738, -0.012929, -0.0159507375, 0.0275374558, 0.0115825012, -0.0565639287, -0.0510573797, 0.0006954117, 0.0176391471, 0.0193409529, -0.0069677504, 0.0430670157, -0.0797549039, -0.0284975879, -0.0241498351, -0.057280954, -0.0081874495, -0.0276835505, 0.0127117662, -0.0170679316, -0.0421967767, 0.0077064415, -0.0428768881, -0.0142448628, 0.0192264859, 0.0652018115, 0.0188566949, 0.0373275392, 0.0155948661, 0.0761190206, 0.0330483019, -0.0252658296, 0.0184144974, 0.0154222874, 0.0431045257, 0.0194482356, -0.0513381958, 0.0080965012, -0.0368157253, 0.0215095524, 0.0294461958, -0.0544201359, -0.0204210021, -0.0315249227, 2.494e-7, 0.083843939, 0.0450035743, 0.0078931302, -0.007840991, 0.0252021272, 0.0280044712, -0.0160249975, 0.0539807342, -0.0527849756, -0.002988823, 0.0007894273, -0.0260438025, 0.0228709504, 0.0025078717, -0.0678271726, 0.0142511623, -0.0525441207, -0.0339663774, -0.0871812552, -0.019615652, 0.0157886855, 0.1341020167, -0.0111749433, -0.0213881712, 0.0215875767, -0.05032745, 0.056712348, -0.0456736088, -0.0287042446, 0.0671277866, 0.0086272471, -0.0040346631, 0.0047602612, 0.0182364527, -0.0210462864, 0.009641883, 0.0151530551, 0.0826358274, -0.0552659929, -0.0114290528, 0.0394033268, -0.0465563461, 0.0116024008, -0.040639475, 0.0213774648, -0.0212714467, -0.0497337878, -0.0672870278, 0.0069799111, 0.0065555866, 0.0624057613, -0.0404809788, 0.0175234638, 0.0380557179, 0.0143054752, 0.0009002985, -0.0167638678, -0.0750649199, -0.0199472699, 0.0350898653, -0.042787455, 0.0078226188, -0.0113416323, -0.0123068513, 0.040072754, -0.0459185205, 0.0210119411, 2.196560886e-34, 0.062388096, -0.0202588402, 0.0244281106, 0.0418811813, 0.0544824675, -0.0147057278, -0.0210987274, 0.0245484058, -0.0315809324, -0.1108609512, 0.0005864941 ]
http://arxiv.org/abs/2406.18034v1
http://arxiv.org/pdf/2406.18034v1
LLMs for Doctors: Leveraging Medical LLMs to Assist Doctors, Not Replace Them
2024-06-26T03:08:24
2024-06-26T03:08:24
Wenya Xie, Qingying Xiao, Yu Zheng, Xidong Wang, Junying Chen, Ke Ji, Anningzhe Gao, Xiang Wan, Feng Jiang, Benyou Wang
cs.CL
cs.CL
The recent success of Large Language Models (LLMs) has had a significant impact on the healthcare field, providing patients with medical advice, diagnostic information, and more. However, due to a lack of professional medical knowledge, patients are easily misled by generated erroneous information from LLMs, which may result in serious medical problems. To address this issue, we focus on tuning the LLMs to be medical assistants who collaborate with more experienced doctors. We first conduct a two-stage survey by inspiration-feedback to gain a broad understanding of the real needs of doctors for medical assistants. Based on this, we construct a Chinese medical dataset called DoctorFLAN to support the entire workflow of doctors, which includes 92K Q\&A samples from 22 tasks and 27 specialists. Moreover, we evaluate LLMs in doctor-oriented scenarios by constructing the DoctorFLAN-\textit{test} containing 550 single-turn Q\&A and DotaBench containing 74 multi-turn conversations. The evaluation results indicate that being a medical assistant still poses challenges for existing open-source models, but DoctorFLAN can help them significantly. It demonstrates that the doctor-oriented dataset and benchmarks we construct can complement existing patient-oriented work and better promote medical LLMs research.
LLMs for Doctors: Leveraging Medical LLMs to Assist Doctors, Not Replace Them: The recent success of Large Language Models (LLMs) has had a significant impact on the healthcare field, providing patients with medical advice, diagnostic information, and more. However, due to a lack of professional medical knowledge, patients are easily misled by generated erroneous information from LLMs, which may result in serious medical problems. To address this issue, we focus on tuning the LLMs to be medical assistants who collaborate with more experienced doctors. We first conduct a two-stage survey by inspiration-feedback to gain a broad understanding of the real needs of doctors for medical assistants. Based on this, we construct a Chinese medical dataset called DoctorFLAN to support the entire workflow of doctors, which includes 92K Q\&A samples from 22 tasks and 27 specialists. Moreover, we evaluate LLMs in doctor-oriented scenarios by constructing the DoctorFLAN-\textit{test} containing 550 single-turn Q\&A and DotaBench containing 74 multi-turn conversations. The evaluation results indicate that being a medical assistant still poses challenges for existing open-source models, but DoctorFLAN can help them significantly. It demonstrates that the doctor-oriented dataset and benchmarks we construct can complement existing patient-oriented work and better promote medical LLMs research.
[ 0.0764166489, 0.0348533057, -0.014561424, -0.0036556849, -0.0175022949, 0.0013680938, 0.0027785318, 0.0332120582, 0.0046411804, -0.0032452438, -0.0254424997, -0.0400383659, 0.0033915513, 0.0166074149, 0.0363069177, -0.0605636425, 0.0184955169, 0.0073048109, 0.0340454429, -0.0083721383, 0.0150448876, 0.0139170308, -0.0523268543, 0.0651104972, 0.0362131484, -0.0154415676, 0.0166312456, -0.0017986866, 0.0274242032, -0.0327280611, 0.0018316923, 0.0296324119, -0.0090803616, 0.0011635042, 0.0000021077, -0.0417138413, 0.0195560362, 0.0103661083, -0.0660271943, -0.03405945, 0.0175295062, -0.0276273601, -0.0077259066, 0.0371164866, -0.039674826, -0.0239432771, 0.0075262175, 0.040091332, 0.057156872, 0.0715801567, -0.0047824476, 0.0362756625, 0.0017085067, -0.0246590823, 0.0334928297, -0.0637372956, 0.0183314886, 0.0102587892, 0.0516390279, -0.0096413381, 0.0285557564, 0.0233078729, -0.0214122832, -0.0108238272, -0.0084766354, 0.0440654643, -0.0099856136, -0.0028834261, -0.010418077, 0.0726929307, 0.015439501, -0.0006595133, 0.0106900977, 0.0067168502, 0.0084731989, -0.0060934178, -0.053057272, 0.0131687857, -0.0186731443, -0.02027441, -0.0025893294, -0.0131207099, 0.0350125283, 0.0059636757, 0.0111838924, -0.0186587125, -0.0059588496, -0.0232192371, -0.0135436272, -0.0182038397, 0.0280397162, -0.0398632437, 0.0473564342, 0.0516623668, -0.0844675004, 0.0174062867, -0.0119544221, -0.0160696618, 0.0800097883, -0.0293684937, 0.0414712168, 0.0235084016, -0.0058355494, 0.0396827832, 0.0130453398, -0.0072324988, -0.0198685452, 0.0296126511, -0.0446090959, -0.0195138324, -0.052466929, 0.0021578956, -0.0194220506, 0.0084789721, 0.005666934, 0.0129095744, -0.0579666235, -0.0061485516, -0.0137720173, 0.0525842868, 0.0298407339, 0.0413716361, -0.03736002, -0.0035920218, -0.0193843897, 0.0226630028, 0.004435949, 0.040040005, 0.0073686466, -0.0491235144, 0.0177476779, 0.0044882945, 0.0035645256, 0.0173940957, 0.0017417643, 0.0540334173, 0.0019000525, 0.0029608537, 0.0025571533, -0.0469988137, 0.0007809942, -0.0079678418, 0.0130810197, -0.0608962476, 0.030745659, 0.0106133912, 0.0135978451, -0.0299940165, 0.0277764723, 0.0344958864, -0.0399784967, 0.0531324632, -0.0512754656, 0.0523302369, 0.0381476693, 0.0630877763, 0.0006556032, 0.0306102987, 0.0281461347, -0.0381723866, 0.03132727, -0.0151131721, 0.0403568968, 0.0279537421, 0.0081431894, -0.0032655308, 0.0297873598, 0.0326993279, -0.0075302268, 0.0800736621, -0.0289188512, 0.0518180765, -0.0144049693, 0.0058286488, 0.0334452875, 0.0527064428, 0.0135049541, -0.0071864715, -0.0000195499, -0.026519997, 0.0033143573, -0.0361887813, -0.0134430174, 0.0132437041, -0.0081391856, -0.0210935697, -0.0367255025, -0.0086709335, -0.0618598759, -0.0401751883, -0.0501381196, -0.024019653, -0.0588341318, 0.0029114077, 0.0216454435, -0.0205489956, -0.0141949272, 0.0836039409, 0.0257526338, -0.0799262971, -0.033086028, -0.0631944165, -0.0147060221, 0.0984209329, 0.0660306737, -0.0533739701, -0.0475180186, -0.0215385277, 0.0166929532, 0.0277168583, 0.0642613769, -0.0259238575, -0.0000964573, -0.0544879995, -0.0089722434, 0.0341092274, 0.0045019616, -0.0307350289, -0.004706176, 0.0445650741, -0.0116692921, -0.0409877487, -0.0502924025, 0.0251687746, -0.0947713181, -0.0077038542, 0.0158929769, 0.0003327166, 0.0305782445, 0.0553052202, -0.0060196514, 0.0437339917, 0.0260801874, -0.0302072838, -0.0072723865, 0.0393323377, -0.0333777852, -0.0046911207, 0.0232658759, -0.0370799452, -0.0067592473, -0.0004187068, -0.0299014375, -0.0439816266, -0.0419984087, -0.0575024933, 0.0403067023, 0.0052854689, 0.0274120197, 0.0469184965, -0.0041874489, 0.0057580369, -0.0147257354, -0.033722952, 0.0080977343, 0.0133562647, 0.0186744649, 0.0040914542, -0.0556159504, -0.0051435465, 0.0097607113, 0.0110437199, 0.0323665254, -0.0222258233, 0.0271722265, -0.0352626964, -0.0475250632, -0.1230032295, -0.0304199774, -0.0016146749, -0.0049172295, 0.0379596278, 0.0311103296, -0.0051438883, -0.0224948619, -0.0739469603, 0.0237631369, 0.0312887877, -0.0044619231, 0.0248072241, 0.0244783107, 0.0127436165, -0.0415204801, 0.0129271848, -0.0364156924, -0.0346296728, 0.0107115591, -0.0219988097, -0.0328946039, -0.0143227242, -0.0420607999, -0.0098762326, -0.0342893712, -0.0470678285, 0.0174610075, -0.0105444891, 0.0068549858, 0.0148339821, -0.0065742275, -0.0067915954, 0.0423794799, -0.0041424832, -0.0174991339, 0.0107073532, 0.0163369291, -0.045030117, 0.0081331292, -0.069232814, -0.0128294416, 0.0546701662, 0.0126668308, 0.0236966498, -0.0090099275, -0.0114904977, -0.0187212303, 0.0275367703, -0.0167135969, 0.0341026112, -0.0341242403, 0.0535349883, 0.0041002296, -0.0197970588, 0.0715936348, -0.0127559453, 0.0149734067, -0.0655428618, -0.0491034538, -0.0546989143, 0.0114555005, -0.0571460612, 0.0092024961, -0.0018743338, -0.0214577839, -0.0052859187, 0.0375259332, -0.0360161886, -0.0816567615, 0.0594394654, -0.0360821486, 0.0045260135, -0.0203189291, -0.0312804766, -0.0567494929, -0.0141984383, -0.0611874759, 0.0179338548, 0.0160102583, -0.0002674717, -0.0624945536, -0.0014602947, 0.0029419521, 0.0856752992, 0.0520663001, -0.0081687402, -0.0056051719, -0.0063633523, 0.0265338831, 0.0254777875, -0.0130229173, -0.0167740993, 0.0577223115, 0.0731582344, 0.046254389, -0.0052446262, 0.0263668839, -0.0466783531, 0.0277336016, -0.0201250426, -0.0132731451, 0.0287773926, 0.0444096178, -0.052922599, -0.0148986485, 0.1106747314, 0.0216558054, -0.0793101639, 0.0300103184, -0.0512533076, -0.0414825641, 0.0063773589, 0.0100817541, -0.1006983519, -0.027366003, -0.0252382327, 0.0446111932, 0.004929001, -0.0096663637, 0.0306525026, -0.0814812705, 0.0505570769, 0.003959517, -0.0250394437, -0.0268464535, -0.0267961174, 0.0144487424, 0.0090726065, 0.0279773567, -0.0096491147, -0.0099781891, -0.0267823283, -0.0019799918, -0.0485643037, -0.058950901, 0.058390554, -0.0107378531, 0.059880197, 0.076308392, 0.0379002579, -0.0124018472, -0.0577204376, -0.0360315405, 0.0215803981, 0.0163665358, 0.0092513189, 0.0733918548, 0.0046366043, 0.0039264476, -0.0134105086, 0.0404140018, -0.0910484418, 0.0051643657, 0.0631754026, 0.0568873696, -0.0367255621, -0.020193344, -0.0442706533, 0.0516888388, 0.0488550849, 0.052654516, 0.0013129571, -0.0112137077, -0.0098302057, 0.0062644854, 0.0390969142, 0.0126442648, -0.0189966038, 0.0283382908, -0.0126000959, 0.0007236854, -0.0712913796, -0.0377941765, 0.0450056978, -0.011898078, 0.0192106068, 0.0312430002, 0.0179569181, 0.0143756308, -0.0320004746, 0.0025896588, 0.00118905, 0.052299004, -0.0786261335, -0.0602341779, 0.0591196045, 0.0071328757, 0.0272201654, -0.0072760684, -0.0279313419, -0.0303815305, -0.0092651779, 0.0555248484, 0.0211208817, 0.0059006969, -0.0197159853, -0.03772166, -0.0032700964, 0.0383548476, -0.0574430302, -0.0057553309, 0.0124256415, 0.0286767464, 0.0006926642, 0.0195874758, -0.0352801383, -0.0119256778, 0.010187326, 0.0267092045, -0.0533158742, -0.0738963038, 0.0060666525, 0.0497751608, 0.0204402152, 0.0064983116, 0.0159806963, -0.0218859855, 0.0511514135, -0.0450602584, 0.0009141243, -0.0437647812, 0.0015927262, -0.010554784, -0.055558458, 0.0395839624, 0.0134095447, -0.0824395418, -0.0485863015, -0.0677738786, 0.0391097143, -0.0001691999, -0.0076928148, 0.0304602254, -0.0248158891, -0.0057742656, -0.0027887686, 0.0154191516, 0.0157838371, 0.0286283605, 0.0878581554, -0.0163049474, 0.0139779644, 0.0096732182, 0.0467827469, -0.0408348739, -0.106341064, -0.0029390561, 0.0614493564, 0.054852888, 0.0106342109, 0.0353042483, -0.0440764129, 0.0320786163, -0.0285511091, 0.0261438601, -0.0379095636, 0.0133605842, 0.0677028149, 0.0119306948, 0.0134481164, -0.0120304199, 0.1151314601, -0.0367123336, 0.051554475, 0.0038892755, -0.0229642317, -0.0547059551, -0.02427819, 0.0545672961, 0.0269738249, -0.0632098168, 0.0115776975, -0.0607560612, -0.051455345, -0.0105532687, -0.0188687295, -0.0268455055, -0.0355762243, 0.0490253828, 0.0421775356, 0.0135217328, 0.0038889074, -0.037025068, 0.0510671884, -0.0379551984, 0.0019541446, 0.0316462703, -5.945213657e-33, 0.0114410389, -0.0413684025, 0.0411998816, -0.0149561372, 0.0189690329, 0.0278678667, -0.0327018127, 0.0158672854, -0.0342442282, 0.0220233276, -0.0513167493, 0.0232047066, 0.0129354633, 0.0301903673, 0.0207555182, -0.012463497, 0.033741504, 0.0149688609, -0.0000087301, 0.046988856, -0.0120521327, 0.0314946771, 0.0317717083, -0.0921682939, 0.0139847538, -0.0289775282, 0.0218591485, 0.0261011645, -0.0464487001, 0.0564671867, -0.0083406959, 0.0498864353, 0.0464035384, 0.0152982688, -0.0158436224, 0.0029859152, -0.0340729244, -0.0153008541, 0.0132708987, -0.0066341977, -0.0577764176, -0.0189026389, 0.0155613702, -0.0194830317, -0.0062109563, -0.0003327789, 0.0276460126, -0.0413164161, -0.0178080089, 0.0060310406, -0.0053548184, 0.0269434042, -0.0039211055, -0.0345959365, -0.0500401929, -0.0077448706, -0.0336565599, -0.0040696426, -0.0508581586, 0.0121591548, 0.0347165912, 0.0258824471, 0.0214940589, -0.0343494155, -0.0197285842, 0.0239361767, 0.0177583657, 0.0139936479, -0.0434639119, 0.0102094077, -0.0654944479, 0.0628412068, 0.0420072041, 0.0372231789, -0.0055739866, -0.0335079879, -0.0580110997, 0.0392053314, -0.015058266, 0.0679908544, 0.0002631136, 0.0114991609, -0.0607640073, -0.0050637769, -0.0067757326, 0.0028289971, -0.0194634851, -0.025284186, 0.0215450712, 0.005879364, 0.0139254574, -0.0168502219, -0.0236807559, -0.0163808577, 0.0319178067, -0.0245300978, 0.0218176171, 0.0003471643, -0.0418360308, -0.0174811278, -0.1086352244, -0.0038166957, -0.0103065856, 0.0036055113, 0.0374836065, 0.0504550673, 0.0124913007, -0.0283984933, -0.0434470475, -0.0055000451, 0.005051008, -0.0066888784, 0.0024271605, 0.0062775919, 0.0320422091, -0.0148641756, -0.0102439811, 0.0687406287, -0.0047488734, -0.0248301011, -0.0634227321, -0.0698725581, -0.0042611477, -0.0155406296, -0.083901614, 0.0125752157, -0.0494654924, 0.0527174622, 0.0644052252, -0.0143906428, -0.0072329873, -0.0301136337, 2.863e-7, 0.0485923924, 0.0323781148, 0.0099794772, 0.0148678925, 0.0303370282, 0.0137369521, -0.0188322794, 0.022864813, -0.0117841633, 0.0033095775, 0.0302400198, -0.0126740159, 0.0564955063, 0.0409160256, -0.078242287, -0.0268615447, -0.0902081132, -0.0122215888, -0.0092764264, -0.0419278368, 0.029362319, 0.0611629561, -0.0317819342, -0.0341462232, -0.0035634195, -0.0120960902, 0.042565614, -0.0400753021, -0.0019522618, 0.0639704615, 0.0826055706, -0.0028739802, 0.0153369699, 0.032792829, -0.0335329846, -0.0164211858, 0.018640317, 0.1219064742, -0.0712663829, 0.0118308533, 0.0349319726, -0.0075093023, 0.0178498719, -0.0409677178, 0.0235241596, -0.0191875845, -0.0557402968, 0.0037679092, 0.0050900448, -0.0281865783, 0.0528959371, -0.0190916676, -0.0309046991, 0.0311582536, 0.0124896253, 0.0258491058, -0.0467517525, -0.0463942289, -0.020109728, -0.0002086685, -0.0238048583, -0.058260601, 0.0479175486, -0.0411975868, -0.0470971726, -0.064846307, 0.0080445716, 2.025830204e-34, 0.0023483755, 0.0242131799, 0.0037393447, 0.0663161352, 0.0146375652, -0.0118015604, 0.0355164781, 0.0392734408, 0.0384304933, -0.0182089061, -0.0374385752 ]
http://arxiv.org/abs/2406.18027v1
http://arxiv.org/pdf/2406.18027v1
Automated Clinical Data Extraction with Knowledge Conditioned LLMs
2024-06-26T02:49:28
2024-06-26T02:49:28
Diya Li, Asim Kadav, Aijing Gao, Rui Li, Richard Bourgon
cs.CL
cs.CL, cs.AI
The extraction of lung lesion information from clinical and medical imaging reports is crucial for research on and clinical care of lung-related diseases. Large language models (LLMs) can be effective at interpreting unstructured text in reports, but they often hallucinate due to a lack of domain-specific knowledge, leading to reduced accuracy and posing challenges for use in clinical settings. To address this, we propose a novel framework that aligns generated internal knowledge with external knowledge through in-context learning (ICL). Our framework employs a retriever to identify relevant units of internal or external knowledge and a grader to evaluate the truthfulness and helpfulness of the retrieved internal-knowledge rules, to align and update the knowledge bases. Our knowledge-conditioned approach also improves the accuracy and reliability of LLM outputs by addressing the extraction task in two stages: (i) lung lesion finding detection and primary structured field parsing, followed by (ii) further parsing of lesion description text into additional structured fields. Experiments with expert-curated test datasets demonstrate that this ICL approach can increase the F1 score for key fields (lesion size, margin and solidity) by an average of 12.9% over existing ICL methods.
Automated Clinical Data Extraction with Knowledge Conditioned LLMs: The extraction of lung lesion information from clinical and medical imaging reports is crucial for research on and clinical care of lung-related diseases. Large language models (LLMs) can be effective at interpreting unstructured text in reports, but they often hallucinate due to a lack of domain-specific knowledge, leading to reduced accuracy and posing challenges for use in clinical settings. To address this, we propose a novel framework that aligns generated internal knowledge with external knowledge through in-context learning (ICL). Our framework employs a retriever to identify relevant units of internal or external knowledge and a grader to evaluate the truthfulness and helpfulness of the retrieved internal-knowledge rules, to align and update the knowledge bases. Our knowledge-conditioned approach also improves the accuracy and reliability of LLM outputs by addressing the extraction task in two stages: (i) lung lesion finding detection and primary structured field parsing, followed by (ii) further parsing of lesion description text into additional structured fields. Experiments with expert-curated test datasets demonstrate that this ICL approach can increase the F1 score for key fields (lesion size, margin and solidity) by an average of 12.9% over existing ICL methods.
[ 0.0225942433, 0.0497610867, 0.0245044399, 0.0310858339, -0.0132240411, 0.0142612057, -0.0454547592, 0.0660628825, 0.0002679226, -0.021308979, -0.0120373927, -0.0165618695, 0.0027075452, 0.0370211005, -0.0169721618, -0.032580696, 0.0128476489, 0.0112122903, 0.0638020188, 0.0390217751, -0.0096191615, -0.0024553959, 0.0070632035, 0.0001520275, 0.0082673412, -0.0110361483, 0.0185359139, -0.0288747717, 0.0459575951, -0.0504755005, 0.0181865729, -0.0077613615, -0.0097674858, 0.0026081016, 0.000002175, -0.0527849533, -0.0601332784, 0.042678453, -0.0252267253, -0.0425550602, 0.041991286, -0.084171854, -0.0080913808, -0.0327307954, -0.052232895, -0.0004368245, 0.0120873768, 0.0750220865, 0.0589730367, 0.0663659573, -0.0339351632, 0.0133487079, 0.0209433883, 0.0142169986, 0.0034147678, -0.0344353579, 0.0541060865, 0.0465556011, -0.0076565864, -0.0232721884, 0.0389356539, 0.0626331121, -0.0202626362, 0.0038468374, 0.0206887908, 0.0579183586, -0.0125274798, 0.0031455962, 0.0035112146, 0.0416784436, -0.0410111323, 0.010635769, 0.0199135132, -0.0433964618, -0.0136826541, 0.0098770233, -0.0302004423, 0.0004835888, -0.054593645, -0.0029121616, 0.0283129122, -0.0016802684, 0.0121077113, -0.0184568577, -0.0377559625, 0.0138259595, -0.0364013389, -0.0220753159, 0.0368959829, 0.0184743237, -0.0065074572, -0.0472383052, 0.0393149033, 0.0520040914, -0.0029115095, 0.0283557307, -0.0083245663, -0.0319252051, 0.0855147317, -0.0629082173, -0.0152710564, -0.0348567888, -0.0406007841, 0.0606338866, -0.0321199186, -0.02679733, -0.025562603, 0.0449119508, -0.0207464322, 0.0304241087, -0.0190967843, -0.0317294933, -0.0010466392, 0.0397776738, -0.0158799998, 0.0254814941, -0.0474308021, -0.0059527135, -0.0266462117, 0.0584503151, 0.0444404855, -0.0141300531, -0.0509449467, 0.0255071241, -0.0232165735, 0.0252419263, 0.0243541691, 0.0021843405, -0.0171357486, -0.0050676079, 0.0150034949, 0.0184008796, 0.0327089205, 0.0005036718, 0.0142982258, 0.0442368947, 0.010944508, 0.0031528678, -0.0389206596, -0.075724043, -0.0065483418, -0.0130172595, 0.0072118118, -0.0731810778, 0.0104536256, 0.0072767311, 0.0040802117, -0.055991739, 0.0060271439, -0.0088333935, -0.0167314447, 0.0495213307, -0.0807990879, -0.0232477318, 0.0102300122, 0.0202254951, 0.0024740603, 0.0123607917, -0.0011484621, -0.044049453, -0.0040051052, 0.0102224397, 0.0306211952, 0.028124705, 0.0184652675, -0.0648218021, -0.0387274362, 0.030396454, 0.0137222512, 0.0195138305, 0.0133956587, 0.0377591029, 0.0099522462, 0.013037133, 0.0202520229, 0.0509308763, 0.0524344705, -0.0166632328, -0.0035854024, -0.0030845909, 0.0086178519, -0.012040657, 0.0221438389, -0.0099887531, 0.0180473235, 0.0357423536, -0.0385291614, 0.0101945447, 0.0163867008, -0.0668150857, -0.0300905984, 0.0010828458, -0.022014806, 0.0252619535, 0.0764745101, -0.0074569006, -0.0276266579, 0.0192323793, 0.0582145452, 0.0030882256, -0.0347215347, -0.0437183939, -0.0105958926, 0.085684754, 0.0535514094, -0.0705417544, -0.0452701226, -0.0325494409, -0.0370033272, 0.0009147461, -0.0219043586, -0.0286417324, 0.0258481205, -0.0144770062, -0.0112319477, 0.0312342402, 0.0109198997, -0.0399475358, 0.0071401526, -0.0044276607, -0.0404548608, -0.0439727642, -0.0379884429, 0.0692504272, -0.0692289323, -0.0177876502, -0.0307678692, 0.012790422, 0.0083550755, 0.0066087409, -0.0261300169, -0.0118384184, 0.0398816057, -0.0211311169, -0.0487968028, 0.0188980307, -0.0613733493, -0.0434089825, 0.0354295224, 0.0054939389, -0.0364223197, -0.0834738314, -0.0201783367, 0.0314419828, -0.0583027937, 0.0033929592, 0.0678154305, -0.0089248959, 0.014097291, 0.0490903966, -0.0248810481, -0.0590896904, 0.064691633, -0.0149750924, -0.0257059634, 0.0049036457, -0.077846922, 0.0120838135, -0.0110429591, -0.0132732093, 0.0200299174, -0.0008287239, 0.0159963835, -0.0105626527, 0.0138319703, -0.0278219264, -0.0311150607, -0.0756595284, 0.0008286308, 0.0100469226, -0.0073519833, 0.0254532099, 0.0227354188, -0.0085405828, -0.0368658192, -0.0432973877, -0.0509704351, 0.0288767982, -0.0294796601, 0.0748125538, 0.0650317743, -0.0227388423, -0.0445858017, 0.0234830528, -0.0462201647, -0.0009370383, 0.0536310449, -0.0618761294, -0.0198371336, 0.1011869609, -0.0266116001, 0.0399237871, -0.0257585086, 0.0196223091, 0.0007957304, -0.0559059493, -0.0463105403, 0.0831077397, -0.0218458306, 0.0417788588, 0.0371884257, -0.0161866099, -0.0574992485, 0.0164600592, -0.0163337104, -0.0132279461, -0.0288365334, -0.0261985641, 0.031730637, 0.0629625618, 0.008510083, 0.005998156, -0.0031502156, 0.0096855201, -0.019913638, 0.0111170448, -0.0513465405, -0.0078801997, -0.0089912461, 0.0506481603, -0.0170820467, -0.0289387479, 0.0934240669, -0.0180010535, 0.0379485488, -0.0659360811, -0.0012179686, -0.0353076793, 0.032734938, -0.0136631113, 0.0127986614, 0.023306448, -0.017793268, 0.0180853624, 0.0106069669, 0.0326875187, -0.0850237459, 0.0662030205, -0.0316764228, 0.0532136448, 0.0255890526, -0.0283265151, -0.0613189153, 0.0091902297, 0.0087081445, -0.0012888616, -0.0246293508, 0.0162132476, -0.0309326407, 0.0286591053, 0.0207422692, 0.0604283288, 0.0396121182, -0.024913514, -0.0031049701, 0.0197942667, 0.0246982966, -0.0257441401, -0.0408528708, -0.02302159, 0.0200674552, 0.0850498602, -0.0382650718, 0.0252397489, 0.0124167483, 0.0029730443, -0.0146700582, -0.0249252375, 0.017460173, 0.0138254594, 0.0480585434, -0.083060354, -0.0312044173, 0.0993877426, 0.0324798599, -0.0403491519, -0.0041860612, -0.0115045728, -0.0778540745, -0.0110993944, 0.0353079028, -0.0327317156, 0.0148295164, 0.0175765045, 0.0148747936, -0.0230728164, -0.0500461198, 0.0375856385, -0.0594202206, 0.0512029231, -0.0015626283, -0.0548811741, -0.0635730773, -0.0617523901, 0.0015320993, -0.0066930172, -0.0291066505, -0.0006100956, -0.0148541499, -0.0207142178, 0.0013606694, -0.0412870273, -0.0867006928, 0.0187102482, -0.0397851691, -0.0014000009, 0.0617659874, 0.0171797965, -0.0351239629, -0.0349482596, 0.0037042743, 0.055354964, -0.0084329369, -0.0052907462, 0.0730771124, 0.0004745358, -0.0072417366, -0.0351473205, 0.0178502388, -0.0061556087, -0.0104834335, 0.0103018666, 0.0752373934, -0.0391263962, -0.0058865044, -0.0191920046, 0.0018934993, 0.0313568674, 0.0527111962, -0.0075420761, 0.0091786673, 0.0056069978, -0.0335788019, 0.0507364087, 0.007986322, -0.0172517113, 0.0733435974, -0.0568496585, -0.0144806579, -0.0638552606, -0.0199606288, 0.0636565909, -0.0044038193, 0.0385554433, 0.0436530598, 0.0057983925, -0.0113921138, -0.1001529843, 0.036845766, 0.0220137239, 0.0011095938, -0.0178062189, -0.0335161388, 0.0329403579, 0.0016738027, -0.0191431697, 0.030339634, -0.0109241148, 0.0115466891, 0.0049515101, 0.0185172185, 0.0181895886, 0.0448562019, -0.0040330505, 0.0290694013, 0.0736767873, 0.0258807335, -0.0819035172, -0.0310353525, -0.0102033708, -0.0276767183, 0.0135170333, -0.010741693, -0.0251898654, 0.0230966583, 0.0185361691, 0.0358235613, -0.0449018702, -0.0017337115, 0.0572383255, 0.0592527986, 0.0232084077, 0.0170884337, -0.0278300345, -0.0449820235, 0.0133862831, -0.0213503744, -0.0140866563, -0.0266781133, 0.0250351131, 0.0089867096, 0.0027784309, -0.0103333211, 0.0397936888, -0.0240039639, -0.0102052633, -0.0318130367, 0.0529280193, 0.0370404832, 0.0044147759, 0.0768625811, -0.0185294021, -0.0098729953, -0.0356218889, 0.039106857, 0.0442982428, -0.001084843, 0.0301051401, 0.0200058036, 0.0429337993, -0.0389318317, 0.0112787914, -0.0107402094, 0.0395488888, -0.0640403703, -0.0162408575, 0.001828387, -0.0159685928, 0.0321141705, -0.0472935736, 0.0073706266, -0.0119852992, 0.0280326102, -0.0217284206, -0.0030962212, 0.0527009107, 0.020537708, -0.0085738432, -0.0598717779, 0.0638158247, -0.0260364898, 0.04877498, -0.0687474087, -0.0542699434, -0.0755207762, 0.0082829949, 0.0326319262, 0.005622949, -0.0558305159, 0.0021976053, -0.049260661, -0.0401020795, 0.0341249667, 0.0090274867, -0.0401219502, -0.0093027661, 0.0037345644, 0.0778468773, 0.0256007295, -0.0575319268, -0.0280748624, 0.0412344374, 0.0052543078, -0.025482187, 0.0124939969, -6.066216841e-33, -0.0177664105, -0.0560604706, 0.0477007292, 0.0071517583, 0.0131699201, 0.0433705673, -0.0290718414, -0.0395320728, 0.0149017451, 0.0219620802, -0.0303835999, -0.0235439371, 0.016369639, -0.0017331017, 0.0255263075, -0.0119239548, 0.0262336489, 0.0054201563, -0.0086072758, -0.0346633494, 0.0051799077, 0.07072898, 0.0698442757, -0.1182882488, 0.0158710256, 0.0420494415, 0.0177220758, 0.0089122271, 0.0116958302, 0.0502031669, -0.039557334, -0.0031324243, 0.0436553359, -0.0318194665, -0.0066981991, -0.020309126, -0.1149700284, 0.0041609011, -0.0048162299, 0.0110887671, -0.0158960875, -0.010439638, 0.059654586, -0.023711348, 0.0460925773, -0.0176901538, 0.0484550372, -0.041894801, 0.0038574135, -0.0533932969, -0.0548453704, 0.0758248642, 0.0154998973, 0.0320877247, 0.018976612, -0.0400328115, -0.0059685586, 0.0425951853, -0.0203414503, -0.0123907095, 0.0442539155, 0.0461411327, 0.0033290517, 0.0121906754, -0.0150231831, 0.0871463045, 0.023691982, -0.0173831265, -0.0269044004, -0.004491922, -0.002226172, 0.0483284518, -0.0344973058, 0.0121077606, 0.020485146, -0.0025932258, -0.0931464136, 0.0581110492, -0.0489929691, 0.0292887539, -0.0124399103, -0.0244497601, -0.0544007793, 0.0086519867, 0.0073097693, -0.0050115725, 0.0012656482, -0.008018082, 0.0162243489, 0.0278364625, 0.0049171033, 0.0277120005, -0.0128319338, 0.0076662125, 0.0067250505, -0.0453484766, 0.0543233715, 0.0136442678, -0.0604082383, -0.0191284604, -0.0336451419, -0.0259876009, -0.0161776915, 0.0076638493, 0.0543641374, 0.00741272, 0.0636891723, -0.0174116921, -0.0757220536, 0.0092186611, -0.0107712848, 0.0161755141, 0.0007243024, 0.0407169312, 0.0403457209, 0.031065058, -0.0230410937, 0.0348667242, 0.0039746715, -0.0099505382, -0.0305968579, -0.0146377617, -0.0037092285, -0.0106645888, -0.0432331376, -0.0370529778, 0.0022994252, -0.0123478882, 0.0636032149, -0.0343679525, 0.0163114872, 0.0073324069, 2.906e-7, 0.1124250814, 0.0457593054, 0.0353050791, -0.0190527905, 0.0160462707, -0.0049836366, -0.0363474749, -0.0306777097, 0.0010556636, -0.0012669575, 0.0339021236, -0.0039127255, 0.0006975412, 0.0619222149, -0.088620849, -0.0303605366, -0.0120637584, -0.0638985112, -0.0028840778, -0.0132024847, 0.0390669927, 0.0412706286, -0.0404958352, -0.0303074177, 0.0075552086, -0.0388538502, 0.0114976503, -0.0390422642, 0.0218345206, 0.0326948352, 0.0270593818, -0.0127810044, 0.0247254297, 0.0680279657, -0.0066062259, 0.017829638, 0.002330693, 0.1095741168, -0.0228422247, 0.0177593101, -0.0034593875, -0.0119920922, 0.0332661197, -0.0230970271, 0.0213017333, -0.0425831564, -0.0249415599, -0.0154917641, -0.0162750576, -0.019966213, -0.0050092442, -0.0252589546, -0.0171883181, 0.065323025, 0.0058378088, 0.0324113667, -0.0343262665, -0.0643806458, -0.0321075544, 0.0252255984, -0.0036224241, -0.0288286377, 0.0682998225, -0.0260247551, -0.0199671257, -0.0584218018, 0.0370240547, 2.932582669e-34, 0.0228755362, 0.0173556246, -0.0335112736, 0.0448010974, 0.0191350542, -0.0046852366, 0.0022551317, 0.0366326049, -0.0296791103, -0.0356180854, 0.0112595893 ]
http://arxiv.org/abs/2407.01599v1
http://arxiv.org/pdf/2407.01599v1
JailbreakZoo: Survey, Landscapes, and Horizons in Jailbreaking Large Language and Vision-Language Models
2024-06-26T02:20:23
2024-06-26T02:20:23
Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, Haohan Wang
cs.CL
cs.CL, cs.CR, cs.CV, cs.LG
The rapid evolution of artificial intelligence (AI) through developments in Large Language Models (LLMs) and Vision-Language Models (VLMs) has brought significant advancements across various technological domains. While these models enhance capabilities in natural language processing and visual interactive tasks, their growing adoption raises critical concerns regarding security and ethical alignment. This survey provides an extensive review of the emerging field of jailbreaking--deliberately circumventing the ethical and operational boundaries of LLMs and VLMs--and the consequent development of defense mechanisms. Our study categorizes jailbreaks into seven distinct types and elaborates on defense strategies that address these vulnerabilities. Through this comprehensive examination, we identify research gaps and propose directions for future studies to enhance the security frameworks of LLMs and VLMs. Our findings underscore the necessity for a unified perspective that integrates both jailbreak strategies and defensive solutions to foster a robust, secure, and reliable environment for the next generation of language models. More details can be found on our website: \url{https://chonghan-chen.com/llm-jailbreak-zoo-survey/}.
JailbreakZoo: Survey, Landscapes, and Horizons in Jailbreaking Large Language and Vision-Language Models: The rapid evolution of artificial intelligence (AI) through developments in Large Language Models (LLMs) and Vision-Language Models (VLMs) has brought significant advancements across various technological domains. While these models enhance capabilities in natural language processing and visual interactive tasks, their growing adoption raises critical concerns regarding security and ethical alignment. This survey provides an extensive review of the emerging field of jailbreaking--deliberately circumventing the ethical and operational boundaries of LLMs and VLMs--and the consequent development of defense mechanisms. Our study categorizes jailbreaks into seven distinct types and elaborates on defense strategies that address these vulnerabilities. Through this comprehensive examination, we identify research gaps and propose directions for future studies to enhance the security frameworks of LLMs and VLMs. Our findings underscore the necessity for a unified perspective that integrates both jailbreak strategies and defensive solutions to foster a robust, secure, and reliable environment for the next generation of language models. More details can be found on our website: \url{https://chonghan-chen.com/llm-jailbreak-zoo-survey/}.
[ 0.0698621422, 0.078648977, -0.0055694249, 0.0488077924, -0.0809420198, -0.0188770182, 0.0305150189, -0.0020183546, 0.0740283355, -0.0371022336, -0.0234739929, -0.003870446, 0.0116106812, 0.0406741388, 0.029882513, 0.0076726526, 0.0523399226, -0.0362896584, -0.015668299, 0.0028455204, 0.0142310709, -0.0078733573, 0.0557678007, 0.0072194324, -0.0052653095, -0.014048188, -0.0289014839, 0.0146877393, 0.0064293798, -0.0213122945, -0.0163655113, 0.0122164674, 0.0030801145, 0.0801119655, 0.0000023205, -0.0435899906, -0.0519994348, 0.04224528, 0.0094321882, 0.0356620029, 0.019474471, -0.095088847, -0.021063108, -0.0175648052, -0.052269496, 0.0188254882, 0.0446519926, 0.055016771, 0.0785369948, 0.0377043448, -0.0041346396, -0.0252892543, 0.0583946221, 0.0050092745, -0.0199231505, -0.0630326122, 0.0193949789, -0.0197651871, 0.0003103226, 0.0166486874, -0.0013923276, 0.0133566409, 0.0353509858, -0.0247587115, -0.0509560592, -0.0151769826, -0.0415756404, -0.0562589541, -0.0051853247, 0.0283594504, -0.0407408215, -0.0245035198, 0.0103792381, 0.0311450977, 0.0177453831, 0.0199557468, 0.0067658266, 0.0625897273, -0.0067010159, -0.0185305309, 0.0494309142, -0.0201475285, 0.044227384, 0.012837342, 0.0023612368, 0.0009841217, 0.0146592101, -0.0384371281, -0.0090086386, -0.0447801128, -0.0069677075, -0.0569421165, 0.0182295181, -0.0018588399, 0.0265138578, 0.0052309735, -0.09240973, -0.0256277397, 0.0147539489, -0.0574272051, -0.0022321525, 0.0134009542, -0.0146687534, 0.0175672527, 0.0157540552, 0.0099316314, -0.0423879735, 0.0239046197, 0.0111832088, -0.0107452255, -0.0064783082, -0.0475328192, -0.0000359722, 0.0667138249, -0.0407428406, -0.0063029914, -0.0103420103, 0.0243165325, -0.0462892801, 0.0292365383, -0.0596559346, 0.0188458953, 0.0271499027, 0.01140022, 0.0410353877, -0.0392018184, 0.0037747249, -0.0272282287, -0.0147177186, -0.0642106012, -0.0058761071, -0.0018751579, -0.0951455459, -0.0033379043, -0.0113988221, 0.125031203, 0.043387372, -0.0089395642, -0.08182735, -0.0920694247, -0.0345970131, -0.0426036008, 0.0191018246, -0.075708203, -0.0246129036, -0.0379214026, -0.0032718491, -0.0481105782, 0.0209302418, -0.0037745631, -0.0583458804, 0.0390489325, -0.0190935973, -0.0160213746, 0.002812322, 0.0524011962, 0.0678068697, 0.0384374112, 0.0019962322, 0.0510342419, 0.071314089, -0.0094599677, 0.0614498071, 0.0291973613, -0.0298800822, 0.0082336105, -0.0471945032, -0.0010519086, 0.0447792672, -0.0125235468, -0.0565499924, 0.0477840491, 0.0034026331, -0.0419208929, 0.0963315815, 0.1058846414, -0.0465030186, 0.0276456941, -0.0139706237, -0.0468206555, 0.0442313813, -0.0145117249, -0.0476513132, 0.0071634548, -0.0626756623, -0.0393394902, -0.0855514407, 0.0577564351, 0.0063029639, -0.011443655, 0.0120819611, -0.0413779505, -0.0127771739, 0.0272053238, 0.0126163829, 0.0135338269, 0.040264152, 0.0104424991, 0.0065337582, -0.0428550318, -0.0354818776, -0.0932925344, -0.0032414969, 0.0803096145, 0.0429561995, 0.0210563075, 0.004897601, 0.0070887776, -0.0143842818, 0.0042298087, -0.0156897344, 0.0173592232, 0.0151745658, -0.0012859516, 0.0075017214, -0.0333015993, 0.0161097348, 0.0099263191, -0.0271645412, -0.0071699196, -0.0111993114, -0.0228281934, -0.0355921052, 0.0032694044, 0.0252657253, 0.0242037084, 0.0105629293, 0.0220370796, -0.0122449892, 0.023355484, -0.0558384135, 0.0262191761, 0.0750222579, -0.0066928389, 0.0684677139, -0.0034341221, -0.0140011674, 0.0215023216, 0.0276116375, -0.0631549656, 0.0304747913, 0.0896018371, -0.0279370993, 0.0024137604, 0.0457386263, -0.0395474993, 0.0218152292, -0.0195030458, 0.0281431489, -0.0351553112, 0.0161748268, -0.0422740541, -0.0036548735, -0.0342317224, -0.0327318087, -0.021975683, 0.0176363848, 0.0345467031, 0.0182321742, -0.0462727509, 0.0467437878, -0.0239847004, 0.0141443741, 0.0061599491, 0.0205622241, -0.0050143823, -0.0407425873, -0.0233796947, -0.0182617903, -0.0123255746, 0.0101329871, 0.025304364, 0.0138965454, -0.0272232946, -0.0258102175, 0.0052163573, -0.0204863138, 0.0068757967, -0.0000634995, -0.0096932817, -0.0216230247, -0.05075147, 0.0144359171, -0.0218195803, 0.0078164376, 0.030132547, -0.0256648213, -0.0359483622, -0.0106607853, -0.0198507309, -0.0223873127, 0.0184306353, -0.0087749055, -0.0071414285, 0.0277909432, 0.0093645863, -0.0383997709, 0.0718492791, 0.0087224003, 0.0691370741, 0.0565689802, 0.0021644717, -0.0007335056, -0.0535054281, 0.0771679133, -0.0429907404, -0.0409244336, -0.0442831218, 0.0181417093, 0.0105706388, 0.0196915492, -0.0009755462, -0.0159593131, 0.0590246506, -0.0041110474, -0.0511592776, 0.026324993, -0.0482233912, -0.0370238423, -0.0103849825, -0.0258382969, -0.0873170346, 0.0562043898, -0.0587708838, 0.0230813101, 0.0050217672, -0.0649505854, -0.0271047074, 0.0602176636, 0.0125891762, 0.0457116216, 0.0542802364, -0.0246097315, -0.0277032126, -0.0015573623, -0.0047616963, -0.0274291039, 0.0412797444, -0.0604092441, -0.0243853964, -0.0160399806, -0.0332250968, -0.025879534, -0.0392866582, -0.0089674219, -0.0086414889, 0.0352545381, -0.0371753462, 0.0214507002, -0.0301765595, 0.0430276394, -0.0209523588, 0.0016964636, -0.0030820579, -0.0367028303, 0.0252352878, 0.0044201491, -0.0589243472, -0.0074228477, 0.0312342271, 0.0369080119, 0.0922902077, 0.0499881916, 0.032650549, 0.0265528318, 0.0596053936, 0.0360883661, 0.0281891301, -0.0239977222, 0.0064491793, 0.0158046596, -0.0222784169, -0.0154888397, 0.0317485221, 0.0198495165, -0.0176889673, 0.0304817893, 0.0131992074, -0.0052762409, 0.0177946966, 0.0276709944, -0.0738637671, 0.0010652078, 0.0112085938, 0.0704801679, -0.0494519807, -0.02278159, -0.0409948416, -0.0731552318, -0.0044095474, 0.0448555984, 0.0046513951, -0.0196322054, -0.0291211773, -0.0462399572, 0.0238818731, 0.0027114295, -0.031393908, -0.0096482038, 0.0011661425, 0.0187937487, -0.0465130061, 0.0276727602, 0.0177801978, -0.0182464365, 0.0422555394, 0.0913966298, -0.0058460608, -0.0435853377, 0.0274659153, 0.0132448785, 0.0218205284, 0.0092977611, 0.0057074642, -0.0179810245, -0.0057291798, -0.0159690809, -0.0048570996, 0.0484729074, 0.0238717087, -0.0527887605, 0.0374089479, 0.0411937311, -0.021182077, -0.0276165418, -0.0290606301, -0.047192242, -0.0092172362, 0.0280221626, 0.0018594145, -0.0363982059, -0.0672904179, -0.0538393371, -0.0402458198, -0.0584179759, -0.0176775288, 0.0185660012, 0.0307124425, -0.0315751284, -0.0215290133, 0.0028125329, -0.0291185621, 0.0346010737, 0.0579993017, -0.0119942175, 0.0047482643, -0.0028420913, -0.0302960817, -0.0220951866, 0.0023363479, 0.0453025103, -0.0213877242, 0.0353965089, -0.0085181976, -0.0205961186, 0.0051651336, 0.0171290133, -0.025999194, -0.0546134599, 0.0416466855, 0.0669005513, 0.0191732645, -0.0245355181, -0.0334704071, -0.0442845821, 0.0394281074, 0.0918294713, -0.0412357114, 0.0361655392, 0.0625011548, -0.0251962114, 0.0128588201, 0.0133232968, 0.0007912036, -0.0273375995, 0.0827575028, -0.0005923709, -0.058768332, 0.0119749987, 0.0147587359, 0.0509181209, -0.0293706395, -0.0129783507, 0.0438306034, 0.0439813435, 0.0230308138, -0.0067583355, -0.0336982012, -0.0275326576, 0.0080649843, 0.049921006, 0.0024842818, 0.0313670896, 0.0411073267, 0.0325298421, -0.0451073572, 0.0147526283, 0.030636372, 0.0301627796, 0.0520126, 0.0415348001, 0.0390441343, -0.0336132683, 0.0653791949, -0.0417633988, 0.0392101891, -0.0218504965, 0.0594511442, -0.0039884318, -0.0148482639, -0.0378174558, 0.079940781, -0.0782578737, -0.0477646142, 0.0100802304, 0.1137196422, -0.0142414672, 0.0266687237, -0.0282380488, -0.016859306, 0.0079101464, -0.043464534, -0.0348282792, 0.0021577559, -0.0563358888, -0.0110052712, 0.0064246305, -0.0203765333, 0.0243912265, 0.0478101447, 0.0397013314, 0.0498300306, 0.0137780868, -0.0095093995, -0.017743187, -0.001835975, 0.0674258396, -0.0108945053, -0.0400017649, -0.0394192487, -0.0752403364, -0.0771218166, -0.0096623013, 0.0115700355, 0.0356933661, -0.0028772259, -0.0430594534, 0.0538196675, 0.0368365906, 0.0112687517, -0.0634869635, 0.0119190048, -0.0090451585, -0.0544304587, 0.0362846367, -5.615267454e-33, 0.0404971279, -0.0339400135, -0.0667208806, -0.0250565633, -0.0074421465, 0.0007042529, 0.0049607926, 0.004731731, -0.0318867266, -0.0441887416, -0.0230971035, -0.0096206069, -0.0101067666, 0.0133234421, -0.0369441062, -0.0215887278, -0.0255189706, 0.019574469, -0.0014644908, 0.0110112149, 0.03802168, 0.0212083757, 0.011847374, -0.0366354287, 0.0663529932, -0.0251551084, -0.0244086124, 0.0363264941, -0.0531381592, 0.0527448803, 0.014436054, 0.0060874834, 0.0109853139, -0.0142334066, -0.0114622647, 0.0891413763, 0.0081940368, 0.0233085621, 0.0102504557, 0.0123659, -0.0010088575, -0.0367895924, 0.0096855722, -0.0068958295, 0.0024854373, -0.0016022443, -0.0270059109, -0.0368220508, -0.0175567195, 0.0115896212, -0.0000576561, -0.0208585896, -0.0455783978, 0.01632379, 0.0212864485, -0.0442961045, -0.0123511665, -0.0012021422, -0.008933844, 0.05045316, 0.0414116159, -0.0253963061, 0.057243973, 0.0004543072, -0.0013567489, 0.0518258736, 0.0088528385, -0.0500706583, -0.0026460662, -0.0259334724, -0.011319506, 0.0836079046, 0.0587347709, 0.0712533146, 0.0162589457, -0.0132317208, -0.0415466614, -0.0368986204, 0.0016504374, 0.0318564437, 0.0305738356, -0.0067903879, -0.0319392234, 0.0044418885, -0.0154554313, -0.0309276078, -0.0375060551, -0.010345432, 0.0239449125, 0.0049268622, -0.0406130962, -0.0069495211, 0.0099909082, -0.0927559137, 0.0112237697, 0.0106896898, 0.0139897503, -0.0034702185, -0.0146065066, 0.0351470858, -0.0137662115, 0.0024008558, 0.0134316776, -0.031728439, 0.0437899344, -0.018550938, 0.029936092, 0.0186560992, -0.0308979992, 0.0004700374, 0.0017598988, 0.0417567715, 0.0286416858, -0.0000393663, 0.0141192125, -0.0290139448, -0.0025923613, -0.0262964144, 0.0320388451, 0.0152128944, -0.0500423871, 0.0285111591, -0.0458385795, 0.0295181312, 0.022447234, 0.0183970109, -0.0063776583, 0.0353211574, 0.0497109853, -0.0349565856, -0.0007234908, 0.0143954745, 3.133e-7, -0.0423265435, -0.0331620239, 0.0101030627, -0.0493264571, 0.0180396345, 0.0755407885, 0.002652742, 0.0773704723, -0.033388447, -0.0288408007, -0.0387165509, -0.0116433501, 0.0689482242, -0.0079965899, -0.0061212122, 0.0469171144, -0.0337713324, -0.0277011208, -0.0680200681, 0.0198938083, 0.0288303569, 0.0526443273, -0.0092510711, -0.0440995581, -0.0010224548, -0.0820604414, -0.0174423307, 0.0060093501, -0.0102095073, -0.0325374305, -0.0066986987, 0.0604028143, 0.0032310858, -0.0012793372, -0.016524842, -0.023356868, -0.0398043618, 0.0296257213, -0.0072153299, -0.0048721964, -0.0574367978, -0.0391827673, -0.0320773125, -0.0478357002, 0.0686268732, 0.0322059244, -0.0375057384, -0.0136899576, 0.0408507697, -0.0399248488, 0.0023463154, -0.0341549516, 0.0226382222, 0.0041432665, 0.0315243974, 0.025927119, 0.0214032643, -0.0766609162, 0.0064508864, 0.0121829249, 0.0140185198, -0.0301221721, -0.0048648915, 0.0563892536, 0.0557434633, -0.0130756469, -0.0043644239, 2.963256643e-34, -0.0278221276, -0.0609511957, 0.0178880338, 0.0342987292, 0.018834075, 0.0177970398, -0.0160909463, 0.0156163611, -0.0009677608, -0.0256516635, -0.0065287696 ]
http://arxiv.org/abs/2406.18002v1
http://arxiv.org/pdf/2406.18002v1
Decoding with Limited Teacher Supervision Requires Understanding When to Trust the Teacher
2024-06-26T01:16:12
2024-06-26T01:16:12
Hyunjong Ok, Jegwang Ryu, Jaeho Lee
cs.CL
cs.CL, cs.AI
How can sLLMs efficiently utilize the supervision of LLMs to improve their generative quality? This question has been well studied in scenarios where there is no restriction on the number of LLM supervisions one can use, giving birth to many decoding algorithms that utilize supervision without further training. However, it is still unclear what is an effective strategy under the limited supervision scenario, where we assume that no more than a few tokens can be generated by LLMs. To this end, we develop an algorithm to effectively aggregate the sLLM and LLM predictions on initial tokens so that the generated tokens can more accurately condition the subsequent token generation by sLLM only. Critically, we find that it is essential to adaptively overtrust or disregard the LLM prediction based on the confidence of the sLLM. Through our experiments on a wide range of models and datasets, we demonstrate that our method provides a consistent improvement over conventional decoding strategies.
Decoding with Limited Teacher Supervision Requires Understanding When to Trust the Teacher: How can sLLMs efficiently utilize the supervision of LLMs to improve their generative quality? This question has been well studied in scenarios where there is no restriction on the number of LLM supervisions one can use, giving birth to many decoding algorithms that utilize supervision without further training. However, it is still unclear what is an effective strategy under the limited supervision scenario, where we assume that no more than a few tokens can be generated by LLMs. To this end, we develop an algorithm to effectively aggregate the sLLM and LLM predictions on initial tokens so that the generated tokens can more accurately condition the subsequent token generation by sLLM only. Critically, we find that it is essential to adaptively overtrust or disregard the LLM prediction based on the confidence of the sLLM. Through our experiments on a wide range of models and datasets, we demonstrate that our method provides a consistent improvement over conventional decoding strategies.
[ -0.0070892451, 0.0758449063, 0.0219942015, 0.0297140777, 0.007445985, 0.0087307338, 0.0200376939, -0.0103850653, -0.041102279, -0.0085482122, -0.0036673564, 0.0045237723, 0.0042673051, -0.0328044891, 0.0361328274, -0.005507193, 0.0122816181, -0.0636858493, -0.0008140132, -0.0200005379, -0.0377829857, 0.0179063305, -0.0145087214, 0.0289139096, -0.0347037986, -0.0034426122, -0.0133494884, 0.0505507626, 0.0270577427, -0.0146667548, -0.0138773657, 0.0505408123, 0.0310886148, 0.024912443, 0.0000018635, -0.0111424262, -0.0342618451, -0.0161737744, -0.0406052917, 0.0076551507, 0.0092518823, -0.0383105874, -0.0362799391, 0.013376018, -0.0338633098, 0.0086129401, 0.0334650837, 0.0071738269, 0.038995076, 0.0764878839, 0.0144290756, -0.0248736218, 0.0217234287, 0.0292200483, 0.022285521, -0.0454889387, 0.0239116047, 0.0060184165, -0.0566064529, -0.0287432857, -0.0412347391, 0.0469147451, 0.033137694, -0.0236241221, 0.0160304531, 0.0198945142, 0.0041483659, -0.030834334, -0.0413202047, 0.0087058162, 0.013914315, -0.0565815978, 0.0058533149, 0.0030656138, 0.0244460516, -0.0193821471, -0.0434987918, 0.0643825457, -0.0343101583, 0.0737520903, 0.0336285718, 0.0246924553, 0.0226427075, -0.0167724155, -0.0255245492, 0.0559152178, 0.0218005963, -0.0303254053, -0.0130095789, 0.0625705943, 0.0181420781, -0.0104232263, 0.0004616867, 0.0250591543, -0.0355437733, 0.0343359932, -0.0676131621, -0.0358442403, -0.0021260271, 0.0617880523, 0.0073878206, 0.0216524582, -0.0227930415, 0.0021879633, -0.0155876642, 0.0214444529, 0.015133976, 0.0572637059, -0.038181752, -0.0077064936, -0.02255385, -0.0096337376, -0.0215419792, 0.0387972631, -0.0267424341, 0.0329010822, -0.0565072559, 0.0107510686, 0.0213389099, 0.0248980131, 0.0071958276, -0.0121493386, -0.0502966046, -0.0021537109, 0.0205814783, 0.018795846, 0.0088928528, 0.0398417935, 0.0184082892, -0.0785918981, -0.0241461862, 0.0249600131, 0.0128456689, 0.0113152536, -0.0426646993, 0.0596797094, -0.0317498781, -0.0059387204, -0.0934546068, -0.0653233752, -0.0178431459, -0.0363300703, 0.016496053, -0.0826330036, 0.0152513068, -0.0204854235, 0.0169817694, -0.0325729623, -0.0181474965, -0.027577322, 0.0375372395, 0.0287741777, -0.0479635037, 0.0061002998, -0.011220688, 0.0119079622, -0.0085322559, 0.0874292254, 0.0015580991, 0.0531372391, 0.021964049, 0.0080378912, 0.0219547059, 0.0607975274, 0.0547227748, 0.0357547924, 0.0111621208, -0.0344565772, 0.0088640377, -0.0077230176, 0.0608566962, 0.0503502116, -0.071424298, -0.0024780619, 0.0725022554, 0.0720148981, 0.0290665831, 0.007943538, 0.0343969427, 0.0128673445, 0.012567332, 0.0626185462, -0.0221789461, 0.0000152936, -0.0226054583, 0.0123511218, 0.0239269435, 0.0313456655, -0.016017735, -0.0725349262, -0.0091348849, -0.045298662, -0.0288340934, -0.0000845235, 0.0601440035, 0.022061564, -0.0439630002, 0.0613714345, 0.0241968073, -0.0661711395, 0.0167555921, -0.0801725537, 0.0257841945, 0.0593485609, 0.0451391004, 0.0163519476, -0.0271247476, -0.0106413122, 0.039449133, 0.0464733429, -0.0216566734, -0.0061198641, 0.0496998057, -0.0550064929, -0.0560022928, 0.0592428111, 0.0233723726, 0.0330989882, 0.0216330178, -0.0058392617, -0.0007189691, -0.035182301, -0.0297583397, 0.0560128018, -0.0063988464, -0.0435939059, 0.0070141093, 0.010099696, 0.0408444852, 0.0465350784, 0.0039855503, 0.014748442, 0.0493927486, -0.0407264605, -0.0358034186, -0.0247686841, -0.0706774145, -0.0686608776, 0.0273656808, -0.0562874377, -0.0308300145, 0.001873867, -0.0221091863, -0.0516275167, -0.0037271881, -0.0572372489, 0.0629566908, -0.0554015338, 0.0127197793, 0.0247636251, -0.0160875097, -0.0264872871, 0.0252139363, -0.033122234, -0.0261666495, 0.0284797698, 0.0133071095, -0.0041876542, -0.0703773797, 0.0290284045, -0.000899661, -0.0492534637, 0.0365190618, -0.0067767203, 0.0475833416, -0.0501839556, -0.0674663484, -0.0308268573, -0.0509814546, -0.0526305884, 0.0235142577, -0.000498666, -0.0476508439, -0.0178037751, -0.0103282183, -0.0464134626, -0.0383249708, -0.0169047769, -0.0301933419, -0.0041678506, -0.0224583335, -0.0401837528, -0.0194779057, -0.0111291893, 0.0077860397, 0.0270364042, 0.0104966098, -0.0311716348, -0.0716351345, 0.0291678868, -0.0184476189, -0.0165277924, 0.0069907201, -0.0241384897, 0.0461300798, -0.0423737504, 0.0138731804, 0.0206089951, -0.0232447069, 0.0065750219, 0.0238319784, -0.0175067596, -0.0835939646, -0.0147167901, 0.0036324295, -0.0549247563, -0.060829442, -0.0240795556, 0.0227665361, 0.0395788997, 0.0509981476, -0.0009344296, -0.0587831847, -0.0098798061, 0.0029930067, -0.0858339593, 0.0118402671, 0.0011129939, -0.0122123193, 0.0402142294, 0.0081444699, -0.0868213102, 0.0483437702, -0.0422301441, -0.0188158937, -0.0149163334, -0.0113441264, -0.0315494351, 0.0084308162, -0.057380259, 0.0184296966, -0.0237092394, -0.0675998926, -0.0201453138, 0.0105397711, 0.0071161608, -0.0388353057, 0.024215525, -0.0544486567, 0.0102520119, -0.024629565, -0.0145164691, -0.0494019315, 0.0126683433, -0.0139785325, -0.0063317241, 0.0060179569, -0.0322223566, -0.0260077473, -0.0386995748, -0.0294775069, -0.0075797006, -0.0200751871, 0.0035694933, -0.0319677293, -0.0017640588, 0.0069343783, -0.0008694656, 0.003510738, -0.0696863607, 0.0366373174, 0.0622605458, 0.0611479096, 0.0282292385, -0.0244483966, 0.0177857969, 0.0618923455, 0.0199089926, 0.0465613678, 0.0579041131, 0.0466753058, -0.0205972642, 0.0136878183, 0.1252538711, 0.0163909644, -0.0360304005, -0.0181983691, -0.0151453279, 0.0420545563, -0.0087214997, -0.0222409926, -0.0742485896, -0.0300380532, 0.0149600627, 0.071193926, -0.0001828582, -0.0094826249, -0.0193130951, -0.0603436157, 0.0542697497, 0.0669595972, -0.0529392771, -0.0157498047, -0.0399900936, -0.0687310398, -0.0121298507, 0.0614588484, 0.0257422123, -0.0057864604, 0.0069078645, 0.0517930537, -0.0160850529, -0.0384293161, -0.0464638844, -0.0325237587, 0.0736461654, 0.0602204576, -0.0060784509, 0.0097046513, -0.0502864532, 0.019497199, 0.0019977754, -0.0307658724, -0.027172735, 0.019654708, -0.0334925391, 0.0085831983, -0.0088101486, 0.0885502622, -0.0509853177, 0.010349323, 0.016672628, 0.0532552451, -0.0090072835, -0.0347113945, -0.0004469127, 0.0020658944, 0.027039364, 0.0870554, 0.0536275133, -0.0104561262, 0.0090990458, -0.0369685814, 0.0154860038, -0.0528009832, -0.0058175665, 0.058561936, -0.0030374997, -0.0209932271, -0.0577840321, -0.0255321153, 0.0372248851, 0.0301022734, -0.0154495388, 0.0447287373, -0.021077184, -0.0592444502, -0.0274924506, -0.0005163691, -0.0406158082, 0.0016669942, -0.0824583322, -0.0158139598, 0.0030954198, -0.0399865173, 0.0281742364, 0.0065496806, -0.1015740037, -0.0732183531, -0.0097800521, 0.029645795, 0.0257113241, 0.0288838185, -0.004021652, -0.0205833856, 0.0912611559, 0.0572097413, -0.0048218775, 0.0387002379, 0.0642454848, 0.0262201186, 0.0158584304, 0.0443309955, -0.0524943061, 0.0048872815, 0.043732658, 0.0328548253, -0.0347346403, 0.0280933101, 0.0424221121, 0.0292085838, 0.0142067987, -0.0586144403, 0.0043328688, -0.0175004359, 0.0468099043, -0.0081608053, 0.0246588998, -0.03588783, 0.0232764743, 0.0268831626, -0.004511361, 0.0127484975, 0.0021355203, 0.0034323279, 0.0511741489, -0.0245092381, 0.0222023576, -0.0145942336, -0.0210893806, 0.0236553662, -0.0159377493, -0.021467302, 0.0254006162, 0.0002026784, 0.0077048927, 0.0400970131, -0.0221883785, 0.0093751177, 0.0650918782, 0.0011940394, -0.00246646, -0.0486036465, -0.0541299209, -0.0211126655, 0.0263101235, 0.0659852475, 0.0428090841, 0.0611939058, -0.0180078652, 0.0308416691, -0.0343794338, 0.0193616617, -0.0209510308, 0.0071941796, 0.0342852101, -0.0060378546, -0.0127886878, -0.0180435106, 0.0505751148, 0.0151866069, 0.0568447374, -0.0466348641, 0.0292586964, -0.0445719585, -0.0423421934, 0.0303072277, -0.0150119746, -0.015567177, -0.0209443532, -0.0276767537, -0.0490873605, 0.0138405096, 0.0259520598, 0.0210802183, 0.0004062839, 0.0351151191, 0.0310078971, -0.0522232428, 0.0120856967, 0.0077559673, -0.0476358086, 0.0574898385, -0.0473634973, -0.0009885668, -6.086191062e-33, -0.034287326, -0.047280211, 0.0379962251, 0.01173825, 0.0287535284, 0.0373226367, -0.0123703294, -0.0171312988, 0.0502408184, 0.0139353368, -0.0031592203, 0.0045873164, 0.0465807915, 0.0460353568, 0.0202066079, -0.0061207269, 0.0562217236, -0.0246302914, -0.0173423477, -0.0189037155, 0.0241054539, 0.036437761, 0.0505365171, -0.0195308458, 0.0427537449, -0.0179088172, -0.0157787669, -0.0064206747, -0.030638136, 0.0712854266, 0.0000031872, 0.0241524521, 0.0157144815, -0.0045940494, -0.0618642978, 0.00252072, -0.0694240853, -0.0119292736, 0.0002563344, 0.0070982534, -0.0349519849, -0.0457038172, 0.0244311728, 0.016111169, -0.0775902495, -0.0189357828, 0.0099378908, -0.0269703139, -0.0381875522, -0.0826586112, -0.0143356519, 0.0298644993, 0.0052862894, 0.000870403, 0.0480440073, -0.0012568312, -0.0103998277, 0.006749826, -0.0030142197, 0.067332983, 0.050460007, 0.0254194196, 0.0845933333, 0.0104260789, -0.03243443, 0.0554265566, 0.0324838832, -0.0357277691, -0.0252801348, -0.009364468, 0.0299645606, 0.0521385111, 0.0157583114, 0.0284875203, 0.0236315448, 0.0099062799, -0.0469982624, 0.0406634733, 0.0088885948, 0.0195772015, -0.0059790136, 0.0393774137, -0.0140592977, -0.0160960145, -0.044394251, -0.0362046212, -0.0031814689, 0.0018164566, 0.0249005686, 0.0388386175, -0.0450946391, -0.0624080338, 0.038563475, -0.0334686488, -0.0067327432, -0.02301584, 0.0273566227, -0.0287499484, -0.037490055, 0.0020887, -0.0318573229, -0.0269630328, 0.0724284351, -0.0428229943, 0.0063588084, 0.0042475318, 0.0490260907, 0.0167142283, -0.0531740673, 0.0152639924, -0.0333764441, 0.0147357294, 0.0491286255, -0.0310781673, 0.0495656244, -0.0552432276, -0.0080834841, 0.013680934, -0.008427117, 0.0248117335, -0.0122149186, 0.0290526617, -0.0047880383, 0.0128616672, -0.0379481688, -0.0014892713, -0.0057148742, -0.0387121327, 0.0508485511, -0.0137624973, -0.008136645, 0.0110080438, 2.446e-7, 0.0130252643, 0.0281630978, 0.0735992491, -0.0185921285, 0.0210175179, 0.005235312, -0.0354592428, 0.0136793274, -0.0257346854, -0.0415628925, 0.0005071357, 0.0363225378, 0.0615135133, 0.0474074483, -0.0907196924, 0.0168792345, -0.0057676402, 0.0224469937, -0.0368203335, -0.0022425579, 0.0644051433, 0.0336565264, -0.014181409, -0.055038508, 0.044179488, -0.0010286525, 0.0118922107, -0.0014251537, -0.010082054, 0.006073344, 0.0579720512, -0.036774423, 0.0329445153, 0.0045800069, -0.0097506521, -0.016750142, -0.0201160107, 0.0494382419, 0.0200234242, -0.0513515733, 0.0209194999, 0.0061508645, -0.0068867831, 0.0396826714, 0.0300059114, 0.0133723179, 0.0076984745, 0.0340377949, -0.0257309936, -0.0146786273, -0.0178655107, -0.0340400115, 0.0155252432, 0.0531275086, 0.0107672932, 0.0171500836, 0.0028603559, -0.0382147878, -0.0207366049, -0.0939743593, -0.0521586016, -0.0955674797, -0.0150134768, 0.0691648424, 0.0023876817, -0.0556027666, -0.0112214191, 2.208558276e-34, 0.0425596945, -0.0210199729, 0.039541591, 0.0657616556, 0.0021923543, -0.0330870077, -0.0609309338, 0.0462374613, 0.0023149971, -0.0243990179, -0.017774431 ]
http://arxiv.org/abs/2406.17992v1
http://arxiv.org/pdf/2406.17992v1
Catching Chameleons: Detecting Evolving Disinformation Generated using Large Language Models
2024-06-26T00:21:39
2024-06-26T00:21:39
Bohan Jiang, Chengshuai Zhao, Zhen Tan, Huan Liu
cs.CL
cs.CL, cs.AI
Despite recent advancements in detecting disinformation generated by large language models (LLMs), current efforts overlook the ever-evolving nature of this disinformation. In this work, we investigate a challenging yet practical research problem of detecting evolving LLM-generated disinformation. Disinformation evolves constantly through the rapid development of LLMs and their variants. As a consequence, the detection model faces significant challenges. First, it is inefficient to train separate models for each disinformation generator. Second, the performance decreases in scenarios when evolving LLM-generated disinformation is encountered in sequential order. To address this problem, we propose DELD (Detecting Evolving LLM-generated Disinformation), a parameter-efficient approach that jointly leverages the general fact-checking capabilities of pre-trained language models (PLM) and the independent disinformation generation characteristics of various LLMs. In particular, the learned characteristics are concatenated sequentially to facilitate knowledge accumulation and transformation. DELD addresses the issue of label scarcity by integrating the semantic embeddings of disinformation with trainable soft prompts to elicit model-specific knowledge. Our experiments show that \textit{DELD} significantly outperforms state-of-the-art methods. Moreover, our method provides critical insights into the unique patterns of disinformation generation across different LLMs, offering valuable perspectives in this line of research.
Catching Chameleons: Detecting Evolving Disinformation Generated using Large Language Models: Despite recent advancements in detecting disinformation generated by large language models (LLMs), current efforts overlook the ever-evolving nature of this disinformation. In this work, we investigate a challenging yet practical research problem of detecting evolving LLM-generated disinformation. Disinformation evolves constantly through the rapid development of LLMs and their variants. As a consequence, the detection model faces significant challenges. First, it is inefficient to train separate models for each disinformation generator. Second, the performance decreases in scenarios when evolving LLM-generated disinformation is encountered in sequential order. To address this problem, we propose DELD (Detecting Evolving LLM-generated Disinformation), a parameter-efficient approach that jointly leverages the general fact-checking capabilities of pre-trained language models (PLM) and the independent disinformation generation characteristics of various LLMs. In particular, the learned characteristics are concatenated sequentially to facilitate knowledge accumulation and transformation. DELD addresses the issue of label scarcity by integrating the semantic embeddings of disinformation with trainable soft prompts to elicit model-specific knowledge. Our experiments show that \textit{DELD} significantly outperforms state-of-the-art methods. Moreover, our method provides critical insights into the unique patterns of disinformation generation across different LLMs, offering valuable perspectives in this line of research.
[ 0.0303570889, 0.0293432828, -0.0218203384, 0.0617221668, -0.0240811165, 0.0034576773, -0.0071037877, -0.0130919246, -0.0118137319, -0.051656045, 0.0077871205, -0.0187409371, 0.0118580209, 0.0535772592, 0.0259188283, -0.0109657561, 0.0902048722, -0.0063282666, 0.0313881226, -0.0301075056, 0.031769421, 0.0191513039, 0.0039913789, 0.0275991727, -0.0354624093, -0.0297166742, 0.0203132685, 0.0023056709, 0.0237767082, -0.013529432, 0.0310705639, 0.0390508063, 0.0469369069, 0.113761805, 0.0000018568, -0.0171290748, -0.0055975667, 0.0081539042, -0.0631179512, 0.0387899131, 0.0637786537, -0.0398332141, -0.0109236669, 0.0124021433, -0.0205808915, -0.0216264371, 0.0348739438, 0.0529988706, 0.0941913947, 0.0838366225, 0.0124714365, -0.0087071415, 0.0081497766, -0.0280217156, -0.0193036068, -0.0097047482, 0.0548196137, -0.0328500234, -0.0651482418, 0.0359281935, 0.0043161362, 0.0067979163, 0.0208407659, -0.0074235755, -0.0463558994, 0.0190110784, 0.0056299684, -0.0327506438, -0.0232091565, 0.0649783313, -0.0313447081, -0.0556642935, 0.0041301199, -0.0050103655, 0.0351204015, 0.0183587223, 0.0236025918, 0.0230285991, -0.00813042, 0.0050931871, 0.0245062057, -0.0205926485, 0.0580350496, 0.0319836996, -0.0085244821, 0.0422206745, 0.0075180777, -0.0294522792, -0.0071057831, -0.0532561392, -0.0090506021, -0.0867701024, -0.0179094393, 0.0118017327, -0.0379581824, 0.02526737, -0.0867680386, -0.0256444495, 0.04469008, -0.0680850148, -0.0051388415, -0.0118513899, -0.0235799197, 0.030301135, -0.0232792851, 0.030564623, -0.0231194738, 0.0215963777, -0.0059385081, -0.0080278683, -0.002374894, -0.0034957121, -0.0006221746, 0.0357533656, 0.0065851435, 0.0095542241, -0.0168442894, 0.0082723126, -0.0382907353, 0.0380933434, -0.0478274934, 0.0091949878, -0.0217050444, -0.0232769288, -0.0280824918, 0.0510488339, -0.0065481151, 0.0051586176, 0.0027894266, -0.0442657694, -0.0174011178, 0.0280550029, -0.0567083955, 0.015904827, -0.0133700315, 0.0158581808, -0.0147940228, 0.0180768427, -0.0448425598, -0.0402040146, -0.047933653, -0.0463847592, -0.0038834584, -0.0597197451, 0.0011639398, -0.031331256, -0.0214023832, -0.0631007701, 0.0099275419, 0.0009513276, -0.0565110445, 0.0735741928, -0.0410184711, -0.006790502, 0.0753275901, 0.034490224, 0.0082299598, 0.0825106502, -0.0182167124, 0.0586772636, 0.0513815358, 0.0096989004, 0.0120468885, 0.0136195878, 0.052900143, 0.0754500777, 0.004735033, -0.0079030152, 0.0252678599, -0.0132799447, 0.024108341, 0.0418779962, -0.0143605834, -0.0667815655, 0.0791049972, -0.0494008586, -0.0345730931, -0.0246969238, 0.0591002628, 0.03556576, -0.0249716192, 0.0212384798, -0.0455465652, -0.0211761948, -0.0597484522, -0.0319527574, 0.011044085, 0.0583369322, -0.004412292, -0.0362496972, -0.014248224, -0.0092222374, 0.0008863329, -0.0100580119, -0.0120966975, 0.038542904, -0.0244349614, 0.0875790492, 0.0620064735, -0.0016624588, -0.004078744, -0.0551234372, 0.0168726537, 0.040143691, 0.0432569981, -0.0131768361, -0.0386514366, -0.0439445861, 0.0190549269, 0.0406162776, 0.0668822601, -0.0342178307, -0.0285340194, -0.001700762, -0.0010771777, -0.0357660055, 0.0451161005, 0.0207613539, 0.020603884, 0.0147350943, 0.010250818, -0.0349697024, -0.0496079326, 0.0355973989, -0.0024917887, -0.0160174146, -0.0039591552, 0.021811299, 0.0215557814, 0.0107672401, -0.0167680196, -0.0085414955, 0.0479550436, -0.0042245006, -0.0381917246, -0.0249081254, -0.0087518869, -0.0318951569, 0.0509362929, 0.021133028, -0.0298102368, 0.005604913, -0.0431831814, 0.0135450996, 0.0254294612, -0.0337572619, 0.033825621, -0.0312945731, 0.0090798596, -0.0288489629, 0.0286995489, 0.0048169228, 0.0673153698, -0.0527915694, -0.0194732603, 0.0240059532, 0.0193962604, 0.0024687515, 0.0181817934, 0.0059217019, -0.0009436868, -0.0035189113, 0.0015710149, 0.0349593461, -0.0178127214, 0.0033303357, -0.0329711884, -0.0084542567, -0.0244740136, -0.0340099968, -0.0240164772, 0.0476033166, -0.0386138409, 0.0067292717, 0.0144178942, -0.0201507732, -0.0022444758, 0.0037419365, 0.024891939, 0.0137092276, 0.004282136, -0.035841953, -0.0229804553, 0.0458094515, -0.0140714701, 0.0488069504, 0.0125864474, -0.0263144374, -0.0532454289, -0.0167813171, -0.0638976544, 0.044048883, -0.0337450393, -0.0332011469, 0.0124099022, 0.0393466428, 0.0478157476, 0.001434489, 0.0417133495, 0.0628507733, 0.0583309084, 0.0163754728, -0.0272387583, -0.022521954, -0.036137633, 0.0014848842, -0.0556437522, -0.0619271584, 0.0265170261, 0.0588620715, 0.0382283852, -0.0229984485, -0.0211585853, 0.0462284423, -0.0172092021, -0.0689963475, 0.0307319015, -0.0544704571, -0.0196839161, 0.0240494683, -0.0395959988, -0.0558740646, 0.0835269094, -0.0519770049, 0.0409727804, -0.0255461372, -0.0097658588, 0.020979099, 0.0111545818, -0.013867815, 0.0435242914, 0.0364916362, 0.0108537963, -0.0518081374, -0.0303706396, 0.0467624031, -0.0255890545, 0.0177893192, -0.0527524576, 0.0223461054, 0.0392788872, -0.0206814576, -0.0790617839, -0.1047935113, 0.0184861142, -0.0083238054, 0.0068873432, -0.0297598019, -0.0560092218, -0.0501221791, 0.0193666331, 0.006862923, 0.0008563882, 0.0116112847, -0.0150083806, 0.0046722284, 0.0568210855, -0.0156723522, 0.0843723342, -0.0392533503, 0.0198909435, 0.0495122746, -0.0240751542, 0.0090835383, -0.0096247708, -0.0667041019, 0.0072829169, -0.0213743597, 0.0072436696, 0.0184516646, 0.0103430403, -0.0638907477, 0.001058348, 0.0928115845, 0.0191998519, -0.0158588979, 0.0065294579, 0.0160934366, 0.0247290656, 0.0330239572, -0.0150056612, -0.0942199901, -0.0054513942, 0.0127602946, -0.0087823756, -0.0374329053, -0.0106647536, -0.0409275666, -0.1061412916, 0.0110501163, 0.0080738906, -0.0010438695, -0.0274691861, -0.0241899025, -0.0240896847, 0.0530635342, 0.0248042364, -0.0126121957, 0.0306164511, -0.025965156, -0.010172681, -0.0327142514, -0.0631402507, 0.0028129471, 0.0123688476, 0.0640969202, 0.0543632768, 0.0203908253, -0.0058253487, -0.0574053116, 0.0513019301, 0.0070646042, -0.0472933948, -0.0186891612, 0.0565463752, -0.0358360037, -0.0249632057, 0.0059867366, 0.0231090859, 0.0333049446, 0.0102325911, -0.0195857715, 0.0347688086, 0.00372957, -0.0564076342, -0.0147388121, 0.0117504103, 0.0629246235, 0.0593641736, 0.0148323383, 0.0014467076, -0.0005315805, -0.0440017506, 0.020994382, 0.0270135384, -0.0189847387, 0.0286180805, 0.0009399334, -0.0435463041, -0.0580843426, -0.0050029173, -0.0099734683, -0.00960798, 0.0270753354, 0.0068024611, -0.0002957715, -0.0221965648, -0.01596017, -0.032886494, -0.0510620996, 0.0468633361, -0.0285603367, -0.0167267751, 0.0004784648, 0.0081111547, -0.0137515366, 0.0315318517, -0.089793019, -0.0861143097, 0.0260311887, 0.0265226066, 0.0111163938, 0.0622285791, -0.0376142301, -0.0468579531, 0.101308912, 0.0286299121, -0.0063731456, 0.0348093845, 0.0391561724, -0.0146702854, 0.0200487413, 0.0457588583, -0.0223281924, 0.0278814044, 0.0218635891, -0.0257990528, -0.0633555278, -0.0226590782, 0.0364045352, 0.0067298086, 0.0463967472, -0.0185519923, 0.0049426672, -0.039650593, -0.0045643169, -0.0081796367, -0.0604990721, 0.0130890962, 0.0279060118, 0.0221580621, -0.0080130063, 0.0361549594, 0.0291186329, -0.0268426295, 0.0376879387, -0.0112388572, -0.0456001647, 0.0077360291, 0.0369119756, 0.0619508587, -0.0099772168, -0.0058466396, 0.0393946841, 0.0008423454, 0.0130190263, 0.0025406065, 0.0107329264, 0.0224901065, -0.025153771, -0.0092683192, -0.0164789725, -0.0094304541, -0.0335771367, -0.0173691865, 0.0231901091, 0.0404653661, 0.0438632369, 0.0329233445, 0.005271215, -0.032665804, -0.0155643662, -0.0210947804, 0.0112983519, -0.0401306525, 0.039813742, -0.0057552676, -0.0543076806, 0.034940023, 0.0699863359, 0.0099100564, 0.0605605468, -0.003156445, -0.0357962325, -0.0273421872, -0.0314654633, -0.010031742, 0.0313388817, -0.0153837828, 0.018200295, -0.0953384042, -0.0656384379, 0.0164040569, 0.0115972804, -0.0081035849, 0.009010274, -0.0178080704, 0.0719604418, 0.0061442261, 0.0081934445, -0.0567345545, -0.002565936, 0.0210539568, -0.0221397504, 0.0350736044, -6.517400593e-33, -0.0015466112, -0.0609430633, -0.027840104, -0.0379114673, -0.020128008, 0.0352310985, 0.0011696642, -0.0261966046, -0.0360170715, 0.0611459427, 0.0574783199, 0.033149343, -0.0184097271, -0.0024698286, -0.0511578843, -0.0364914685, 0.0876847282, 0.0009369199, -0.0507532805, 0.027216807, 0.0669368133, -0.0072793663, 0.0141326841, -0.0656853244, 0.0636744276, -0.0104571497, -0.0499170125, 0.037417464, -0.04146, 0.0252526831, -0.00009935, 0.072999157, 0.0045487867, -0.0571653321, -0.0176315829, 0.0112591628, -0.0135505022, -0.0198146719, 0.0406919122, -0.0210579522, 0.0068296352, -0.0187074281, 0.0350023583, 0.0041208253, -0.0062768711, 0.0231696442, 0.0126735177, -0.0125618717, 0.0025099837, 0.0178199131, -0.0698893815, -0.0054226145, 0.0126173189, -0.0222556964, -0.0402722023, 0.0265885293, 0.0382824913, 0.022880204, -0.1037770882, 0.0162980389, 0.0272065364, 0.0092475088, 0.0385263301, -0.0069272318, -0.0140912477, 0.0734506696, 0.0196463168, 0.0642140433, -0.0110163912, -0.0098676309, -0.0005043977, 0.0214717574, 0.0440408587, 0.0799906105, -0.0053542908, -0.0493124239, -0.0748951808, -0.0074234917, 0.0222558342, -0.0103958715, 0.0041050171, -0.0317732506, -0.005844688, -0.022192128, -0.0685764104, 0.0170017779, -0.0222655628, 0.0013252263, 0.0160742868, 0.0491075404, -0.0245975461, 0.0290014911, -0.0011347883, -0.038415499, -0.0156598557, 0.0124738179, -0.0110627217, -0.0391330309, 0.0224066935, 0.0486158319, -0.0540900715, 0.0057579838, 0.0333859622, -0.0321147107, -0.0106256334, -0.0016480507, 0.0571061037, -0.0047912174, -0.0680166334, 0.03368273, -0.0114879478, 0.0445675552, -0.0063156383, -0.0403768905, 0.0328833871, -0.0081735635, -0.0246628858, -0.007158157, 0.0447500125, 0.0076152398, -0.0110952631, 0.0457924828, -0.0348975807, 0.0545030981, -0.0421563871, 0.0056791101, -0.0286489688, 0.0389015228, 0.015581388, -0.0799793899, -0.0029613143, -0.0471804291, 2.675e-7, -0.0020584629, 0.0563703179, -0.014095637, -0.0022148022, 0.0458426662, 0.0653516054, 0.0123908035, -0.0141598433, -0.0305257924, -0.081305027, -0.0333133824, 0.0431367867, 0.0052338, 0.0483800359, -0.0174817629, 0.024804797, -0.0538207665, 0.00788166, -0.0596262962, -0.021375984, 0.0168702696, 0.0928511694, 0.0412424579, -0.0303299911, -0.0095629003, -0.0191226397, 0.0627545267, 0.0255836509, -0.0391345657, 0.0307699032, 0.0329672955, -0.0129782474, 0.0065271719, 0.0085418308, -0.0346364751, 0.0020747124, -0.0588440523, 0.0178687554, 0.0096769733, -0.0115942415, -0.0483664386, -0.0246326346, -0.0101387808, -0.0534581207, 0.0778727531, -0.039764218, -0.0038598801, -0.0683681294, -0.026613716, 0.0090146754, 0.0095529705, -0.0370642506, -0.0576819964, 0.0228533316, -0.0134999938, 0.0563942716, 0.0117555913, -0.0603254735, 0.0040616263, 0.0238491129, -0.0484929793, -0.0496378168, 0.0132019632, -0.0258946232, 0.0397567675, -0.0110558253, 0.0662020668, 2.553680662e-34, 0.0608994365, -0.0206708629, 0.0008632141, -0.0129754478, 0.0446423069, 0.0016602906, -0.0348325484, -0.0258412044, 0.0138588063, 0.0016761402, 0.0017279497 ]
http://arxiv.org/abs/2406.17990v1
http://arxiv.org/pdf/2406.17990v1
Explicit Diversity Conditions for Effective Question Answer Generation with Large Language Models
2024-06-26T00:12:08
2024-06-26T00:12:08
Vikas Yadav, Hyuk Joon Kwon, Vijay Srinivasan, Hongxia Jin
cs.CL
cs.CL, cs.AI, cs.LG
Question Answer Generation (QAG) is an effective data augmentation technique to improve the accuracy of question answering systems, especially in low-resource domains. While recent pretrained and large language model-based QAG methods have made substantial progress, they face the critical issue of redundant QA pair generation, affecting downstream QA systems. Implicit diversity techniques such as sampling and diverse beam search are proven effective solutions but often yield smaller diversity. We present explicit diversity conditions for QAG, focusing on spatial aspects, question types, and entities, substantially increasing diversity in QA generation. Our work emphasizes the need of explicit diversity conditions for generating diverse question-answer synthetic data by showing significant improvements in downstream QA task over existing widely adopted implicit diversity techniques. In particular, generated QA pairs from explicit diversity conditions when used to train the downstream QA model results in an average 4.1% exact match and 4.5% F1 improvement over QAG from implicit sampling techniques on SQuADDU. Our work emphasizes the need for explicit diversity conditions even more in low-resource datasets (SubjQA), where average downstream QA performance improvements are around 12% EM.
Explicit Diversity Conditions for Effective Question Answer Generation with Large Language Models: Question Answer Generation (QAG) is an effective data augmentation technique to improve the accuracy of question answering systems, especially in low-resource domains. While recent pretrained and large language model-based QAG methods have made substantial progress, they face the critical issue of redundant QA pair generation, affecting downstream QA systems. Implicit diversity techniques such as sampling and diverse beam search are proven effective solutions but often yield smaller diversity. We present explicit diversity conditions for QAG, focusing on spatial aspects, question types, and entities, substantially increasing diversity in QA generation. Our work emphasizes the need of explicit diversity conditions for generating diverse question-answer synthetic data by showing significant improvements in downstream QA task over existing widely adopted implicit diversity techniques. In particular, generated QA pairs from explicit diversity conditions when used to train the downstream QA model results in an average 4.1% exact match and 4.5% F1 improvement over QAG from implicit sampling techniques on SQuADDU. Our work emphasizes the need for explicit diversity conditions even more in low-resource datasets (SubjQA), where average downstream QA performance improvements are around 12% EM.
[ 0.0530610122, 0.0509117655, -0.0071520451, 0.0152883595, -0.0006215092, -0.0102417702, 0.0374410599, -0.0187421925, 0.0142333582, -0.0454719551, -0.0122885117, -0.0430126116, -0.0048163035, 0.0205946378, 0.036644727, -0.0283785388, 0.0585170165, -0.0148631949, -0.0437916815, -0.0396436676, -0.0003432177, 0.0186363067, 0.0036886733, 0.0284672044, -0.0161589775, -0.0255436525, -0.0395487659, -0.0008469417, 0.0157360379, -0.0263788961, 0.0244048033, 0.0591152124, 0.0153787769, 0.0227974337, 0.0000019254, -0.0114490027, -0.0245834608, 0.0066091577, -0.0282170679, 0.0074190614, -0.0492004044, 0.0248075761, 0.0506125018, 0.0368290879, -0.0514552742, -0.0240534935, 0.0557455495, 0.0289659016, 0.0904615745, 0.0540677086, -0.0042568371, 0.0012123366, 0.0130454814, -0.0053513059, 0.0404936597, -0.0160167851, 0.0118167298, -0.0066307727, 0.0330226831, 0.0600931905, -0.004638562, 0.0157391261, 0.0283287857, 0.010727332, 0.0038339107, 0.046330817, -0.0157847088, 0.0214019213, -0.0084097879, 0.1009919271, 0.0138485041, -0.0067003407, -0.001804598, -0.0052496572, 0.0083753597, 0.020194646, -0.0465184189, 0.0127289528, -0.0182867646, -0.0085222572, 0.0096109742, 0.016965311, -0.015726272, 0.0010076187, -0.0230136979, 0.0411918201, 0.0527733974, -0.027249625, -0.0530578941, 0.0340273306, -0.0366557315, -0.0553044714, 0.0011237339, 0.0572745316, -0.0551872775, 0.0190486182, -0.0695687681, -0.0443105698, 0.0236878023, 0.0117620071, 0.0245707128, 0.0210369769, 0.0090018865, 0.0121705243, 0.0050927196, 0.0377494842, -0.0325398035, -0.014314767, -0.0893673822, -0.0583355241, -0.0076723052, -0.0318439119, -0.0512560755, 0.0427819528, -0.0877210349, -0.0224357843, -0.0267471336, 0.0320453048, 0.0155650759, -0.0433374085, -0.0216029044, 0.068969503, -0.0601562969, -0.0060533308, 0.0065566967, 0.0108059738, -0.0646806508, 0.0260136053, 0.0341860615, -0.0860432386, -0.0261958558, 0.0221153554, 0.0462842733, -0.039320223, -0.0289610196, 0.0304578692, -0.0273766853, -0.0173084065, 0.0020306038, -0.0369412825, -0.0330970809, -0.0420293808, 0.011124826, -0.0646789968, 0.0553774498, 0.0074531869, -0.0317266546, 0.0091723269, 0.0171770807, 0.033160679, -0.0121284062, 0.0240231641, -0.0360455923, 0.0356501974, 0.0608979091, -0.0085600335, 0.0360242091, 0.0725478306, 0.0022441256, -0.0080187786, 0.0271456055, 0.02100881, 0.0198990647, 0.0149242925, 0.0474490896, 0.0374677964, 0.0128282253, 0.0553349964, 0.0133536113, 0.0323802121, 0.0157963112, 0.0504997224, -0.0528400466, -0.0084000695, 0.0369578414, 0.026213089, 0.0348591246, 0.0134164644, 0.0586033426, 0.0194093045, -0.0030994327, -0.0100830141, -0.0048874519, 0.0485500582, -0.0161179285, 0.009154005, 0.0663233101, 0.0563475601, -0.0569125861, -0.0153784305, 0.008864644, 0.0309443697, -0.0433325432, 0.0375269465, 0.0058090189, 0.0199689474, -0.0169840828, 0.0889056846, -0.0065908283, -0.0454734489, -0.0027390646, -0.017584255, 0.0583110601, 0.0346667469, 0.0337639675, 0.0031810617, -0.0542901643, -0.0071830996, 0.0003737484, 0.0407067016, 0.0551273189, -0.0209744573, 0.0101873949, -0.0541014634, -0.045463834, -0.0087210368, 0.0062620216, 0.0449720807, 0.0582035668, 0.0414122157, 0.0332914628, -0.0414914489, -0.0118932705, 0.0250993259, -0.0390555188, -0.0353534818, 0.0434030332, 0.0041400366, -0.0059756991, 0.0277944896, 0.0010493386, 0.023401361, 0.0098377103, -0.0244035944, -0.0312294867, 0.0023029395, -0.0540437438, -0.0899158344, 0.0126374671, -0.0428344756, -0.0006445438, 0.0408218727, -0.0286726523, -0.0473617949, -0.0097952057, -0.0644245669, -0.0462344401, -0.06322328, -0.0198406093, 0.0445880964, 0.0345713161, -0.0584099852, 0.0662115216, -0.0046199956, 0.0137631726, 0.0174004938, 0.0498434789, 0.0423011631, -0.0732683167, 0.0142824491, -0.0176314712, -0.0062865675, 0.027131794, -0.0080458792, 0.0321176201, -0.0447546355, -0.0135612898, -0.0701180398, 0.0219308939, -0.0305688605, 0.0046916027, 0.0294247586, 0.0427883714, -0.0207193047, -0.0139111625, -0.0816728249, -0.0140360715, 0.0506451987, 0.0479646884, 0.0326656066, 0.0115826624, 0.0275657754, -0.0240378827, -0.0182516482, 0.0093443394, 0.0480444245, 0.0168554243, 0.0190630648, -0.058574222, -0.0002431811, -0.0412091166, 0.0179400276, 0.0257201325, -0.0087997504, 0.026012525, -0.0251376517, 0.0494988114, -0.0068937652, 0.0381033309, -0.0345827527, 0.0157672372, 0.0159881432, -0.0642849505, 0.0407228842, -0.0471341647, 0.0180073865, 0.0260796919, -0.0488596559, 0.0345398225, 0.0035585021, 0.0271733515, -0.0400114134, -0.0393905044, 0.0537228957, 0.015801318, -0.05693857, -0.0410947353, 0.0022425242, -0.0040636719, 0.0215628892, -0.0261215735, -0.0157792699, 0.0266693179, -0.0147983329, 0.0108489897, -0.0863435492, 0.0199422464, -0.0563268587, 0.0214697588, -0.029327495, 0.0261789411, -0.0027084905, -0.0132596102, -0.0385823771, 0.0059441524, -0.0362556837, -0.0379546583, 0.0481094122, -0.0193669628, 0.0140149854, -0.0173013452, -0.0272018518, -0.0919629782, -0.0412457325, 0.0373276845, -0.0084924232, 0.0226314776, 0.0005752966, 0.0167494416, 0.0203052703, 0.0269374251, 0.0357538313, -0.0729445368, -0.0016061374, -0.0275047421, 0.105637826, 0.0987448618, 0.0875712484, 0.038104333, -0.026924435, -0.0050706142, 0.0295116566, 0.0251121987, -0.0033765945, 0.027070608, 0.0021576623, 0.0154772131, -0.0114653222, -0.0108598424, 0.01000378, 0.0764241591, -0.015674755, -0.0656859577, 0.0759274364, 0.0550321266, 0.0025349006, 0.0045858081, -0.0206475854, 0.0382581614, 0.0241146069, 0.0630894676, -0.0722446442, -0.0121763907, -0.0458384231, 0.0161894336, 0.0011744575, -0.019026367, -0.0332885049, -0.1438507587, 0.039413698, -0.0215586927, -0.0929550603, -0.040623989, 0.0096108764, -0.0099996598, 0.0300420411, 0.0202330351, 0.007129651, -0.0037210246, 0.0389875881, 0.0333061181, 0.0300862323, -0.0366555192, -0.0064125373, -0.0404822603, 0.0328742377, 0.0449577346, 0.0204692893, 0.0152662629, -0.0074653011, -0.0063094269, -0.0729187354, 0.0107264128, 0.0009818185, 0.0312433261, 0.0112646688, -0.017233165, -0.0072539803, 0.031030545, -0.0162675809, 0.0059081758, 0.0493135564, 0.0502024218, -0.0259022079, 0.0240883175, -0.0337830745, 0.0297613423, 0.0612232164, 0.0681609437, 0.0433743596, -0.0349558443, -0.0235112272, -0.0257578865, -0.0086954534, -0.1073597148, 0.0002332186, 0.0400695913, 0.0090532573, 0.0245627463, -0.0964707732, -0.0685945079, 0.0238110926, 0.0489935502, -0.02683069, 0.0057306364, -0.0363701992, -0.0132570602, -0.0060325596, 0.0233763345, -0.0359759815, -0.0132443374, 0.0061818422, -0.0170520041, -0.0234101135, -0.0899116322, -0.0086868275, 0.0156075889, -0.0815370902, -0.0378280692, 0.004487168, 0.0286342856, 0.0134482337, 0.0461856425, -0.0255871341, -0.0053676176, 0.0612844527, 0.068800956, -0.0192952286, 0.0254726596, -0.0068205809, -0.0092860404, -0.0111690192, 0.0399128869, -0.0446558073, 0.0064449096, 0.0197145417, 0.073946178, -0.0131354332, -0.0291188098, 0.035129901, 0.0257511195, -0.0302220937, -0.0115850251, -0.0038965885, -0.026496077, 0.0230461322, -0.033952672, -0.0113464165, -0.039064128, 0.010100781, -0.0083517935, -0.0199619941, 0.0062646372, -0.0023646364, -0.0063265781, 0.0690679923, -0.050228145, -0.033278726, 0.0030320177, 0.001510535, 0.0115155885, -0.0123684471, -0.0282167662, 0.0185406487, -0.0011166735, 0.043470528, 0.0041613155, 0.0347707011, 0.0019085949, -0.0091408417, -0.0415330864, 0.016941553, -0.0742724836, -0.07243938, -0.0369149372, 0.0346877836, 0.0253810864, 0.0103457272, 0.0236703325, -0.0061296537, -0.0083411885, -0.0100699756, -0.0120233968, 0.021856891, -0.0116384272, 0.0411902703, 0.0110380929, -0.0106624514, -0.0088590775, 0.0812616423, -0.0064214249, 0.0511819385, -0.0174728259, 0.019605618, -0.0345122702, -0.0195047054, 0.0014478412, 0.0053551388, -0.02322427, 0.0189490244, -0.0373502523, -0.0357153788, 0.0129085947, 0.002010108, -0.0143225808, 0.014859858, 0.04471967, 0.0454169288, -0.0184569191, -0.0189053603, -0.0467279181, 0.043447569, -0.051324185, -0.0213434491, 0.0154632973, -6.651967513e-33, -0.0331507251, -0.0500569791, 0.0573255271, -0.0940835252, -0.0286210887, 0.0241768956, -0.0121190315, -0.034950316, -0.0316402875, -0.025791185, -0.0195830874, 0.0014050745, 0.0079515483, 0.0513706394, 0.0048636664, -0.0229102038, 0.0164986178, -0.0074312217, -0.0117686968, 0.0451650247, 0.0423989929, 0.0023503057, 0.003181247, 0.008902166, -0.0166636873, -0.0494484641, -0.0060763108, 0.0144310053, -0.0539239496, 0.0659452677, -0.0460090451, 0.0343869068, 0.0198804904, -0.0480297655, -0.0146768959, 0.0201270059, -0.044399552, -0.0522828288, 0.0007706884, -0.0223900583, -0.0604008585, -0.0405120812, 0.002678246, -0.0164377596, 0.0251342952, -0.012099077, 0.0223625004, -0.0183425359, -0.0132869584, -0.0248378478, -0.043551743, 0.0068237875, 0.0122622671, -0.0007968259, -0.0841102377, 0.0781189203, 0.0082476158, 0.000679217, -0.0122822728, 0.0353713706, 0.0936826766, 0.01487354, 0.0311314203, 0.0250703972, -0.0014112268, 0.012275232, 0.0190963801, -0.0321400911, 0.0058313254, -0.0090682013, -0.02950521, 0.0016623271, 0.0368085504, 0.0859927312, 0.0077635739, 0.0359817445, -0.0672919601, 0.0565241501, -0.0221266542, -0.0393166505, 0.0242878739, -0.0464623831, -0.0334715247, -0.0006599742, -0.0154920854, 0.0436373167, -0.0270680916, -0.0301875882, 0.0684142858, 0.0227661747, -0.0452352166, 0.0054283454, 0.057088647, -0.0257391427, -0.0174900852, 0.0203993395, -0.0283047874, -0.0087562148, -0.0049298164, 0.0349618606, -0.0478076451, 0.0036312013, 0.0562878661, 0.0008305998, -0.049509801, 0.0462689362, 0.002805514, -0.0034065067, -0.058099024, 0.0127477134, -0.0192648694, 0.037087936, -0.0303488728, 0.0087549295, 0.0164380968, 0.0002151873, -0.0058909692, 0.0450808406, 0.006242868, -0.0012415769, -0.005818645, -0.0513414182, -0.0011142647, 0.0440020636, -0.0247980338, 0.0610910319, -0.0342252813, -0.0385734588, -0.0303929541, -0.050996542, -0.0133757489, 0.0235907342, 2.614e-7, 0.0431904644, 0.0069803088, 0.0320523456, 0.03210181, -0.0012004117, 0.009846014, 0.0322834328, 0.0462692752, 0.0076160207, 0.0426460281, 0.0046580043, 0.035354007, 0.0156507399, 0.0269850958, -0.0675493628, 0.0447113849, -0.053702224, -0.0044513317, -0.063694492, -0.0449389368, 0.0583842434, 0.0731575787, -0.03163369, -0.0237789769, -0.0181944873, -0.0068475204, 0.047232613, 0.0470807068, -0.0516518764, 0.0437795967, 0.0434447117, -0.0162678901, 0.0195061285, 0.0114783077, -0.0154687818, 0.0110717881, 0.0410962887, 0.0354579985, 0.0142848538, -0.0670009404, -0.013501591, -0.0182437208, 0.0045412784, -0.0613601655, 0.0740704387, -0.036973536, 0.0181722492, -0.009919499, -0.053092774, 0.0039987117, 0.0149797136, -0.042166464, -0.0072176321, -0.0064140158, 0.0048099747, 0.0053029768, 0.0074961428, -0.0396782644, -0.0034648245, -0.0114563005, -0.0429439768, -0.0405109636, -0.039873451, 0.0043139425, 0.027136052, -0.0451421775, -0.0069672358, 2.520492691e-34, -0.0003452345, 0.0166952964, 0.0127411084, 0.0021131122, -0.0080768829, -0.0243097823, -0.0190507099, -0.0148571748, 0.0275195576, -0.0675269365, -0.0380280018 ]
http://arxiv.org/abs/2406.17987v1
http://arxiv.org/pdf/2406.17987v1
Multi-step Knowledge Retrieval and Inference over Unstructured Data
2024-06-26T00:00:45
2024-06-26T00:00:45
Aditya Kalyanpur, Kailash Saravanakumar, Victor Barres, CJ McFate, Lori Moon, Nati Seifu, Maksim Eremeev, Jose Barrera, Eric Brown, David Ferrucci
cs.CL
cs.CL, cs.AI
The advent of Large Language Models (LLMs) and Generative AI has revolutionized natural language applications across various domains. However, high-stakes decision-making tasks in fields such as medical, legal and finance require a level of precision, comprehensiveness, and logical consistency that pure LLM or Retrieval-Augmented-Generation (RAG) approaches often fail to deliver. At Elemental Cognition (EC), we have developed a neuro-symbolic AI platform to tackle these problems. The platform integrates fine-tuned LLMs for knowledge extraction and alignment with a robust symbolic reasoning engine for logical inference, planning and interactive constraint solving. We describe Cora, a Collaborative Research Assistant built on this platform, that is designed to perform complex research and discovery tasks in high-stakes domains. This paper discusses the multi-step inference challenges inherent in such domains, critiques the limitations of existing LLM-based methods, and demonstrates how Cora's neuro-symbolic approach effectively addresses these issues. We provide an overview of the system architecture, key algorithms for knowledge extraction and formal reasoning, and present preliminary evaluation results that highlight Cora's superior performance compared to well-known LLM and RAG baselines.
Multi-step Knowledge Retrieval and Inference over Unstructured Data: The advent of Large Language Models (LLMs) and Generative AI has revolutionized natural language applications across various domains. However, high-stakes decision-making tasks in fields such as medical, legal and finance require a level of precision, comprehensiveness, and logical consistency that pure LLM or Retrieval-Augmented-Generation (RAG) approaches often fail to deliver. At Elemental Cognition (EC), we have developed a neuro-symbolic AI platform to tackle these problems. The platform integrates fine-tuned LLMs for knowledge extraction and alignment with a robust symbolic reasoning engine for logical inference, planning and interactive constraint solving. We describe Cora, a Collaborative Research Assistant built on this platform, that is designed to perform complex research and discovery tasks in high-stakes domains. This paper discusses the multi-step inference challenges inherent in such domains, critiques the limitations of existing LLM-based methods, and demonstrates how Cora's neuro-symbolic approach effectively addresses these issues. We provide an overview of the system architecture, key algorithms for knowledge extraction and formal reasoning, and present preliminary evaluation results that highlight Cora's superior performance compared to well-known LLM and RAG baselines.
[ 0.0547376052, 0.0854258835, -0.0337433703, 0.0290511735, -0.0527970344, 0.0081210472, 0.012830548, 0.0098395869, 0.0214371961, -0.0001060583, 0.0186451171, 0.0368270725, -0.0349054821, 0.0707079247, -0.0001281729, 0.0003959525, 0.0402353816, -0.008671985, 0.0948852599, 0.0035981324, -0.0186181739, 0.0211366247, -0.0245229844, 0.0005476804, -0.0321310945, -0.0043257466, 0.0054992535, -0.0295273419, 0.021213131, -0.0585470535, -0.0295771379, 0.0357356854, 0.0553317964, 0.0432677194, 0.0000023676, -0.0713421181, -0.0307889767, 0.0060487217, -0.021569455, -0.0488511957, 0.0136741446, -0.0245670211, 0.0159127079, -0.0033067209, -0.0595306531, 0.0330905728, 0.0293924864, 0.0203326438, 0.009973756, 0.0413611904, -0.0193594787, -0.0185403079, 0.0126509154, -0.0257878583, -0.0034964925, -0.0050062705, 0.0526322536, 0.0162678696, 0.0289292429, -0.0253446475, 0.02777512, 0.0385684967, 0.0180083588, -0.0013745697, -0.0262425523, 0.0385890268, -0.0553750768, -0.0252101067, -0.0272809118, 0.0695366785, 0.0473401062, -0.0119740786, 0.0375561267, 0.0204501208, -0.0239114426, 0.0042357487, -0.0395418294, 0.0648564771, -0.0405076407, -0.0006655681, 0.0121547049, -0.0028605496, 0.043756716, 0.0348423421, -0.041751679, 0.0926436111, -0.0054105772, -0.0288318768, 0.0222924352, -0.0245620254, -0.0348451473, -0.0216946639, 0.0509579144, 0.0562036894, -0.0342714414, 0.0422825292, 0.0017839674, -0.0432254337, 0.0465681627, -0.0356919952, 0.0099388771, 0.0298810434, 0.0226010662, 0.0554892831, -0.0325298198, -0.003487291, -0.0353585109, 0.0483123064, -0.0086039081, -0.0109226825, -0.0432640351, -0.0207987241, -0.029045118, 0.0467319936, -0.0473911799, -0.0281993169, -0.0235431772, -0.0090439525, -0.0379874595, 0.0568229295, -0.0083558289, 0.0370266326, -0.0024443937, -0.0251227543, -0.0643755794, 0.0050626802, -0.0439815894, 0.0298230909, 0.0084225014, -0.0070717637, 0.0070896107, -0.0024412393, 0.0060928175, -0.0196259618, 0.023902107, 0.0795681328, 0.0310226548, -0.0149018513, -0.0436613485, -0.0510788187, 0.0038982495, -0.1013061255, 0.0133276097, -0.0576060787, 0.0774914548, 0.012530541, -0.0486284606, -0.0819399208, 0.0302198827, 0.0028593377, 0.0124184974, 0.0391380191, -0.0135313813, -0.0346768089, 0.0364396647, 0.0164167397, -0.0052892449, 0.0252821632, 0.0534502938, 0.0325381607, 0.0260965712, 0.0277842674, -0.0385482647, 0.0027196582, 0.0215702076, -0.0305156223, -0.0132835694, 0.0204050113, -0.0156292114, 0.0458736047, -0.0271678306, 0.0555812232, 0.0026516444, -0.0193547606, 0.0451880917, 0.0766318887, 0.0267633591, 0.010208738, 0.0354150645, -0.0188897084, 0.0079920515, -0.0770289004, -0.0319732651, 0.0345995314, -0.0465957113, -0.0019388386, -0.016770428, 0.0697334558, 0.0022204339, -0.0574811958, -0.0209608022, 0.0174013376, -0.0064875102, 0.0101063494, 0.0759987012, -0.0379583947, -0.0074168416, 0.0140534872, 0.0406111553, -0.0185160898, -0.0238365401, -0.0209343825, 0.0614645183, 0.0510397255, 0.0670102537, -0.0557249524, -0.0190232862, 0.0052643986, -0.0407414772, 0.0559484325, -0.000776578, -0.0015091691, -0.0138610722, -0.032044176, -0.0129644666, 0.0141975004, 0.0130795697, 0.0363994651, 0.023938559, 0.0307804663, -0.0102859763, -0.0173770674, -0.0060433024, 0.044358436, -0.0564055443, -0.0261829887, -0.0086501697, 0.0168624483, 0.0462343469, 0.0100699347, 0.0068465672, 0.0325674824, 0.0239543654, -0.0356334522, -0.0534889065, 0.0192057062, 0.0016848979, -0.0166120995, 0.0315393806, -0.0277978536, -0.0107721714, -0.011847862, -0.0458858721, -0.0221745037, 0.0462213196, -0.0022796111, -0.0069397087, -0.0450026132, -0.0253006648, 0.0374682955, -0.024341736, -0.0197136551, 0.0889435634, 0.0173611026, -0.0264273975, 0.0465763099, -0.0258968491, 0.0067238975, -0.043888256, -0.0024851442, -0.0149156917, -0.0096503105, -0.0125521729, 0.0035463579, -0.0490354635, -0.0368622541, 0.0040828241, -0.0473194346, 0.0563888401, 0.0179518685, 0.0172901787, -0.0238203406, -0.015747454, -0.009922076, -0.005462708, -0.0228109658, 0.0307270903, 0.0211954527, -0.0062923632, 0.0326439254, 0.0520510897, -0.01175295, -0.0346674211, 0.0009614346, -0.0682138354, 0.0293992404, 0.0532974042, -0.0086808382, -0.0145055605, 0.0077874046, -0.0706184655, 0.0565316342, -0.0255862884, 0.022763744, -0.001136633, -0.0353362486, -0.0273304209, 0.057208471, -0.0013305825, 0.0151899867, 0.0489575788, -0.0084061958, -0.0569293573, 0.0448623523, -0.0413210057, 0.0516541824, 0.0066468739, -0.0021901249, 0.018717818, 0.0566455275, 0.0083392346, 0.0018085899, -0.0332519189, 0.0348922275, 0.0101560429, -0.004497075, -0.028528804, 0.0046436135, 0.0267857611, 0.0616455376, -0.0231183488, -0.043362543, 0.1052986309, -0.0018034246, 0.0441192612, -0.0448093005, -0.0281067807, -0.0683065727, 0.0725930333, -0.0061815991, 0.0156779308, 0.0138101485, 0.0161093641, -0.0092224432, -0.0124125173, 0.0074402741, -0.0426198058, 0.0285509303, -0.0678559989, 0.0395498425, 0.0083548399, -0.0568798073, -0.031190969, -0.004213592, -0.0335373022, 0.0192272998, 0.0235692319, -0.0215705652, -0.0490427092, -0.0014548376, 0.0220408905, 0.0312352907, 0.0015859471, -0.0106881279, -0.0399393961, 0.0013029908, 0.0709342584, 0.0036108815, 0.0348611511, -0.009740456, 0.0080153095, 0.0485433638, -0.0110663101, -0.0116968555, -0.0438998342, 0.0075559183, -0.0093924887, -0.0416169725, 0.0428087413, 0.0385740139, -0.0406715535, -0.0742702708, -0.007394908, 0.0761131272, 0.0352759622, -0.0309431199, 0.0003025019, -0.0358155482, -0.0055388738, 0.0117005371, 0.0716324002, -0.1264135391, -0.0283010565, -0.0056134542, -0.0029930635, -0.0093891099, -0.067657873, -0.023523014, -0.0946722478, 0.0629123524, -0.0140891531, -0.0296524148, -0.0131232413, -0.0429350957, 0.0102127185, -0.0086850571, -0.0462664552, 0.0148763582, -0.0057946052, -0.0173455607, -0.0129566835, -0.0083648516, 0.0034291144, 0.023324037, -0.0490252301, 0.0084551424, 0.0149337687, 0.0327901319, 0.0010292727, -0.0257219784, -0.0325477943, 0.0165473111, 0.0229447465, -0.03280003, 0.0164874867, 0.0176853202, 0.0112559674, -0.0543240309, 0.0579084121, -0.0605428554, -0.0341311209, 0.0280201957, 0.0832171589, -0.0442234837, -0.0108309835, -0.0505948104, 0.0073933462, 0.0079747997, 0.0749880224, -0.0469385944, 0.0136393085, -0.0714437962, 0.0087821633, 0.0056835744, -0.0382150151, 0.0034256403, 0.0118035842, -0.0211513545, 0.0298863165, -0.0664453879, -0.0063617188, 0.0414402969, 0.0334760435, 0.0153930662, -0.0145611726, 0.0422250256, -0.0191964097, -0.0600427836, 0.0585877225, -0.0462495759, 0.015611128, -0.0053467965, 0.0232133195, 0.0010323859, -0.0523986146, -0.0339565836, 0.0388649069, -0.04526316, -0.0445110314, -0.043117594, 0.0460463092, 0.0250714254, 0.0228428394, -0.0237325076, 0.0261828415, 0.0422754139, 0.0511418544, -0.0285122581, -0.0348163061, -0.0177820567, 0.0118920961, -0.0173134133, 0.0036455342, -0.0332570858, 0.0292309765, -0.0108445249, 0.0760593712, -0.0394411981, -0.0156475008, 0.0299746599, 0.0399855301, 0.0334880948, 0.010778592, -0.0339155644, -0.040690355, 0.0041702571, -0.0296779145, -0.0157437548, -0.0192845389, 0.0110376356, -0.0169055909, -0.0047302204, -0.0033128762, -0.0045689228, 0.0286041442, -0.0087329708, -0.0253283437, 0.0044533554, 0.031515643, -0.0017267553, 0.0187345762, -0.0367388614, -0.0038888673, 0.0353125967, -0.011104282, 0.0742357597, -0.033319924, 0.0650923848, 0.0369057097, -0.0128778461, -0.0619389862, -0.0008635251, -0.0378167443, -0.0407073125, 0.015893966, -0.0126463957, 0.0170055218, 0.0262607653, -0.0015021291, -0.0767567605, 0.0037841368, 0.0179318972, 0.0030156577, -0.0156403687, -0.0253246352, 0.0512056574, 0.0284531265, -0.0000649909, -0.024641633, 0.0372448303, -0.0236022063, 0.0589501783, -0.0272624604, -0.0470021442, -0.0249033738, -0.0628419891, -0.0027885807, -0.0126282806, -0.0332606956, -0.0307663325, -0.0785877034, -0.0354328007, 0.0112994965, 0.0513794273, -0.0248699524, -0.0074071344, 0.023637224, 0.0441474095, 0.0194721092, -0.0625394732, -0.0055389437, 0.0184483472, -0.0689847022, -0.0252696276, -0.0079479404, -6.791372327e-33, 0.0130710499, -0.03828419, 0.0809435099, -0.0110761141, -0.0098986859, -0.002167704, -0.0593787879, -0.0342853256, 0.0009691204, 0.0200186707, -0.0049003651, 0.005422072, -0.001445792, 0.0084618498, 0.0226742066, 0.0106425956, -0.0204497334, 0.0352143571, -0.028453121, -0.0244801585, 0.0785466656, 0.0190097447, -0.0059121228, -0.0346766114, -0.007724646, 0.0243276302, 0.0174412634, 0.0340364911, -0.0645813271, 0.0407300554, -0.0101923738, 0.0258111954, 0.0024775141, -0.0236449465, 0.0332740098, 0.0274232868, -0.0840241686, -0.0444226488, 0.0111279497, -0.0506724082, -0.0416827388, -0.030240085, 0.0400056764, -0.0010285885, 0.01250224, 0.0053928718, 0.0367382243, -0.0486509465, 0.0081420746, -0.114199169, -0.0855216309, 0.0262517743, -0.0205886438, 0.0320828333, 0.0006067313, 0.0238938667, -0.0118422089, 0.0471678413, -0.0479487516, 0.0242847055, -0.0284699257, -0.0248207748, 0.0558128431, -0.0242676698, 0.0052280049, 0.0468506366, -0.0355814993, -0.0189149715, -0.0497307926, 0.0236116834, -0.0098651014, 0.107654199, 0.0261196047, 0.0307499077, 0.0303022321, 0.0294834245, -0.1082524881, 0.0476660207, 0.0317235291, 0.08439105, 0.036906641, 0.0196230225, -0.037537951, 0.0009742327, -0.0187399015, 0.0134236524, -0.0467226431, -0.0261717588, 0.019823378, -0.0004527015, -0.0168711487, 0.0286464114, 0.0021507193, -0.0259719342, -0.0009649115, 0.0240014493, 0.0439795181, -0.0047285687, -0.0247691516, 0.0284524299, -0.050279323, -0.0363902301, 0.0035576972, -0.0136157926, 0.0319760777, 0.0216443632, 0.0243108347, 0.0008323431, -0.0044519771, -0.0056354254, -0.022993926, 0.0109997774, 0.0067858826, -0.0217049941, 0.0199651606, 0.0421175398, 0.0071484828, -0.0081675397, 0.0232443623, 0.0312525705, -0.0149428593, 0.0368981101, 0.0138173671, -0.0082475217, -0.0121194823, -0.002096964, -0.0422616079, 0.0421558134, 0.0035224955, -0.07387916, 0.0134036466, 0.0567005761, 3.23e-7, 0.0325917751, 0.0573463328, 0.0328045599, 0.0228127856, 0.0220563449, 0.0052253036, 0.0010233565, -0.0171028953, -0.0056917085, -0.0333466455, 0.0332317427, 0.0535753742, 0.0224537626, 0.058403071, -0.043956276, -0.0019664899, -0.0350101218, -0.0577229112, -0.0876759216, 0.0017213298, 0.0493855476, 0.049650602, 0.0202343613, -0.0237879083, 0.0072722794, -0.0394338034, 0.0472239628, -0.0167448483, 0.0352594554, 0.0157568417, 0.0175441168, -0.0366861634, -0.0050003896, 0.0442519225, -0.0031119408, 0.0284029972, -0.0428463295, 0.0713870451, 0.0007633721, 0.0088122999, 0.0092181247, -0.0339840204, 0.0142407333, -0.0381434485, 0.0706655756, -0.0686179698, -0.0493355319, -0.0008391235, -0.0242491569, -0.0229744241, -0.0167208519, -0.0038008294, 0.025843367, -0.0066648717, -0.0129376035, -0.0064710751, -0.0109182587, -0.090070799, 0.01490629, 0.0342406966, -0.058937002, 0.0040270886, -0.0229378473, 0.0065926658, 0.032893002, -0.0876318514, 0.0153091112, 2.952581915e-34, 0.0028161239, 0.0147438142, -0.0181817673, 0.013379978, 0.0632762387, 0.0161308814, -0.078224808, 0.0347586609, 0.0048387023, -0.1020809785, -0.0116387336 ]
http://arxiv.org/abs/2406.17982v1
http://arxiv.org/pdf/2406.17982v1
EDEN: Empathetic Dialogues for English learning
2024-06-25T23:36:16
2024-06-25T23:36:16
Li Siyan, Teresa Shao, Zhou Yu, Julia Hirschberg
cs.CL
cs.CL
Dialogue systems have been used as conversation partners in English learning, but few have studied whether these systems improve learning outcomes. Student passion and perseverance, or grit, has been associated with language learning success. Recent work establishes that as students perceive their English teachers to be more supportive, their grit improves. Hypothesizing that the same pattern applies to English-teaching chatbots, we create EDEN, a robust open-domain chatbot for spoken conversation practice that provides empathetic feedback. To construct EDEN, we first train a specialized spoken utterance grammar correction model and a high-quality social chit-chat conversation model. We then conduct a preliminary user study with a variety of strategies for empathetic feedback. Our experiment suggests that using adaptive empathetic feedback leads to higher perceived affective support, which, in turn, predicts increased student grit.
EDEN: Empathetic Dialogues for English learning: Dialogue systems have been used as conversation partners in English learning, but few have studied whether these systems improve learning outcomes. Student passion and perseverance, or grit, has been associated with language learning success. Recent work establishes that as students perceive their English teachers to be more supportive, their grit improves. Hypothesizing that the same pattern applies to English-teaching chatbots, we create EDEN, a robust open-domain chatbot for spoken conversation practice that provides empathetic feedback. To construct EDEN, we first train a specialized spoken utterance grammar correction model and a high-quality social chit-chat conversation model. We then conduct a preliminary user study with a variety of strategies for empathetic feedback. Our experiment suggests that using adaptive empathetic feedback leads to higher perceived affective support, which, in turn, predicts increased student grit.
[ 0.0564788394, -0.0215368364, -0.0513057634, 0.008278613, -0.0237103496, 0.026714839, -0.0166413356, -0.036661759, 0.0356718153, -0.0169675723, -0.0084185377, 0.0430875868, -0.01544022, -0.0126243765, 0.0290091466, -0.0777556524, 0.0334555842, 0.0219691377, 0.0342847556, 0.0116442209, 0.0492649898, 0.0249298271, -0.0444328189, 0.066154398, -0.043109104, -0.0337239355, -0.0561996549, 0.0125661576, 0.0076222112, 0.0039332486, 0.0914875716, 0.0344046764, -0.0192183312, 0.0212728884, 0.0000018665, -0.0164835136, -0.0201732162, -0.0526886359, -0.0365568399, -0.0126293264, 0.0490809195, 0.0427362174, -0.0251450464, 0.045106709, -0.0393503793, 0.0214966331, 0.0467140935, -0.0209583063, 0.0823878646, 0.0539714061, -0.0134119261, -0.0008736208, -0.0026534644, 0.0257709362, -0.0181074739, 0.0156580377, -0.0258032884, 0.0242531002, 0.0417498164, -0.0210035462, -0.0293520633, 0.0304586943, 0.0095069902, -0.0571753457, 0.0310148876, 0.0164618567, -0.0469942428, -0.0482620709, -0.005047326, 0.0418221466, 0.0388770625, 0.0131202657, -0.0099044628, 0.011180439, 0.0751108453, -0.0173210632, -0.0587668903, -0.0049295775, 0.0074946433, 0.0164259728, -0.0055709826, -0.0562573746, -0.0657513142, 0.0239174645, 0.0568576977, -0.0421753787, -0.0110133896, -0.0209218767, -0.0290533956, 0.0470228121, 0.0078928154, -0.0650333464, 0.0390589684, 0.0362735465, -0.0370929725, 0.0258942582, -0.0348221399, -0.0123223849, -0.0160515029, -0.0114108883, -0.0498912893, 0.0714417771, -0.014568897, -0.0003298833, 0.0450954735, 0.0253177993, 0.0830123425, 0.0596612021, 0.0053874631, 0.0034592743, -0.0429235138, -0.0103551168, -0.0386221968, 0.044686269, 0.0115276901, 0.0000449816, -0.0817890391, 0.0089962892, 0.0037181862, 0.0258402806, -0.0306034442, -0.0245142244, -0.0238554925, 0.0149774775, -0.0218389798, 0.0315764323, -0.0377820432, -0.0088547952, 0.0339911059, -0.0709913447, -0.0003295966, -0.0045422902, 0.0186623577, -0.0159672983, -0.011877995, 0.0051234327, -0.0113113523, -0.0545127168, -0.0057417946, 0.0169437323, -0.0053945514, -0.0450939611, -0.0107799526, -0.0176504031, -0.0247262102, 0.0017510342, 0.0112558007, -0.0448792614, -0.000425347, 0.0200402029, 0.0040388317, 0.0219603498, -0.0428374037, 0.011633887, 0.0495830774, 0.0064441245, 0.0305597596, 0.0208273288, 0.0235351697, -0.0247030612, 0.029403124, -0.0394581221, 0.0183226112, 0.0179700777, -0.0156194679, -0.0408340544, 0.0299191084, -0.0463710278, 0.0072407988, 0.0254348535, -0.0004082256, 0.0243824422, -0.0238605607, 0.0199795291, 0.054074306, 0.0681784898, -0.0073594199, 0.0610523969, -0.0390451252, 0.0034135298, 0.0399146974, 0.0267265085, -0.0178519133, 0.0134427501, -0.0036244479, -0.0336094126, 0.0157039352, -0.0126345856, -0.0032860006, -0.052704528, -0.0296007581, -0.0227740575, 0.0365266688, 0.0276346561, 0.0213716719, -0.0097026881, -0.0043154326, -0.0157566462, 0.0438312441, -0.101958558, -0.0194501374, -0.0519425496, 0.015638886, 0.1071770191, 0.035286732, -0.029469315, -0.0550494827, 0.0066538216, -0.0087734759, 0.0210061651, 0.066838555, -0.0208917353, 0.0284856092, -0.0168863218, 0.0212440137, -0.0214819964, 0.0013293088, -0.030091919, -0.0108777098, 0.0374858975, -0.0405832976, 0.0538502559, -0.0447790101, 0.0093179317, -0.040477749, -0.0122251641, -0.0416212752, 0.005502949, -0.0147760659, 0.0271842703, -0.004313672, -0.0288944896, -0.0160953645, 0.0297605451, 0.0037172853, 0.074889794, -0.0219713729, -0.0451899245, 0.0592200346, -0.0001054993, -0.0229461975, 0.029420007, 0.0178378932, -0.0049099945, 0.0603109114, -0.0543277003, 0.0909997672, 0.0126353139, 0.0563501418, -0.0772295743, -0.0007500852, -0.0395071581, -0.0435280427, 0.0425683074, -0.0097913826, -0.0159898493, -0.024411466, -0.0297476165, -0.0386876911, -0.0198755786, 0.0166025925, 0.003515068, -0.0115421163, -0.0002443074, 0.013195483, -0.1159928665, -0.0362101309, -0.1031500176, -0.0081083812, 0.0190838855, -0.0234695617, 0.0052860533, -0.0131142, -0.0291251633, -0.0336753167, -0.0363744684, 0.0595199391, 0.0175570324, -0.022620758, 0.0271203909, -0.0376583673, -0.0132367425, -0.1015908644, -0.0090440577, -0.036016278, -0.0130780833, 0.0533107147, 0.0208282545, -0.0386447348, 0.0451799482, -0.0037056028, 0.005924284, 0.0128161535, -0.0898966342, 0.0353299491, -0.0154882306, 0.0488814637, 0.072672911, 0.0176656041, 0.0516110398, -0.0046842513, -0.0460345112, -0.0621361434, 0.0285498984, 0.0251407977, -0.0136279995, -0.0192539878, -0.0519836843, -0.0410832874, 0.0078858053, 0.0298194159, 0.0118846316, -0.0496857092, 0.046614781, 0.0193399116, -0.0049420265, -0.0121930297, -0.0148187755, -0.0064256862, -0.0284114368, 0.018986512, -0.0392451175, 0.0618858226, 0.0351633802, -0.0330651253, -0.0558383837, -0.0110720498, 0.0228683352, 0.0257317238, -0.0609157197, 0.0114726443, 0.0585615709, -0.0250564702, -0.0057420661, -0.0264447909, 0.0353541933, -0.0561197959, -0.0041702311, -0.0282902457, -0.0457316488, 0.0486365743, -0.0142657673, -0.0465574004, -0.0294729378, -0.0536119379, -0.0158325639, 0.0297804028, -0.0100169061, -0.0276528187, -0.0599923655, 0.0233441032, 0.0146392025, -0.038767796, -0.0277898256, 0.0334688425, -0.0481434278, 0.0161957368, 0.0481514893, 0.0385922194, -0.0178978499, 0.112632513, 0.0346007608, 0.0511090569, 0.0162061602, 0.0185460486, -0.0193853769, 0.0359281711, -0.0003977797, -0.0346504413, -0.019167304, 0.08981096, -0.0313433595, -0.0330269076, 0.122531496, 0.0096892621, -0.0271618254, -0.0019658003, -0.0544173643, 0.0233642459, 0.0110081108, 0.0238547008, -0.012074179, -0.0219143964, 0.004712359, 0.0482772402, 0.0064340942, -0.0108873295, 0.0184632, -0.0561369471, 0.0345350988, 0.0013087798, -0.020382354, -0.0148030622, -0.0193081144, -0.0080883382, 0.0205885656, 0.0299699828, -0.0075587975, -0.0028847465, 0.0132018467, 0.0171580054, -0.0004662271, -0.0067041041, 0.0167699866, -0.0433303416, 0.0743592754, 0.0282607563, 0.0597633682, 0.019250989, -0.0041829688, 0.0166541692, 0.0671823323, -0.0123306802, -0.0083917435, 0.0604920164, -0.0141571248, -0.0059223156, -0.047565423, 0.0343698822, -0.0113792028, -0.0092747277, 0.0696450919, 0.0294408835, 0.0074362061, 0.0311263148, 0.0070873396, 0.0297819842, 0.0236786027, 0.0327891856, 0.0208634827, -0.0657058358, 0.0058181793, -0.0249357969, 0.009747724, 0.014574076, -0.0698847249, 0.007754487, -0.0238888189, 0.0272831898, -0.0560431518, -0.0658483133, 0.0565223768, 0.0150569957, 0.0241281893, -0.0357260406, 0.020301627, 0.0485585853, 0.0094673056, -0.0132209444, 0.0134502519, 0.0309371147, -0.0591974854, -0.0395735577, 0.0029807761, -0.0190644991, 0.0234612767, -0.0367095694, -0.0089009935, -0.0477945469, 0.0150522431, 0.0389588885, 0.0137528544, 0.0106067536, 0.0188565403, -0.0623757243, -0.0796657726, -0.0114354705, -0.0149734952, 0.0403998308, 0.0321270227, 0.0334325433, 0.0158888455, 0.005778884, 0.0043662195, -0.0343285389, -0.0009967895, 0.0086908424, 0.0190317612, -0.0286656842, 0.0487832129, 0.0469131172, 0.0081021991, -0.0013320541, -0.0420928597, -0.0241816398, 0.0179031584, -0.0532788634, -0.0086092353, -0.0323381834, -0.0226866826, -0.0269606691, -0.0046897945, -0.0014362662, -0.0659553632, -0.0805946216, 0.0097728958, -0.0613466874, 0.0396816246, -0.0104811424, -0.0152986525, -0.0225273389, 0.0161320958, 0.0031151418, -0.0016180385, 0.0197847411, 0.0465724207, 0.0036477856, 0.0921758711, -0.0171478186, -0.0429940857, 0.0603942387, -0.0058421497, -0.0341261066, -0.0196207929, 0.0096759805, 0.0625916198, 0.0136366468, 0.0133043434, 0.0275339466, -0.0392894261, 0.0285106655, -0.0068097105, 0.0234700255, -0.0186322015, -0.0284839496, -0.0224718563, -0.0244972315, -0.0123524787, -0.0203472134, -0.0007041281, -0.0452385619, 0.0504594445, -0.0177167505, -0.0119094141, -0.0303368066, 0.0021875063, 0.0359645113, 0.0116604446, 0.0127681717, 0.0412670039, -0.0300408453, -0.0418004096, -0.033788424, 0.028619824, 0.0192563646, -0.0596067235, 0.0112938797, 0.0350978449, -0.0636734813, 0.0017871422, -0.0497744866, 0.0325601362, -0.0466767475, -0.0075538526, -0.0059022228, -5.299945135e-33, -0.0240248386, -0.0039732517, -0.018411126, 0.0655130967, -0.0428000689, -0.0073926747, -0.0559834167, 0.0565394647, -0.0056637069, 0.0124815023, -0.0409009196, 0.0895692855, -0.0040121777, -0.0125218984, 0.0534005649, 0.0126392376, 0.0289312415, 0.0063651851, 0.0207459442, 0.0054391632, -0.010733394, 0.0249130093, 0.0695776269, -0.0616211072, 0.0419483259, 0.0122706508, 0.0038747068, -0.014943325, 0.0257366542, 0.0255358126, 0.0156368334, 0.0377628952, -0.0107090808, 0.0273977686, -0.0155117577, 0.0036932882, -0.0127690649, -0.0481447838, -0.0271786898, 0.0251949392, -0.0314509124, -0.0279981345, 0.0111376485, -0.0102895014, -0.0396951847, 0.0046908245, 0.028025968, 0.0039165793, 0.0565892309, 0.0586730465, -0.0441930071, -0.0082775494, -0.0546469651, -0.0841316655, -0.0349366739, 0.0675702468, 0.013984248, -0.0439966321, 0.0099990368, 0.059743233, 0.0018745436, -0.0039370386, 0.0279062875, -0.0318215415, -0.0127721215, 0.0645198822, 0.02262518, 0.0457891189, 0.0216036756, 0.0084578609, 0.058382567, 0.0094778836, 0.0643642023, 0.0629600734, -0.034432359, -0.0587877482, -0.0182602312, 0.0390045866, 0.1050227061, 0.0147812301, -0.0331185535, 0.0381795354, -0.0525015593, -0.0103843082, -0.0159728713, -0.0091696596, 0.0141207566, -0.0277363807, 0.0252942964, -0.0228625424, 0.0046780137, -0.0910632834, -0.0399635509, -0.0290855207, 0.0089571625, 0.0076648337, -0.0614478178, 0.0282419566, -0.0264404137, -0.0109852524, 0.0070168399, -0.0265592802, -0.0103327865, 0.0329720303, 0.0558513552, 0.0242149197, -0.0283228401, 0.0544938408, -0.0178793799, -0.0060718362, -0.023425635, 0.0586708821, 0.0125388885, -0.0610001832, -0.0210215952, 0.0150726885, -0.0090118283, 0.0314670876, 0.0203271061, -0.0094634863, -0.0744791403, -0.0539748073, 0.026955111, -0.0182357468, -0.0432325192, -0.0423611253, -0.0206789691, 0.014700071, 0.0021916223, -0.0578349307, -0.0255443472, -0.0432909094, 2.634e-7, -0.0367499255, 0.0648830682, -0.0020054486, -0.0235843565, 0.0188443065, 0.0478038453, -0.0156761799, 0.0246652737, 0.0426358022, 0.0189702958, -0.0131897591, -0.0003831361, 0.0639375448, 0.0211829729, -0.0631917939, -0.0347412005, -0.0202485304, -0.0037287192, 0.0003269569, 0.0120606543, 0.0869523287, 0.0168182515, 0.0556819327, -0.002204692, 0.0031658334, 0.0208616219, -0.0176581591, -0.028184602, 0.0134075349, 0.0199903809, -0.012257901, -0.0136131775, 0.0235431921, 0.0108574508, 0.0062626484, 0.0076048318, 0.0123647153, 0.0756592229, -0.0016766178, -0.0069036926, 0.0400609337, 0.0523346178, -0.0330973789, -0.0428187065, 0.0031096698, -0.046869006, -0.0051204539, 0.0088506546, 0.0138276191, -0.0489652045, 0.0355049074, -0.0327613689, -0.0467277169, 0.0278462768, -0.0219912156, -0.0087627079, 0.0199043918, -0.0170392022, 0.0141060632, -0.0055473475, -0.0310666114, -0.0561709739, -0.0134980101, -0.0425776504, -0.035089545, 0.0040518129, 0.0054887296, 1.444883562e-34, -0.0280931853, -0.0030028634, 0.0652540177, 0.0694048628, 0.0526831299, 0.0204667021, 0.0608579181, 0.0330101214, 0.0557263829, -0.007392887, -0.052549582 ]
http://arxiv.org/abs/2406.17975v1
http://arxiv.org/pdf/2406.17975v1
Inherent Challenges of Post-Hoc Membership Inference for Large Language Models
2024-06-25T23:12:07
2024-06-25T23:12:07
Matthieu Meeus, Shubham Jain, Marek Rei, Yves-Alexandre de Montjoye
cs.CL
cs.CL, cs.CR, cs.LG
Large Language Models (LLMs) are often trained on vast amounts of undisclosed data, motivating the development of post-hoc Membership Inference Attacks (MIAs) to gain insight into their training data composition. However, in this paper, we identify inherent challenges in post-hoc MIA evaluation due to potential distribution shifts between collected member and non-member datasets. Using a simple bag-of-words classifier, we demonstrate that datasets used in recent post-hoc MIAs suffer from significant distribution shifts, in some cases achieving near-perfect distinction between members and non-members. This implies that previously reported high MIA performance may be largely attributable to these shifts rather than model memorization. We confirm that randomized, controlled setups eliminate such shifts and thus enable the development and fair evaluation of new MIAs. However, we note that such randomized setups are rarely available for the latest LLMs, making post-hoc data collection still required to infer membership for real-world LLMs. As a potential solution, we propose a Regression Discontinuity Design (RDD) approach for post-hoc data collection, which substantially mitigates distribution shifts. Evaluating various MIA methods on this RDD setup yields performance barely above random guessing, in stark contrast to previously reported results. Overall, our findings highlight the challenges in accurately measuring LLM memorization and the need for careful experimental design in (post-hoc) membership inference tasks.
Inherent Challenges of Post-Hoc Membership Inference for Large Language Models: Large Language Models (LLMs) are often trained on vast amounts of undisclosed data, motivating the development of post-hoc Membership Inference Attacks (MIAs) to gain insight into their training data composition. However, in this paper, we identify inherent challenges in post-hoc MIA evaluation due to potential distribution shifts between collected member and non-member datasets. Using a simple bag-of-words classifier, we demonstrate that datasets used in recent post-hoc MIAs suffer from significant distribution shifts, in some cases achieving near-perfect distinction between members and non-members. This implies that previously reported high MIA performance may be largely attributable to these shifts rather than model memorization. We confirm that randomized, controlled setups eliminate such shifts and thus enable the development and fair evaluation of new MIAs. However, we note that such randomized setups are rarely available for the latest LLMs, making post-hoc data collection still required to infer membership for real-world LLMs. As a potential solution, we propose a Regression Discontinuity Design (RDD) approach for post-hoc data collection, which substantially mitigates distribution shifts. Evaluating various MIA methods on this RDD setup yields performance barely above random guessing, in stark contrast to previously reported results. Overall, our findings highlight the challenges in accurately measuring LLM memorization and the need for careful experimental design in (post-hoc) membership inference tasks.
[ 0.01495991, 0.1391713321, -0.023866253, 0.0533307642, -0.0542635582, 0.0041412972, 0.0241355281, -0.0256790835, 0.0011719387, -0.0553536527, -0.0013607751, -0.0650284439, 0.0104936063, -0.0170778129, 0.0125235692, 0.032733053, 0.0672267079, -0.0525487401, -0.0322337449, -0.0486888327, -0.0258436818, -0.00273772, 0.0031732195, 0.0042242217, -0.0041058403, -0.001017804, -0.0121407267, -0.0036130743, 0.0068711946, -0.0196913742, 0.0480940491, 0.1043960825, 0.013540579, 0.1133909523, 0.0000021565, -0.0357908495, -0.0523572527, 0.0224234033, -0.0535851344, 0.0166980904, 0.0060290699, -0.0010478731, 0.0270276424, 0.01381128, -0.0008323616, -0.0092654973, 0.0682884455, -0.0136825005, -0.0132311732, 0.0915874541, 0.0041902279, -0.0320335962, -0.0031136291, -0.0196786132, 0.0126017798, -0.0017090373, 0.031112548, -0.0489667431, -0.007227465, -0.0467259772, -0.0311766453, 0.0226882752, 0.0132078398, 0.0383017957, 0.0013345005, 0.0510805286, -0.0830036327, -0.0241864659, 0.0041467617, 0.0379234403, 0.0084159859, -0.0288546588, -0.0026249271, 0.0225157831, 0.0018720191, 0.048863776, -0.0232027527, 0.0091727199, 0.0135276131, 0.0162470061, -0.0108867232, 0.0190369114, 0.0499505587, 0.0131070521, -0.0660118237, 0.0381760597, 0.0224870592, -0.0327522121, -0.010835249, -0.0088695334, -0.0462994091, -0.0248001982, -0.0327529833, -0.0272874217, -0.013618229, 0.0281328205, -0.0310416408, -0.0290553458, 0.0252819415, -0.0380373411, 0.0344858989, 0.0143287042, -0.0158107933, 0.0291668698, -0.0199361239, 0.018323794, 0.0265417118, 0.029607391, -0.0242413413, 0.011510306, -0.033533752, 0.0032698216, 0.0400396734, 0.0304762684, 0.0295391381, -0.0012314419, -0.0522126667, 0.004419636, -0.0276105329, 0.0043300223, -0.0389201604, -0.0132744471, -0.0175734479, 0.0167322867, 0.0080686845, 0.0108873332, 0.0035146854, 0.0205393303, -0.0111287627, -0.0568747446, -0.0296672825, 0.0081891995, -0.0176285766, -0.0324772298, 0.0176969655, 0.0651756823, 0.0079064714, -0.085594818, -0.0290843714, -0.0119185178, -0.0297995079, -0.0951278508, 0.0319275558, -0.0820012763, 0.0351799615, 0.0161795691, 0.0002768301, -0.0385047905, 0.0050540329, 0.0115148546, -0.0601128675, 0.0841689408, -0.0149156218, 0.0097503336, -0.0341504142, -0.0009593507, 0.0192695316, 0.1128140688, -0.0040347902, -0.0015469565, 0.0664328858, 0.0110679306, -0.0015393613, 0.0322285369, 0.0318059735, 0.0147774201, -0.0320508815, -0.0154678058, 0.038849514, -0.0348858237, -0.0066704294, 0.0173481312, -0.0011674071, -0.0219336729, 0.0741749331, 0.0527908057, -0.0151540823, -0.003974475, -0.0206499957, -0.0075390446, -0.0098111024, 0.0148091614, -0.0837555155, -0.040608827, -0.0782747567, -0.0136373825, 0.0269213375, 0.0338871256, -0.010262033, -0.0166464467, -0.0243434105, -0.0114057558, 0.0087309899, 0.0384409837, 0.0594955459, -0.0206771828, 0.017799357, 0.0651724637, 0.0624731705, -0.0413830206, -0.0045843283, -0.0532663539, 0.0369133428, 0.0057420791, 0.0396587588, -0.0513618663, 0.0172841791, 0.0010559261, 0.0522782691, 0.0009828282, 0.0315606333, 0.0138381859, 0.0304427911, -0.0364979729, -0.0245234054, 0.0278197769, 0.0616959408, 0.0381411351, 0.0583860762, -0.0019793164, 0.0142436624, -0.0554093346, -0.0222706534, 0.0445427187, -0.0147770057, -0.0182248, 0.0209483672, -0.00845055, 0.0402959622, -0.0093183815, -0.0519536287, -0.0158930812, 0.0178419519, 0.0030784726, -0.0059359558, -0.0239686407, 0.0215654466, -0.0336886644, 0.0332037471, 0.0181977581, -0.0085316589, 0.0655523986, -0.050626833, -0.0128500154, 0.0199847072, -0.0159866456, 0.0383522846, -0.034731511, 0.0324798264, -0.0378080569, 0.0517993681, -0.0300588515, 0.050159581, -0.0226084776, 0.0256578308, -0.0018599764, -0.0233816113, 0.0128243277, -0.0435322113, -0.0802083388, 0.0511272773, -0.0224472135, 0.0152542563, 0.0193390828, -0.0310455132, -0.0136759691, -0.0144212591, -0.0545298532, 0.0221056063, -0.0395490602, 0.0022693702, -0.0243189074, -0.0212165136, -0.0369939953, -0.0159735121, 0.0106542939, -0.0023772928, -0.0294467267, 0.0533251241, -0.0194896124, -0.0180993099, -0.0398493111, -0.0201387722, -0.0216250774, -0.0032139844, 0.0947839022, -0.005453852, 0.0091437194, -0.0149619468, -0.0028045687, -0.0544527136, 0.0149988718, -0.0086061265, -0.0408658944, -0.0027945405, -0.0084777158, 0.0070740995, 0.0409065448, 0.0163955819, 0.0095172534, 0.077857241, -0.0158447046, -0.0207849685, -0.0189386476, -0.0205020402, -0.0140970154, -0.0167773422, -0.0340797938, -0.0038380856, 0.0594427548, -0.002914872, -0.0394221209, 0.0050965729, 0.014437139, -0.0279356204, -0.0042632795, -0.0194530878, 0.0135224853, -0.0291221067, 0.0223320853, -0.0107020624, -0.0852332041, 0.0404613949, -0.0408204906, 0.0212312918, -0.0334153399, -0.0299353711, 0.007362105, 0.0435012989, 0.0329707712, 0.0105363615, 0.0494821444, 0.0037992424, -0.0460895672, -0.0209094267, 0.0486699417, -0.0341773443, 0.0336736329, -0.0574233048, 0.0416497365, -0.0058468934, -0.0005394626, -0.0681163594, -0.0697681755, -0.0062506916, -0.0056885146, -0.0274014864, -0.0283289831, -0.0273562931, -0.0021036873, -0.0015257604, 0.0070073199, -0.0169824064, 0.0203540009, 0.0061098128, 0.0143610276, 0.0326999091, -0.0123851905, 0.0356006585, -0.0739572793, -0.0112447152, 0.0444169976, 0.019239787, 0.0405908339, -0.0185750779, -0.0293264296, 0.0431838334, -0.0206744745, -0.0079211574, 0.0296535846, 0.0235291142, -0.0489507467, 0.0023379175, 0.1017261297, 0.0259846654, -0.0643710569, 0.0121497195, 0.0355937146, 0.1020279378, 0.0071827322, 0.0023822756, -0.0742704868, -0.0044544763, 0.0258932319, 0.0238257982, -0.0474774092, -0.0403646752, -0.0115541779, -0.1190766022, 0.0382211134, 0.0114675248, -0.0834714845, 0.0192710534, -0.0382378846, 0.0333784632, -0.0205175895, 0.0233631171, -0.0024376742, -0.0216745064, -0.0128641985, 0.0136116268, -0.0665415376, -0.075166367, -0.0014316157, -0.0368397534, 0.0732629225, 0.0293464419, 0.0159357674, -0.0458528884, 0.0135160852, 0.0798939914, -0.0042490256, -0.0546375997, 0.0126490975, 0.0806012154, -0.0159356836, 0.0073724878, 0.0158849191, 0.0300437231, -0.0021370626, -0.0343445577, 0.0073135579, 0.0457757749, -0.0027721373, -0.0291073192, 0.0009844142, 0.004210474, 0.0046423562, 0.0815752521, 0.0210696515, -0.0176717937, -0.010781019, -0.0328085348, -0.0021048014, 0.0275471322, -0.0280857608, 0.0312147085, -0.0396079086, -0.0647671893, -0.0091535663, -0.0202350542, 0.0362888835, 0.0338893719, -0.0105435643, 0.0391446315, -0.0381163359, 0.0106418766, -0.0210878663, 0.0091589093, -0.0669350326, 0.0293328632, -0.0312477164, -0.0091618197, -0.019926589, 0.0139748324, -0.0147911143, -0.0182824377, -0.0884718895, -0.0832222477, 0.0392618626, 0.0348722711, 0.023547437, 0.0444344282, 0.0072818073, -0.0578111485, 0.0794762298, 0.0465969332, 0.0097558768, 0.0271606911, 0.0389128514, -0.0441012271, -0.0044577359, 0.0353539735, -0.0217074342, 0.0117952283, 0.0652145371, 0.0075731263, -0.0200627483, 0.0323426984, 0.007463498, 0.0064219404, 0.0564254038, -0.0376246013, -0.0114023136, -0.0318165906, 0.0307544172, -0.0195557512, -0.0219668206, -0.0005948565, -0.0448643193, 0.0296080373, 0.0284676291, -0.0302878264, -0.0412606858, -0.0267352443, 0.0735951439, -0.0257036593, 0.0318159983, 0.0553644486, -0.0089865811, 0.0565735176, 0.0066852779, -0.0088069923, 0.0117836241, 0.0082973968, 0.0428638682, -0.0022361667, 0.0454527885, -0.0049977354, -0.0339676552, -0.0497802012, -0.0044195089, -0.0647490174, -0.0209916104, -0.0025413691, 0.0411229208, 0.0242913254, -0.0394004621, 0.0068873288, 0.0073194108, -0.0084285717, -0.0465672351, 0.0270405337, -0.0135554606, -0.0217144322, 0.0167017765, 0.0190088525, -0.0259733107, 0.0484344997, 0.0750616118, 0.0391614325, 0.0187251586, -0.0313790701, -0.0165436082, -0.0155117819, -0.0289532896, 0.030125346, -0.0013845504, -0.0467752069, -0.0069815926, -0.1052562296, -0.0892194211, -0.0107136508, -0.0313801393, 0.0425611064, 0.0309128333, -0.0153984511, 0.0429465435, -0.0345430598, 0.0323495567, -0.0406452008, -0.0230102949, 0.0013946644, -0.0037830512, -0.017300155, -6.929595034e-33, 0.0019039351, -0.0175919719, -0.0171827655, -0.0206414331, -0.0133144027, 0.017124664, -0.0027594368, 0.0210410077, -0.0112997731, -0.0191682074, -0.0217128173, -0.0035920392, 0.0194829516, 0.0312840231, -0.0204556677, -0.0237383656, 0.0333090611, 0.0125716999, -0.0354376845, -0.0108551597, 0.0598989986, 0.0234230198, 0.0447140634, -0.0224507954, 0.0244465489, -0.0226320755, 0.0092110196, 0.0238160752, -0.0301663242, 0.0217242371, -0.0007899217, 0.0568087623, 0.0187035352, 0.0241580475, 0.011147975, -0.0041752597, -0.0196612738, -0.0108480519, -0.0000541412, 0.0106885759, 0.0467569195, -0.0142350998, 0.0037604438, -0.0196976718, 0.0219583418, -0.0428097174, 0.0277733207, -0.0237976033, -0.0325692594, 0.0226350464, -0.0294335056, 0.0427029803, -0.0079486696, -0.0019524444, 0.0149537455, 0.0169280693, 0.0368657224, -0.014817073, -0.0467339791, 0.0151823405, 0.0380133651, -0.0007601357, 0.0449445285, 0.0581564568, 0.0068463357, 0.0609635897, 0.0682533085, 0.050840728, 0.0237847138, -0.0103783226, 0.0686835274, -0.0107247997, 0.0378875285, 0.0640656129, 0.0843809992, -0.0229283106, -0.0771389082, -0.0294880792, -0.012001764, -0.0016501, 0.0601422824, 0.0104269162, -0.059490297, -0.0479488857, -0.0317201801, -0.0135871181, -0.0370113142, -0.034266755, 0.0293312725, 0.0167671479, -0.0502093844, -0.0444827676, 0.0663808361, -0.0499362536, -0.0145221855, 0.008145228, 0.0006532851, -0.0010589, -0.0163991284, 0.0011303345, -0.054854136, -0.0208141003, 0.0013857255, -0.0203578696, 0.0161305424, -0.0263599679, 0.0551658385, 0.0343252346, -0.0544595905, -0.0000524587, -0.022503417, 0.0211253501, 0.0256353524, 0.0490729772, 0.0386577211, 0.0126455585, 0.0143590253, -0.0076121408, 0.0438816138, 0.0143760946, -0.0423240103, 0.0500695035, -0.0570465736, -0.0047328025, -0.0057092151, 0.0053870105, 0.0023497038, 0.0112311924, 0.0411234237, -0.0670416206, 0.0131708179, -0.0404254831, 2.899e-7, 0.0242431518, 0.0030091007, -0.0298747122, -0.0163661484, -0.008567946, 0.0005993302, -0.033038348, 0.0349173024, 0.0186077673, 0.0027619295, -0.0247087516, 0.0323504955, 0.0726935714, 0.0448662899, -0.0835119411, -0.0152234258, -0.0348493569, 0.0122373858, -0.0741032138, -0.0228759535, 0.0181911588, 0.09421321, 0.0377183259, -0.0072404752, -0.0350454226, -0.0474693328, 0.0454485007, -0.0446313322, -0.0469172746, 0.0295768604, 0.0472739488, 0.0015642943, 0.0185335875, 0.0576964729, -0.0291439705, 0.0299716108, -0.0127695864, 0.010824793, 0.0285110828, 0.0067456448, -0.0428319946, -0.0197298173, -0.0062917545, -0.0414531082, 0.0790361464, -0.062370494, -0.0247872416, 0.0004401722, -0.0262178648, -0.0500947312, 0.0568547472, -0.0143041704, -0.00913426, 0.0518391356, 0.0091931149, -0.013495761, 0.0250729751, -0.060703665, 0.003857235, -0.0111993253, -0.0513580367, -0.0619453229, 0.0320819579, -0.027550729, 0.0199986454, -0.0126938969, 0.0004805911, 3.01930775e-34, 0.0397973135, 0.0009436142, -0.0269808322, 0.0339184403, 0.0547051914, -0.0365719385, 0.0672821403, 0.0188619513, 0.0017196097, -0.0710442811, 0.0120208869 ]
http://arxiv.org/abs/2406.17974v1
http://arxiv.org/pdf/2406.17974v1
Evaluating Fairness in Large Vision-Language Models Across Diverse Demographic Attributes and Prompts
2024-06-25T23:11:39
2024-06-25T23:11:39
Xuyang Wu, Yuan Wang, Hsin-Tai Wu, Zhiqiang Tao, Yi Fang
cs.CL
cs.CL, cs.CV
Large vision-language models (LVLMs) have recently achieved significant progress, demonstrating strong capabilities in open-world visual understanding. However, it is not yet clear how LVLMs address demographic biases in real life, especially the disparities across attributes such as gender, skin tone, and age. In this paper, we empirically investigate \emph{visual fairness} in several mainstream LVLMs and audit their performance disparities across sensitive demographic attributes, based on public fairness benchmark datasets (e.g., FACET). To disclose the visual bias in LVLMs, we design a fairness evaluation framework with direct questions and single-choice question-instructed prompts on visual question-answering/classification tasks. The zero-shot prompting results indicate that, despite enhancements in visual understanding, both open-source and closed-source LVLMs exhibit prevalent fairness issues across different instruct prompts and demographic attributes.
Evaluating Fairness in Large Vision-Language Models Across Diverse Demographic Attributes and Prompts: Large vision-language models (LVLMs) have recently achieved significant progress, demonstrating strong capabilities in open-world visual understanding. However, it is not yet clear how LVLMs address demographic biases in real life, especially the disparities across attributes such as gender, skin tone, and age. In this paper, we empirically investigate \emph{visual fairness} in several mainstream LVLMs and audit their performance disparities across sensitive demographic attributes, based on public fairness benchmark datasets (e.g., FACET). To disclose the visual bias in LVLMs, we design a fairness evaluation framework with direct questions and single-choice question-instructed prompts on visual question-answering/classification tasks. The zero-shot prompting results indicate that, despite enhancements in visual understanding, both open-source and closed-source LVLMs exhibit prevalent fairness issues across different instruct prompts and demographic attributes.
[ 0.0334349796, 0.0686057582, -0.0227391236, 0.0590490401, -0.010020908, 0.0270507634, 0.0650924817, -0.0026422855, 0.0166586116, -0.0732216015, -0.0180113353, -0.0317106284, 0.0204920359, 0.0088465745, 0.0060352515, -0.0591703281, 0.0245578047, -0.014722337, 0.0231562201, -0.0431994647, 0.008070698, -0.0049856342, 0.0198457316, 0.0506581329, -0.0036028423, -0.0312532485, 0.0371903405, 0.0257470645, 0.0151809882, 0.0142826103, 0.0106238807, 0.0725840926, -0.0082544759, 0.0335166268, 0.0000018072, -0.0632679015, 0.0530369319, 0.0069668381, -0.0239354484, 0.0204744898, 0.0199006125, -0.0183491651, 0.0190397203, 0.0198386423, -0.0229140315, -0.0064922455, 0.0568584092, 0.0052911192, 0.030336678, -0.036958877, 0.0031104987, 0.0221858341, -0.0434203334, -0.0122508956, 0.0164786745, -0.0044752941, 0.0066940645, -0.0236681607, -0.0514725335, -0.036245577, -0.053448081, 0.0415876731, 0.0622243583, -0.0140107153, 0.0378549844, 0.0365770124, -0.052240409, -0.0494664572, -0.0290234163, 0.0990745798, 0.0080738282, 0.020796096, -0.0362654403, -0.0034216184, 0.0406607129, -0.021161316, -0.0695559606, 0.0553763583, 0.0030910543, -0.0081771193, -0.0707323551, 0.0082181776, 0.0225801878, 0.0154323475, 0.0277338382, 0.0404217876, 0.0115958117, 0.0012194378, -0.0101436488, 0.0428697541, -0.0149708278, -0.0410204418, -0.0193962548, -0.0209955815, -0.0303349961, -0.0206369516, -0.0386712141, -0.0020989317, 0.0261065345, 0.0040434003, 0.0596298501, 0.0309859738, -0.0556280203, 0.0315256044, 0.0157690179, 0.0857142806, 0.0315569527, 0.0312976092, -0.0265038498, -0.0268162806, 0.0266244952, -0.011278403, -0.006697508, -0.0230596252, -0.0626854151, 0.0303798299, -0.0457601175, -0.0053378572, -0.0521768592, -0.0056353901, 0.0156237548, 0.0459913649, -0.018388385, 0.0108900759, -0.0565674156, 0.008896268, -0.0299273431, -0.0216869358, -0.0141739082, -0.1069306806, -0.0104404921, 0.0121514909, -0.0329019055, -0.041775398, -0.0137183154, 0.0356393866, -0.0060792291, -0.0189427696, -0.0170574766, -0.0267481785, 0.015507197, -0.0253605582, 0.0089902924, -0.0587874837, 0.0076444112, 0.005135552, 0.018340461, 0.0116879027, -0.0173870157, 0.0295999665, -0.0202025566, 0.0793792903, 0.0312239528, -0.0145792533, 0.0160050895, 0.0190465245, -0.005661143, 0.0845795572, 0.010897615, -0.0070207519, 0.057431072, -0.0011468633, -0.0003554913, 0.030655032, 0.00767919, 0.0158882458, -0.0596078001, 0.0475276001, 0.0120173218, 0.0365845598, 0.0618200637, 0.0325445011, 0.0047644041, -0.0895379335, 0.062910594, 0.0745119601, 0.0203268919, -0.0302865803, 0.0227950513, -0.0181368086, -0.0181853194, -0.0028025489, -0.0903209373, 0.0318499655, -0.0235647019, -0.034663219, -0.0169113912, 0.0516463183, -0.0347170867, -0.0361068547, -0.0060593556, 0.00674095, -0.0027643768, 0.062924467, 0.0083815167, -0.0094097089, 0.0223599765, -0.0010934435, 0.0393409878, -0.0439847223, 0.0018873016, -0.0792182088, 0.0405705087, 0.0457191281, 0.0138917444, 0.0116809169, 0.0057708276, -0.0154612353, 0.020120021, 0.0416674688, 0.066172719, 0.006565039, 0.0364862569, -0.0578527898, -0.0532931723, 0.0067212707, 0.0034847627, 0.0237794463, 0.0220624227, 0.0119747967, 0.0101040713, -0.0401932225, 0.0003021879, 0.0234966818, 0.008890436, -0.0119396234, 0.0232059024, 0.0142044462, -0.0001779277, 0.0077246078, -0.0281889439, -0.0016969034, -0.0237324797, -0.0182287917, 0.0072457073, -0.0069576972, -0.0779513642, -0.0164348893, 0.031865567, -0.0711231008, -0.039270103, 0.0512266234, 0.000832865, 0.0203003641, 0.0264289025, 0.0079702139, -0.0535439476, -0.0308209695, -0.010868663, -0.0121413507, 0.0226182416, -0.059649054, 0.03856498, -0.026581049, -0.0178347714, 0.0399161503, -0.0262085926, 0.0093421871, -0.0647469983, -0.0221448112, 0.0298661664, 0.0010644883, 0.0087380596, 0.0002889817, 0.0319648199, -0.0421216562, 0.0365383141, -0.084109053, -0.0035681031, -0.0364844315, -0.0392517895, -0.0185267255, 0.0243780762, -0.021649586, -0.0606688783, -0.0343234092, 0.0210276973, 0.0115102483, 0.0377362147, 0.0113724582, -0.0271053258, 0.0052129943, 0.0012320636, -0.0702342093, 0.0008886983, 0.0806138813, 0.0625831038, 0.0528486259, -0.0084005091, -0.0321906395, -0.0421204045, 0.0586551279, -0.0218040124, 0.0117130773, 0.0212546214, 0.0137863485, 0.0116310734, 0.1068231985, 0.0265624262, 0.0387967788, 0.0402157865, 0.0034181157, -0.0424708612, 0.009111003, 0.0248834919, -0.0059892349, 0.0056299604, -0.0421405174, -0.01742425, 0.011931859, -0.0006019801, -0.0169355068, -0.0826207399, 0.0592579208, -0.0459417775, -0.0866826028, -0.1032113209, 0.0130157135, 0.0120131699, 0.0464842431, -0.0478868224, -0.0922971591, 0.0616979748, 0.0040507475, -0.0083375741, 0.0078255162, -0.002190914, -0.0669478327, 0.079396069, -0.0173248872, 0.0181670133, 0.0126430374, -0.0268903002, -0.087497443, -0.0279744025, 0.0255466774, -0.0359101631, 0.0847363994, -0.0615315363, -0.0049042567, -0.0519163273, -0.0493792295, -0.0544973686, -0.0561290197, -0.0110780764, 0.0265696701, -0.0399936959, -0.0237028413, -0.0186327919, 0.0110772243, 0.02070136, 0.007908605, -0.0251191128, 0.0038183625, 0.0093618594, -0.0224043243, 0.048228927, 0.0476910621, 0.0246131085, -0.0501749776, 0.0279371776, 0.052608449, 0.0339142866, 0.0092810923, 0.0135768186, 0.0400258154, 0.050661359, 0.0083036274, -0.0172896963, 0.01263957, 0.0666579679, -0.033630047, 0.0469013825, 0.0975553021, 0.0186694562, -0.078656204, -0.0078162756, -0.0238763299, 0.0378652997, 0.0346993618, -0.0155477421, -0.0972157717, -0.0146742463, 0.0098639568, 0.0429342613, -0.021236239, -0.0372732095, -0.0048348214, -0.0970656574, 0.0537361577, -0.0057405923, -0.027169954, -0.0313845053, 0.0390648656, 0.0063379086, -0.0169648174, 0.0524849668, -0.0092055975, -0.0031547849, 0.0194176435, -0.0088476567, -0.0011301194, -0.0007912295, 0.0705893859, -0.0597704612, -0.0035931861, 0.0552276969, -0.0148934983, 0.0169517715, -0.0604237951, 0.010144921, -0.0451996215, 0.0213876311, -0.0285834745, 0.0565232038, -0.031974785, -0.0338892043, -0.0523638315, 0.0310512632, 0.0118473042, 0.0038265572, 0.0163503941, -0.0094052861, 0.0066734958, -0.0306810122, -0.0662884042, -0.00877205, -0.0292062741, 0.0452927127, 0.0228564609, -0.0415101573, -0.0445910282, -0.0612073354, -0.0180176813, -0.0292809345, -0.0634722933, 0.0351663642, 0.0432645828, -0.0464963987, -0.0183700453, -0.0392345339, 0.0396708585, 0.010610857, -0.0250733122, -0.0814349353, -0.0149058094, 0.0100281872, -0.048905611, 0.076480113, -0.0507410616, 0.0559619851, -0.0319992453, -0.0024833991, -0.0352073722, -0.0391506329, -0.0201459341, -0.01703736, -0.0333858468, -0.0160693731, 0.0127236936, 0.0447326191, 0.0189862959, 0.0326716639, 0.0079240557, -0.0547168627, 0.0299563594, 0.0565562434, 0.0391528532, 0.0516525432, -0.0124642868, -0.0163197033, 0.018479405, 0.0285434574, -0.0219131578, -0.0098749846, 0.0025939322, 0.0182801802, 0.0078019518, -0.0360302366, -0.0023795997, 0.0250829216, 0.0464946665, 0.0047928356, 0.0019118827, -0.0154946931, 0.0058866693, 0.0123864384, -0.0080753034, -0.0516574793, -0.0159967672, 0.0162441395, 0.0084163984, 0.0008350718, 0.0030548384, -0.0163729917, -0.0130381593, -0.0747290403, 0.0102091692, 0.0192386601, 0.0105858259, 0.0210657902, 0.0344361514, -0.0043568756, 0.015986437, -0.0097992709, 0.0491996631, 0.0490689464, 0.0538723581, 0.021927122, 0.0031644655, -0.0190463085, 0.0181929842, -0.0407385044, -0.0199326072, 0.03524689, 0.046516262, 0.0380556993, 0.0074241045, -0.0127501898, 0.025424445, -0.0002897606, -0.0049451706, -0.009019712, 0.0410818011, -0.0314530134, 0.0875645578, 0.010797387, 0.0042200531, -0.0303947367, 0.0453876369, 0.0311872326, 0.0236255266, -0.0187570974, -0.0126681374, -0.0145742865, -0.0233286843, 0.0617616698, -0.0198280755, -0.0111561362, 0.0284538325, -0.0901448578, -0.0445015766, -0.0246782377, 0.038184084, 0.0238674991, 0.0079213744, 0.0264062136, 0.0512264371, -0.0038635, -0.0862882286, -0.0065130405, 0.0010735494, -0.0613773912, -0.0248392504, 0.0177937187, -5.997447483e-33, -0.0275221579, -0.011799641, 0.0029730098, -0.0325202793, -0.0166018736, 0.0129109677, 0.0147975348, 0.0020257165, -0.0339126512, 0.0018308042, 0.0107709905, 0.0160301439, 0.0075435783, 0.0313722789, 0.0013176708, -0.0585457161, 0.0407938324, 0.0092898309, -0.02491487, -0.02133983, -0.0484755337, 0.0031148961, 0.0207577739, -0.0289602857, 0.0517791919, 0.0104105547, 0.0377752632, 0.0322741158, -0.0277513769, 0.0628118962, -0.0137212751, 0.0825619623, 0.0381396599, 0.0198294837, -0.0180504024, 0.0430683531, -0.0538323298, -0.0235237814, 0.0176870301, 0.0342624038, -0.069821991, 0.0171067938, 0.0057349168, 0.0401112624, -0.027207464, -0.0787249058, 0.0512974374, 0.007337607, 0.0099254968, 0.0365398936, 0.0234478917, 0.0069273924, -0.027455885, -0.008088639, -0.0002829875, 0.0318506695, 0.007355704, 0.0512704588, -0.0108935637, 0.0214618333, -0.0025644954, -0.0307653267, 0.0884776562, 0.0599396527, 0.025525799, 0.0325573161, 0.0719190165, -0.0003332202, -0.0584961213, -0.0715053231, 0.0322571173, -0.0252913889, 0.0351758562, 0.0752118528, 0.0308525823, 0.0270854086, -0.053633634, -0.0050701811, -0.0111720348, 0.0342978649, 0.0024659375, -0.0063175657, -0.0428496413, -0.0228740666, -0.0573945493, -0.0148213999, -0.0561354384, 0.0101516973, 0.0749734566, -0.0010297126, -0.0163745061, 0.0155026233, 0.0180135705, -0.0445064269, -0.0182513557, 0.0043794974, 0.0172269512, 0.0016487424, 0.0279957224, 0.0261721741, -0.0397067629, 0.0537309162, 0.0372707769, 0.0057531041, -0.0163353793, 0.0219662972, -0.0088708876, 0.0046400828, -0.0561375432, -0.0046510929, -0.0334636793, 0.0205218792, -0.0199260805, 0.0216364712, 0.0172304697, 0.0119461576, -0.0227795802, 0.0280751269, -0.0119203338, -0.030914247, -0.0393312238, -0.0227015931, 0.0063206782, 0.0153922178, -0.0206044782, -0.0109576089, -0.0089352392, -0.0090934532, 0.0449424163, -0.0182736553, -0.005928067, -0.014904649, 2.191e-7, -0.0083246389, 0.0187353622, 0.0060611055, 0.0306684822, 0.0001606626, 0.0604452118, 0.0164200291, 0.0458465293, 0.0538647957, -0.0160352495, -0.0421071202, 0.0534007922, 0.0135736074, -0.0025893759, -0.025311511, 0.0195884015, -0.0404608995, 0.005723197, -0.0302915424, 0.0217387881, 0.0387741067, 0.0273087919, -0.0040114713, -0.0133625614, 0.0410912558, -0.0432456285, 0.0252466369, -0.0485802293, -0.0058966605, -0.0462242551, 0.018239107, -0.0065079215, 0.0420979522, 0.0689751506, 0.0170422308, -0.0073340093, 0.0130515778, 0.024138568, -0.0241064131, 0.0027370621, -0.0401053466, 0.0191488899, -0.0253930092, -0.0594578311, 0.0816347301, 0.0264873691, -0.048954837, -0.0487830453, -0.0094248513, -0.0260108039, 0.0719131976, -0.0134332655, -0.0326316953, -0.0038325547, 0.030163452, -0.0239079148, -0.0085971821, -0.0690132231, -0.0182992276, -0.0720488206, -0.013158503, -0.0617928989, -0.0407516398, 0.0382670686, 0.0014545855, 0.0170181841, -0.009514966, 1.756923588e-34, 0.0198619869, 0.0327801891, 0.0466382653, -0.0406185426, -0.0179415029, 0.0045206468, -0.0038182, -0.0338469334, 0.0436029248, -0.0207815617, -0.0012726922 ]
http://arxiv.org/abs/2406.17972v1
http://arxiv.org/pdf/2406.17972v1
LABOR-LLM: Language-Based Occupational Representations with Large Language Models
2024-06-25T23:07:18
2024-06-25T23:07:18
Tianyu Du, Ayush Kanodia, Herman Brunborg, Keyon Vafa, Susan Athey
cs.LG
cs.LG, cs.CL, econ.EM
Many empirical studies of labor market questions rely on estimating relatively simple predictive models using small, carefully constructed longitudinal survey datasets based on hand-engineered features. Large Language Models (LLMs), trained on massive datasets, encode vast quantities of world knowledge and can be used for the next job prediction problem. However, while an off-the-shelf LLM produces plausible career trajectories when prompted, the probability with which an LLM predicts a particular job transition conditional on career history will not, in general, align with the true conditional probability in a given population. Recently, Vafa et al. (2024) introduced a transformer-based "foundation model", CAREER, trained using a large, unrepresentative resume dataset, that predicts transitions between jobs; it further demonstrated how transfer learning techniques can be used to leverage the foundation model to build better predictive models of both transitions and wages that reflect conditional transition probabilities found in nationally representative survey datasets. This paper considers an alternative where the fine-tuning of the CAREER foundation model is replaced by fine-tuning LLMs. For the task of next job prediction, we demonstrate that models trained with our approach outperform several alternatives in terms of predictive performance on the survey data, including traditional econometric models, CAREER, and LLMs with in-context learning, even though the LLM can in principle predict job titles that are not allowed in the survey data. Further, we show that our fine-tuned LLM-based models' predictions are more representative of the career trajectories of various workforce subpopulations than off-the-shelf LLM models and CAREER. We conduct experiments and analyses that highlight the sources of the gains in the performance of our models for representative predictions.
LABOR-LLM: Language-Based Occupational Representations with Large Language Models: Many empirical studies of labor market questions rely on estimating relatively simple predictive models using small, carefully constructed longitudinal survey datasets based on hand-engineered features. Large Language Models (LLMs), trained on massive datasets, encode vast quantities of world knowledge and can be used for the next job prediction problem. However, while an off-the-shelf LLM produces plausible career trajectories when prompted, the probability with which an LLM predicts a particular job transition conditional on career history will not, in general, align with the true conditional probability in a given population. Recently, Vafa et al. (2024) introduced a transformer-based "foundation model", CAREER, trained using a large, unrepresentative resume dataset, that predicts transitions between jobs; it further demonstrated how transfer learning techniques can be used to leverage the foundation model to build better predictive models of both transitions and wages that reflect conditional transition probabilities found in nationally representative survey datasets. This paper considers an alternative where the fine-tuning of the CAREER foundation model is replaced by fine-tuning LLMs. For the task of next job prediction, we demonstrate that models trained with our approach outperform several alternatives in terms of predictive performance on the survey data, including traditional econometric models, CAREER, and LLMs with in-context learning, even though the LLM can in principle predict job titles that are not allowed in the survey data. Further, we show that our fine-tuned LLM-based models' predictions are more representative of the career trajectories of various workforce subpopulations than off-the-shelf LLM models and CAREER. We conduct experiments and analyses that highlight the sources of the gains in the performance of our models for representative predictions.
[ 0.0094538769, 0.1390409172, -0.0136197004, -0.0349908993, 0.0044883783, -0.0054745148, 0.0429925881, 0.0012979503, -0.0051673008, -0.0034865479, 0.0417472981, -0.0022957562, 0.0027442032, -0.0104521019, 0.0218045134, -0.0287623983, 0.0099941911, 0.003954771, -0.0071351486, -0.015128714, -0.0367523767, 0.0328131467, 0.0038292508, 0.0939177945, -0.0163545776, -0.0073035737, 0.0216235146, -0.0085918242, 0.0188260246, 0.0122960964, 0.0326935574, 0.0223574247, 0.0383805409, 0.0023307686, 0.0000020834, -0.0210255142, -0.0248371325, 0.0117194178, -0.0266586151, -0.0877987221, 0.0421281382, -0.0450614132, -0.0067245299, 0.0384790227, -0.0131728547, -0.0091163982, 0.0493434966, 0.0170088988, 0.0434555523, 0.0161792506, -0.0069601918, -0.0685232952, -0.0416914634, -0.0324486755, 0.0181834772, -0.0284421314, 0.0580888242, -0.0447908156, -0.0739038736, -0.0140454937, -0.0495257527, -0.0177736413, -0.0170154888, -0.0387796387, 0.0061987196, 0.0683068782, 0.0032804974, 0.0189771578, 0.014890546, 0.0160594918, -0.0365498066, -0.0468068235, -0.0172251053, -0.0158807542, 0.0455648154, 0.0486220904, -0.0452579558, 0.0216095597, -0.0062544392, 0.0257545784, 0.0288692303, 0.0401481166, -0.0025600474, -0.0100178942, -0.0287953634, 0.1111982018, -0.0156462118, -0.0079526957, 0.0507550351, 0.0148684382, 0.0868657455, -0.0221982189, 0.0254728124, -0.0002058688, -0.0330333412, 0.0203349534, -0.0257873368, 0.0123980306, -0.0225902442, -0.0072093811, -0.0124909859, 0.0443484224, -0.0333711207, 0.0450052693, 0.008856765, -0.0154110799, 0.0083086025, -0.0240766574, -0.0382071696, -0.0395738184, -0.0773435757, -0.0234568641, -0.0445905253, -0.0352668017, -0.0007336328, 0.0485838056, -0.0217458215, -0.0055784094, -0.0201610457, -0.0123785436, 0.0106499996, 0.0239081811, -0.0426733047, 0.0272244476, -0.0666593239, 0.0492069647, 0.0281631723, 0.0546375699, 0.0158067271, 0.0242548268, 0.0317257792, 0.0055880519, 0.0099903243, -0.0204895642, 0.0381968021, 0.045095697, 0.0295654051, 0.0503613837, -0.0320135839, -0.0379359834, -0.0068908976, -0.0158612747, 0.0205967743, -0.0649761111, 0.0481186025, -0.0036943255, 0.0072563281, -0.0651456937, 0.0110378833, 0.0195416473, -0.0480927117, 0.0381405279, 0.02897273, 0.0097226035, 0.0379364043, -0.025453588, -0.0108380122, 0.0218872, 0.0429018661, 0.0416530818, -0.0068001756, -0.0802479461, 0.0343967415, 0.0561646186, -0.0303480141, 0.0061858874, -0.0043046437, 0.0154241053, 0.0276946854, 0.0054776873, 0.0743194148, -0.0337345451, 0.0099012926, -0.0005081933, 0.0456320941, 0.0628941134, 0.0387630463, 0.0420766287, 0.05123486, 0.0322406478, -0.0203739218, -0.047400225, -0.0224696212, 0.0163244177, -0.0232936181, -0.0264386199, -0.0681686625, 0.005403433, -0.0192369074, 0.0002966415, 0.0505917333, 0.0117667597, -0.0074242451, 0.0090692956, 0.0441662185, 0.0151511161, -0.0330378674, 0.0962455496, 0.0411811583, 0.0073238024, -0.0056552961, -0.0273709707, -0.0118807321, 0.0760682896, 0.000047369, 0.0035059813, -0.006134178, -0.0554920845, 0.0514058843, 0.0255316216, 0.0294096023, -0.0002318823, 0.0156823024, -0.014008713, -0.0212193187, 0.0628249496, 0.0270372573, 0.015438511, -0.0047099688, -0.0036163242, 0.0075003644, -0.0677227154, -0.0335876606, 0.0316032469, -0.075424701, -0.0076460796, 0.0640221536, -0.0049884901, -0.0258353408, -0.0360371061, 0.0001914813, -0.0010821579, -0.0144608598, -0.0136203608, 0.0006422815, 0.0220240336, -0.0382582173, -0.0035168189, -0.0212383438, -0.014318129, 0.0044635786, 0.0076022861, -0.0607655384, 0.0050682374, 0.0053160368, 0.0206203777, 0.0214642193, 0.0195302684, 0.0220010318, -0.0130445939, 0.0146496901, -0.0124384826, 0.0252043996, 0.0356197953, -0.004096223, -0.029914014, -0.0263579842, 0.0026358499, -0.0987248793, 0.0185512882, -0.0299697667, -0.0089317719, 0.0194480233, -0.0098163811, 0.0123023838, -0.0140823461, -0.0263209306, -0.0649122223, 0.0159246661, 0.0270607676, 0.0036463146, 0.0056631439, -0.0365995467, 0.0046891719, 0.0084987581, -0.0436791778, 0.0160659812, 0.0254034176, -0.0216065757, 0.0120379049, -0.0154128596, -0.0172901861, -0.0392811783, -0.0540194735, -0.0395601839, 0.0424381644, 0.0193408616, 0.0213742126, -0.0436025299, -0.0435301624, -0.0148138301, 0.0052657104, -0.0199167896, 0.0178016331, -0.0397085361, -0.0146077629, -0.0089186355, 0.0327579379, -0.0055671297, 0.0379072726, 0.0410339125, -0.0332699195, -0.0280197784, 0.0070022689, -0.0248451196, 0.0133712227, -0.0450860821, -0.0580124632, -0.0231059026, 0.0234652217, 0.0345731266, -0.0260246173, -0.0204963405, 0.0193156954, -0.0262753777, 0.0079004001, -0.0633230582, 0.0589453578, -0.0109235542, 0.0443800874, 0.0082975244, -0.0124066491, -0.0136944205, -0.0071277232, 0.0469809175, -0.0022909413, -0.0742177814, -0.0462591462, 0.0309106819, -0.0053366874, -0.0199857727, 0.0418591946, -0.0123106195, -0.0723631382, 0.0505353808, -0.021944046, -0.0288195144, -0.0222997312, -0.0488156714, -0.0018388496, 0.0236276388, -0.0360725895, -0.0383301191, -0.0948436558, -0.0312077925, 0.0489202328, 0.0221224781, -0.0009609342, -0.0760659799, -0.0057684765, -0.0152384033, 0.0266023222, 0.0188751332, -0.0296138935, -0.017187966, 0.009407104, 0.054836262, 0.0532420576, -0.0283976328, -0.0572638735, -0.0036864132, -0.0224694051, -0.0148859499, 0.0438257568, -0.0197376404, 0.0049653761, 0.0391563438, -0.041195143, 0.0090147257, 0.0359849222, -0.0179083291, -0.0522850379, -0.0050411327, 0.0920412019, -0.0105397077, -0.0020446738, 0.0021779269, 0.0215057619, -0.0148034059, 0.0293506552, 0.0698485672, -0.0469669737, 0.0000083054, 0.0113314828, 0.0309856627, 0.0088366056, -0.0300756246, -0.0099343332, -0.106550552, 0.0714468136, -0.0113805272, -0.0325246193, -0.0000573518, -0.0419828296, -0.0040605902, -0.0458155051, -0.0249288641, -0.0360396281, -0.0274458192, 0.0111220535, 0.0471832491, -0.044584848, -0.0273662005, 0.06471356, -0.0313145854, 0.0134951798, 0.0412982255, 0.0489191748, -0.0407036766, 0.0134873316, -0.0093904408, 0.0055837175, -0.0010918631, 0.0435705408, 0.0383400805, -0.0173672363, -0.0219636969, -0.0076398975, 0.0045679547, -0.0262013543, 0.0009949855, 0.0449661091, 0.0735815465, 0.0009356037, -0.0562793948, 0.0092920382, 0.0282450262, -0.0070550037, 0.072099641, -0.0374268629, -0.0759566426, -0.0025238337, -0.0408326536, -0.0125560351, 0.0297733676, -0.0607644394, 0.0395565443, -0.0397550948, 0.001535793, 0.0159178637, -0.0055430192, 0.0643368587, 0.0137991821, 0.0046936534, 0.0129585126, 0.0468571745, -0.0355095603, -0.0568783842, 0.003798445, -0.0844152346, 0.0422578938, -0.033441823, -0.0116284415, 0.0162543003, -0.0316445678, -0.0603774525, -0.0508038886, -0.1136767045, -0.0128878765, -0.0080993213, 0.015529111, 0.0173358005, 0.0199012365, 0.0005779678, -0.0259691775, 0.0692905784, 0.0891548172, -0.0111769205, 0.0007796112, -0.0106942831, 0.0012659632, -0.0348526724, -0.0130105009, -0.0521713793, 0.0033264679, 0.0214516427, -0.0205057804, -0.0157435499, -0.0544980802, -0.0227036569, -0.0293571055, 0.02750469, 0.003032306, -0.0379103683, -0.0485935882, -0.0147171197, 0.0261214487, -0.0117141521, -0.029344229, -0.0297229402, 0.1076539755, 0.0351923071, 0.0421556346, 0.0046228045, 0.0423381664, 0.0258092843, -0.0702747181, 0.0279108565, -0.0356577486, 0.027713079, 0.0044016684, -0.0091986572, 0.0125576314, -0.0372715257, 0.0447367169, 0.0405478962, 0.0367156267, 0.0174370483, 0.0059257802, 0.0336669236, 0.0075864838, -0.0179285947, -0.0400260426, -0.0819973424, -0.0339772776, 0.0191553663, 0.0301696025, -0.0878981724, 0.0174566843, -0.0177352298, 0.0061886348, -0.0204640478, -0.007510778, 0.0304332823, 0.002861968, 0.0751599595, -0.0254169572, -0.0188574046, -0.0091600334, 0.0360332131, 0.1026585475, 0.0319076963, -0.0326310247, -0.0509625897, -0.0288856532, 0.0040222751, 0.0336210392, 0.0195435379, 0.0030361041, 0.0135483826, -0.0617243759, -0.0429591052, 0.0190165192, -0.0560309701, -0.0162854549, -0.0287017394, -0.0119988844, 0.0278277732, -0.0267849527, -0.0000179933, -0.0429321416, -0.020247696, -0.0063696201, -0.0110049061, -0.0708419606, -5.836679529e-33, -0.0238498058, -0.0262172632, 0.0396942198, -0.0009379613, -0.003755148, -0.0125742471, 0.0057634213, -0.0292148627, -0.0069353646, -0.0217291266, -0.0332545042, 0.003102025, 0.0236394424, 0.0496464819, 0.0407431796, -0.0023930157, 0.0452091657, 0.0374659374, 0.0135661196, -0.0103311297, -0.0086443191, 0.0358803011, 0.0851894915, -0.0740861967, 0.0163906924, 0.0160036664, 0.0030578407, 0.0179849081, 0.0032541582, 0.0377776884, -0.0022824358, 0.0350870639, 0.0142999068, -0.0857232958, 0.0013589443, 0.0108443834, -0.0636457652, -0.0119277202, -0.0207425971, 0.0233579781, 0.0583566092, -0.0163089614, -0.0086780172, 0.0258030854, -0.0040414617, -0.0614033043, 0.0730381906, -0.0318360925, 0.0418356508, 0.0043733288, -0.0377750657, 0.0147429816, -0.0118618626, 0.0196831692, -0.0015713637, 0.0523588173, 0.0090201302, 0.0312420614, -0.1116719246, -0.016785156, -0.0020830478, 0.0229155142, 0.0151812462, 0.0269231368, 0.0168329403, 0.0262647662, 0.0499921255, 0.0194059815, 0.0229812507, 0.0066186204, 0.0222089048, 0.0487065651, -0.0064284452, 0.0252293367, 0.0576525852, -0.0434776507, -0.0270509776, 0.0037453345, 0.01442823, -0.0122663407, 0.021279607, 0.0209678542, 0.0079370281, -0.0643904284, -0.0562489107, -0.0406707898, 0.0231702179, 0.0153624536, 0.038808234, -0.0044610845, 0.0324528702, 0.0148699582, 0.0165753886, -0.0593001842, -0.0466129445, 0.0059774397, -0.0603189878, 0.0001007649, 0.0199594721, 0.0043934593, -0.0810042173, -0.0208189283, 0.0006750728, -0.009159985, 0.0482853875, 0.0658312961, 0.0208171494, 0.0333120935, -0.0657583326, -0.0175315142, -0.0152075738, 0.0074769179, -0.0118472334, 0.0231721625, 0.0472789928, -0.0001042525, 0.0377488397, 0.003119421, 0.0128571838, 0.0136507358, 0.0295073278, -0.0218356866, 0.0153710749, -0.0087984176, -0.0134716826, -0.0337831639, -0.0001775394, -0.0238851644, 0.0330496542, -0.0436848477, -0.0268541519, -0.0072739455, 2.618e-7, 0.0494438522, 0.0212674718, -0.0591801591, 0.0191731919, 0.0172032639, 0.0252585225, -0.0124806846, 0.0325437523, 0.0648375824, 0.025924759, 0.0078210123, -0.0062680109, 0.048931282, -0.014126271, -0.1019211039, -0.033679191, -0.0265375264, -0.0192133784, -0.0580946654, 0.0054236939, 0.0847891644, 0.0897402391, 0.0526648797, -0.0347015448, 0.019928949, 0.0074261548, 0.0582564846, -0.0456268787, 0.0093121072, -0.0013737291, -0.0200720169, 0.0296256095, -0.0234803874, 0.039132759, -0.0335603394, 0.0493624173, 0.0193995908, 0.0576032139, 0.0048751258, 0.0543811396, -0.0036424964, 0.0231086146, -0.0137732392, -0.0346378386, 0.0597311333, -0.0619121678, -0.0265761595, -0.0483670533, -0.0805788785, 0.0234363154, -0.0045587583, 0.0132817244, -0.0268136542, -0.0598899983, 0.0333928578, 0.0229738783, -0.0452765673, 0.0128343171, 0.0043105939, 0.0555737354, -0.0602614023, -0.0422383361, 0.0336084515, 0.0302986689, 0.02244821, -0.0440737382, 0.0253592357, 2.529474432e-34, 0.0342487358, 0.063925378, -0.0027624744, -0.0363049507, 0.0246891864, 0.0193767976, 0.0438794009, 0.0374714136, 0.0081867808, -0.0484125987, -0.0071460162 ]
http://arxiv.org/abs/2406.17969v1
http://arxiv.org/pdf/2406.17969v1
Encourage or Inhibit Monosemanticity? Revisit Monosemanticity from a Feature Decorrelation Perspective
2024-06-25T22:51:08
2024-06-25T22:51:08
Hanqi Yan, Yanzheng Xiang, Guangyi Chen, Yifei Wang, Lin Gui, Yulan He
cs.CL
cs.CL, cs.AI
To better interpret the intrinsic mechanism of large language models (LLMs), recent studies focus on monosemanticity on its basic units. A monosemantic neuron is dedicated to a single and specific concept, which forms a one-to-one correlation between neurons and concepts. Despite extensive research in monosemanticity probing, it remains unclear whether monosemanticity is beneficial or harmful to model capacity. To explore this question, we revisit monosemanticity from the feature decorrelation perspective and advocate for its encouragement. We experimentally observe that the current conclusion by wang2024learning, which suggests that decreasing monosemanticity enhances model performance, does not hold when the model changes. Instead, we demonstrate that monosemanticity consistently exhibits a positive correlation with model capacity, in the preference alignment process. Consequently, we apply feature correlation as a proxy for monosemanticity and incorporate a feature decorrelation regularizer into the dynamic preference optimization process. The experiments show that our method not only enhances representation diversity and activation sparsity but also improves preference alignment performance.
Encourage or Inhibit Monosemanticity? Revisit Monosemanticity from a Feature Decorrelation Perspective: To better interpret the intrinsic mechanism of large language models (LLMs), recent studies focus on monosemanticity on its basic units. A monosemantic neuron is dedicated to a single and specific concept, which forms a one-to-one correlation between neurons and concepts. Despite extensive research in monosemanticity probing, it remains unclear whether monosemanticity is beneficial or harmful to model capacity. To explore this question, we revisit monosemanticity from the feature decorrelation perspective and advocate for its encouragement. We experimentally observe that the current conclusion by wang2024learning, which suggests that decreasing monosemanticity enhances model performance, does not hold when the model changes. Instead, we demonstrate that monosemanticity consistently exhibits a positive correlation with model capacity, in the preference alignment process. Consequently, we apply feature correlation as a proxy for monosemanticity and incorporate a feature decorrelation regularizer into the dynamic preference optimization process. The experiments show that our method not only enhances representation diversity and activation sparsity but also improves preference alignment performance.
[ 0.0240752716, 0.0668296814, -0.0110686002, 0.0280555356, -0.0579100065, 0.010792587, 0.0294835344, -0.0473267883, -0.0602705441, 0.0226441082, -0.0564397611, -0.075721927, 0.0391083397, 0.0394579247, 0.0359034128, -0.0183704495, 0.0804702267, -0.0185279846, 0.0022536409, -0.0436441414, 0.0125454543, -0.0185257513, 0.0000714795, 0.0244967639, -0.0066777058, -0.0471853837, -0.0110695865, -0.0001848942, -0.0152070466, -0.0098347999, -0.0228355788, 0.0647953898, 0.0245584883, 0.0179954637, 0.0000020157, -0.0540635139, 0.0253108758, -0.0030863739, -0.0266758073, -0.004683665, 0.001700249, -0.0120914439, 0.0267586671, -0.0336652361, -0.0060353796, 0.0359419137, 0.0534155853, 0.0434771739, -0.0106481006, 0.0470581576, -0.0115181599, -0.0328652598, 0.0147455083, -0.038200777, -0.0148367127, 0.010046388, 0.0258911271, -0.0586708821, -0.0055644778, 0.0323183723, -0.0489894226, 0.0126232719, -0.0007692054, 0.0232048463, 0.0370287187, 0.0179895554, 0.0244531073, -0.0187027827, 0.0231199823, 0.0488831922, -0.0113343447, 0.0189390033, -0.0216809046, -0.0056046173, 0.0040405523, -0.0076724375, -0.0271746945, -0.0286606308, 0.0255392361, 0.026902562, 0.0081285015, -0.0521677211, 0.0446155891, 0.0110371057, -0.0368795283, 0.0618891306, 0.0190892834, -0.0447848402, 0.0095882034, -0.0078313546, -0.0482018739, -0.0062493961, 0.013485495, -0.0236773528, -0.004436776, 0.0052336804, -0.0108123617, -0.0158013459, 0.0189408567, -0.0784822181, 0.0271911435, 0.0602925085, -0.0101133063, 0.0286597274, -0.0215112586, 0.0830012336, -0.0358084701, 0.0286037102, -0.0510904565, 0.0269197468, -0.0098543083, -0.0442332812, 0.0225806478, 0.0637777224, 0.0486066192, -0.0261608195, -0.0401851088, 0.0309607461, -0.0014304341, -0.0132346964, -0.0496089719, 0.0237023886, -0.0022850723, -0.011366698, -0.047622934, 0.0805389583, -0.0335754976, 0.0265834499, 0.0035854625, -0.0095140319, -0.0164136384, 0.0205143094, -0.0197728723, -0.0847158432, 0.0127109038, 0.0482996814, -0.0203224588, -0.0230529867, 0.0124866655, 0.0226619318, -0.0722936541, -0.0245558098, 0.017895082, -0.0597420298, 0.0226893164, 0.0332233571, -0.021604348, -0.0652884245, -0.0162260905, -0.0493266396, -0.0534192398, 0.0347461887, -0.0067300512, 0.0194656029, 0.0096230488, -0.0235617328, 0.0217300281, 0.071840249, 0.0507024974, 0.0123307528, 0.0098791448, 0.0070361439, -0.0683472231, 0.0102037601, -0.0189661663, -0.0070018489, -0.0817014277, -0.027209498, 0.0265399497, 0.0359249413, 0.0794968903, 0.0363378897, -0.0367997363, -0.0724288151, 0.0812739432, 0.0443582423, 0.0215720627, 0.0040695146, 0.0015515197, -0.0235714298, -0.0078267409, 0.057818085, -0.033400137, 0.0599314459, 0.0029020708, 0.030820474, 0.0410647392, 0.0544881113, -0.0352684446, -0.0252624955, -0.014045896, -0.0069900956, 0.0125061693, -0.0199990142, 0.0402820744, -0.0426428393, -0.0124713629, 0.0188792441, -0.0012867785, 0.0098015321, 0.005184717, -0.0455912575, 0.0157457553, 0.0740521252, 0.0090599908, 0.0175526403, -0.0500434674, 0.014102947, -0.0468692295, 0.0682029799, 0.0272050425, 0.0161651503, 0.062179368, -0.0437062234, -0.0320655294, -0.034774188, 0.0019960282, 0.0178981367, -0.0198915955, -0.0172118787, -0.0075578927, 0.0236110799, 0.0220327545, 0.0116369464, 0.0136600658, -0.0022692205, 0.0233106129, 0.0044659362, 0.0135598406, -0.012976842, -0.0206580963, 0.0358422957, -0.0052269315, -0.0277273525, 0.0045762714, -0.0128453262, -0.110463284, -0.0303102657, 0.0472335666, 0.0378697664, -0.0071817227, 0.0329557545, 0.0112018976, -0.036681883, 0.0145470062, 0.0235393234, 0.0027446065, -0.0243832022, 0.0211374424, 0.0016452479, 0.0506525487, -0.0608568713, 0.0409589261, -0.0146766491, -0.0009123532, -0.0160257015, -0.0182422083, 0.0363855325, -0.0293942839, -0.0590163842, 0.0050847051, -0.0092239948, 0.0158907417, 0.0125808688, -0.0325158574, -0.0069239335, -0.0460761823, -0.06406831, -0.0042721103, -0.0020979838, -0.0044392059, -0.0071765915, 0.0131148621, 0.0212790854, 0.0221488923, -0.0229805484, -0.0017726055, -0.0348434374, 0.0735488683, 0.0330656134, 0.0078758104, -0.0249421243, -0.05932476, -0.0356121734, -0.0113881901, 0.0402222797, -0.0557055511, -0.008569777, -0.0065539638, 0.0346039683, -0.0373200327, -0.0142123243, 0.0041032587, -0.0226481613, 0.0317405313, -0.0074639167, 0.0667369962, 0.0509048663, 0.0054869838, 0.0075967982, 0.0700976774, 0.0333922915, -0.0257129278, 0.0022758655, -0.0358862244, 0.0334206335, -0.0513081886, 0.0076297573, -0.0447194017, 0.1044813395, 0.0219850186, -0.0507862531, -0.0474060141, -0.0154325552, 0.0159736332, -0.0406265333, -0.0283050276, 0.00106732, -0.0416552089, 0.0154481549, -0.0423903875, -0.0491806976, 0.0575228781, 0.0112017822, 0.0161968414, -0.0048337588, 0.0069558495, -0.0265117157, 0.0509759039, 0.0399306081, 0.0238192417, -0.0080865612, -0.016387742, -0.1028238386, -0.0109852171, -0.0015628664, 0.0074703638, 0.0972296372, -0.0125597231, 0.0097645586, -0.0015825762, -0.0500361286, -0.0332733728, 0.0047223852, 0.0110584013, 0.0295880158, 0.0587267205, 0.0106259137, -0.041500628, -0.0183951408, 0.0334028788, 0.0335730761, -0.0238751229, 0.0045621768, 0.0193298142, 0.0288328305, 0.0531767607, 0.0694134161, 0.0086442111, -0.038860444, 0.0157183278, 0.0216243993, -0.0173463207, 0.0170233846, -0.0290924981, 0.0219276641, 0.018630892, -0.0206879284, -0.0439607091, -0.0103989374, 0.0400728732, -0.0242997929, 0.0098169735, 0.0859355703, 0.0092757149, -0.0459510162, 0.0193193574, -0.0476552807, 0.0816745684, 0.0373687334, 0.0051944647, 0.0005761692, -0.0011183042, -0.0226161461, -0.0387764126, 0.0323111713, 0.0177675523, -0.0112157632, -0.0527139977, -0.0371873751, 0.0263201743, -0.0421933979, 0.0072188401, -0.0071633891, -0.0097681209, -0.0358625352, 0.0527010001, 0.0406778492, 0.0002504202, -0.0143325897, 0.0237948336, -0.0389442742, -0.0280894022, 0.0048176721, -0.0342144556, 0.030737197, 0.0840650648, -0.0116949435, -0.0214903969, -0.0038590673, 0.0009565013, 0.0475883111, 0.0340506546, 0.0107755614, 0.0241533965, 0.0162208062, -0.0153987566, -0.0285147633, -0.0071306447, 0.0073912214, -0.0289125089, -0.056078352, 0.0291097537, 0.0013810487, 0.0010883247, -0.0302184727, -0.0158666391, -0.0510941781, 0.0540055148, 0.0271076392, -0.0980009809, -0.0046665492, -0.0117063355, -0.0390963852, -0.0418173186, -0.0320928581, 0.0156646688, 0.0632225052, -0.0189149734, -0.0524990708, 0.0078224847, -0.0103285192, 0.0218587611, -0.00850417, 0.022172654, -0.0475175381, -0.0348407663, -0.0299464092, 0.0100048799, 0.0395340025, 0.0037837615, -0.0619240142, 0.0086817769, 0.0189022105, -0.0204977654, -0.010589581, -0.036072202, -0.0887949988, -0.0795290023, 0.023537932, 0.0178909823, 0.0112643875, 0.0275369752, 0.0222373009, -0.0298402384, 0.0458499379, 0.0654348731, 0.00806743, 0.0481376611, 0.0126022575, -0.0048607248, -0.0246834978, 0.0312642157, -0.0132455472, -0.0159129482, 0.0200727116, 0.0570729896, 0.0017159608, -0.0145331239, 0.0135316337, -0.0091205155, 0.0327702463, 0.0033717719, 0.0832648277, -0.0270929746, 0.0171089899, 0.0153722772, 0.0165766198, -0.0109519958, -0.0623236112, 0.0728413388, 0.0054441993, -0.0209175088, -0.0212551393, 0.0424576923, 0.0188604742, 0.0362121239, -0.0455050655, -0.0601535663, -0.0200782716, 0.0206060335, 0.0434412695, -0.0137827052, 0.0917030573, -0.0380797833, -0.015730761, -0.0105963405, 0.020514572, 0.0132899005, 0.0218924787, -0.0343483649, -0.0040748157, -0.0452029184, -0.0405732542, -0.0151697556, -0.046666, 0.0378162265, -0.0081321625, 0.0735770687, -0.0134301959, -0.001145061, -0.043691352, 0.0039969976, -0.0315695964, 0.0085035972, 0.0427281372, 0.0471951067, -0.0140322661, -0.0181700327, 0.0761310682, -0.0070749042, -0.0006429576, -0.0035804517, 0.0293769538, 0.0014467494, -0.0177466031, 0.0277113672, 0.0340410434, -0.0612076931, -0.0018484602, -0.0539800525, -0.0475878753, 0.0460714325, 0.00738513, 0.0359915867, 0.0006601403, 0.0289699882, 0.0291962568, -0.029226942, -0.0229102336, -0.0076494222, -0.0410259068, 0.0308104157, -0.0104956925, 0.01740635, -6.911859028e-33, -0.0423492752, -0.0010211479, 0.003903975, -0.0387796015, -0.0163723323, -0.0811300501, 0.0039477423, 0.0571183525, -0.0448330194, 0.0086024608, -0.0131080924, -0.0009602954, -0.006422814, 0.0195147339, 0.0712751076, 0.0069547147, 0.0326965488, 0.0024130514, 0.0042694025, -0.0408015922, 0.0657344162, 0.0209148321, 0.0603343248, -0.0710287914, -0.015364171, 0.0348774418, -0.0036192141, -0.0098906616, -0.0223314818, 0.0619132742, -0.0281255711, 0.0166129041, 0.0344287492, -0.0162447803, -0.0251807272, -0.0723637417, -0.0296289008, -0.0621047802, 0.0448601767, 0.0046521528, -0.0216765106, -0.0489228182, 0.0528086238, 0.0016173638, -0.0023638743, -0.0464100167, 0.0349825546, -0.0324412212, -0.0679957122, 0.0307497177, -0.0818978995, 0.0123764519, -0.0039208769, 0.0292890109, -0.005570997, -0.0143088922, -0.0194228273, 0.0189186409, -0.0277244113, 0.0243048891, 0.0052384506, 0.0809855461, 0.0320960358, 0.0493830293, 0.0302321333, 0.0529530682, 0.0566555969, 0.0015558578, 0.0223705713, -0.0279525053, 0.055870045, 0.0256686918, 0.0341983773, 0.0273079351, 0.0824161619, -0.013470781, -0.0496929511, -0.0067044864, 0.0342326798, 0.0123947114, 0.0332750157, 0.0342638232, -0.0172118526, -0.0356387869, -0.0141360639, -0.005046471, -0.0698371828, -0.1035255194, 0.0131551353, -0.0166060869, 0.0226772875, 0.0214910619, 0.0499629006, -0.0147152767, -0.0386960544, 0.0549682304, -0.0669217706, 0.0174519531, 0.0197824202, -0.0173687097, -0.0670324787, -0.0263922215, -0.0224073771, 0.0258977395, -0.0036081693, 0.0435452238, 0.0391735397, 0.0265182741, -0.0545251481, -0.0330155268, -0.0098677976, -0.0875299051, 0.0271099415, -0.0036871862, 0.0408799276, -0.0122170355, -0.0172938332, 0.0623427704, -0.0012955344, -0.0254427828, -0.0784521401, -0.0342673473, 0.0074299686, 0.0406913497, 0.017612271, -0.014103814, -0.0281261243, -0.0057188408, 0.0265235621, -0.051463794, -0.0374478363, 0.0216186456, 2.846e-7, 0.0216865018, 0.0253923032, -0.0085379062, -0.0026201003, 0.0289194565, 0.0347899497, -0.0543476567, 0.0185964126, -0.0343338959, -0.0240740683, 0.0242901929, 0.0564875714, 0.019741172, 0.016760001, -0.0587106012, 0.0562961809, -0.0208627637, 0.0605928339, -0.0224328432, 0.0144022191, 0.0172935445, 0.0998390764, -0.0085925953, -0.0026520186, 0.0121805677, -0.0070761768, 0.0771676376, -0.0208843648, 0.0037229229, 0.0338298045, 0.0055614701, -0.0602911226, -0.033711195, -0.0137378722, 0.012071982, 0.0523670502, 0.0117329955, 0.0696674138, -0.026505474, -0.0389843807, -0.0009672754, 0.0238906108, -0.0263297111, -0.0145209227, 0.0445401259, -0.0368694775, -0.0583977848, -0.0253650118, 0.0364892744, 0.0254308116, 0.0315012708, -0.0198143609, -0.0081253713, -0.0134941675, 0.0009347434, -0.0095706154, -0.0272860881, -0.0652571768, -0.0263468977, 0.0550753176, -0.0612803586, -0.0214492846, -0.0580795258, 0.0309923869, 0.0793670788, 0.0116828736, -0.0285356473, 2.511636535e-34, 0.0349927396, -0.0380920582, 0.0217414908, 0.0384594873, -0.0257633757, -0.0481201746, -0.0453334115, 0.0075128577, 0.0186017696, -0.0758946911, -0.0380253233 ]
http://arxiv.org/abs/2406.17967v1
http://arxiv.org/pdf/2406.17967v1
Unmasking the Imposters: In-Domain Detection of Human vs. Machine-Generated Tweets
2024-06-25T22:49:17
2024-06-25T22:49:17
Bryan E. Tuck, Rakesh M. Verma
cs.CL
cs.CL
The rapid development of large language models (LLMs) has significantly improved the generation of fluent and convincing text, raising concerns about their misuse on social media platforms. We present a methodology using Twitter datasets to examine the generative capabilities of four LLMs: Llama 3, Mistral, Qwen2, and GPT4o. We evaluate 7B and 8B parameter base-instruction models of the three open-source LLMs and validate the impact of further fine-tuning and "uncensored" versions. Our findings show that "uncensored" models with additional in-domain fine-tuning dramatically reduce the effectiveness of automated detection methods. This study addresses a gap by exploring smaller open-source models and the effects of "uncensoring," providing insights into how fine-tuning and content moderation influence machine-generated text detection.
Unmasking the Imposters: In-Domain Detection of Human vs. Machine-Generated Tweets: The rapid development of large language models (LLMs) has significantly improved the generation of fluent and convincing text, raising concerns about their misuse on social media platforms. We present a methodology using Twitter datasets to examine the generative capabilities of four LLMs: Llama 3, Mistral, Qwen2, and GPT4o. We evaluate 7B and 8B parameter base-instruction models of the three open-source LLMs and validate the impact of further fine-tuning and "uncensored" versions. Our findings show that "uncensored" models with additional in-domain fine-tuning dramatically reduce the effectiveness of automated detection methods. This study addresses a gap by exploring smaller open-source models and the effects of "uncensoring," providing insights into how fine-tuning and content moderation influence machine-generated text detection.
[ 0.0733409673, 0.044694066, -0.0299805757, 0.0383860283, -0.0499101579, 0.0188977122, 0.0272992775, 0.0236452688, 0.0322231688, -0.0312825181, 0.0224562027, -0.0339960717, -0.0168621726, 0.0374411494, 0.0458197966, 0.0020132775, 0.0252923667, -0.0138163762, 0.0061935754, -0.0052188975, 0.0342380926, 0.0443055518, 0.0223544799, 0.0334128663, -0.0006593284, 0.0020113238, 0.0260612834, 0.0267047696, 0.0047542048, 0.0075878426, 0.035046488, 0.0843489096, 0.0119352629, 0.1047967449, 0.0000021557, -0.0337373391, -0.02317762, -0.0117164645, -0.0153527623, 0.0535023026, 0.1045107096, -0.0097002089, -0.030777622, 0.0432299897, -0.0536525957, 0.0445202552, 0.0361253768, 0.0248808544, 0.0514179058, 0.0732509196, -0.0226411428, 0.0126544135, 0.0336790793, -0.0138041126, -0.0112653468, -0.003051572, 0.0318106711, -0.021413805, 0.0003753744, 0.0230845455, 0.0399602242, 0.0370893814, 0.0032129306, -0.0326300152, -0.0140358889, -0.0039856187, -0.0325612463, -0.0490314439, 0.003346619, 0.0363382809, -0.0246288851, -0.0636407882, 0.0307530779, -0.0164355449, 0.019449994, 0.0447158292, 0.002537427, 0.0369832478, -0.019025933, -0.0051505254, -0.0116063487, -0.0069426089, 0.0314926468, 0.0436873548, 0.0135339219, 0.010036408, 0.0122750904, -0.0499298833, -0.0217853095, 0.0029076058, -0.0380474553, -0.0622864477, 0.0237935856, -0.0054953843, -0.0073542758, -0.0071966848, -0.0732964873, -0.0438769348, 0.038533289, -0.0354369693, 0.0309407581, 0.018267991, -0.000181884, -0.0030851557, -0.0318353027, 0.0567046888, -0.044375658, 0.0013475827, 0.0141379656, -0.0039204224, -0.0193400178, -0.018167451, -0.015019835, -0.0217345674, -0.0188500546, -0.0050337589, -0.0578085519, 0.0310182832, 0.002297, -0.0287613086, -0.0226705894, -0.0356186591, -0.0089103309, 0.0029162606, -0.0043995264, -0.0056339544, 0.0240431856, 0.0178430285, -0.0035986495, -0.1150321215, 0.0030178891, 0.0498650745, 0.0312833562, -0.0246160552, 0.0014188737, 0.0226018708, -0.0066948542, -0.0470419861, -0.0226838887, -0.0480623282, -0.0114943944, -0.0579379313, -0.0126848835, -0.0820408762, -0.0044284649, 0.0194672737, -0.0080758166, -0.034831807, 0.0130422907, 0.0648638457, -0.0287584383, -0.0099380463, -0.0602804758, 0.0482877754, 0.0221512578, -0.0004210211, 0.0034497997, 0.0647285879, 0.0186876189, 0.0789111033, 0.0398329198, -0.0266482588, 0.0553108789, 0.0643987432, 0.0183732696, 0.0146302739, -0.0066599813, -0.0429479629, -0.0280695856, -0.0052039316, 0.0119105736, 0.0686402619, -0.0096551478, 0.0114504127, 0.1055358797, 0.049461931, -0.0168225449, 0.0037519359, 0.0850764066, 0.0640632138, 0.0377922095, 0.0016588119, -0.0344024934, -0.0194505397, -0.0403923541, -0.0201591793, 0.0138379131, 0.0558637455, 0.0007006678, -0.0089906463, -0.0125872195, -0.0100677991, 0.0115205068, -0.023818206, 0.0043839389, -0.0053081266, -0.0534534454, 0.0052217888, 0.0042795697, -0.0616080984, -0.0126846787, -0.0566241331, 0.0143288029, 0.07990852, 0.0347877033, 0.0587583035, 0.0541289747, -0.048999913, 0.0510556959, 0.0861694589, -0.0000889565, 0.0101271858, -0.0140881352, -0.0286186785, -0.0082328757, -0.0241707526, 0.0094030667, -0.0044961656, 0.0475310534, -0.0069137299, 0.0309658032, -0.0120849591, -0.0564924069, 0.0085545164, -0.0493070073, -0.0083038695, -0.0262558237, 0.0043532392, -0.0114173936, 0.030970566, -0.0093090758, -0.0158412289, 0.0174117703, -0.0078882603, -0.0188554171, -0.0003946603, -0.0202366579, 0.0432423949, 0.0266053528, -0.0143103963, -0.0366821699, 0.0311508533, -0.0106174611, -0.0027758221, 0.0474346988, -0.0530434698, 0.0223560128, -0.0708502531, 0.0172062553, -0.0716007352, 0.0391764194, 0.016136229, 0.0594952479, -0.0345770456, -0.0054129143, 0.0592051223, -0.0303579792, 0.0153884785, 0.0010382466, -0.0102037713, -0.0256685372, -0.0331720114, 0.0389674306, 0.0224920399, 0.0075411396, -0.0245378129, -0.038242884, -0.0496913642, -0.0020478, -0.0520245247, -0.000830573, 0.0499191508, 0.0108215827, -0.0402073711, 0.0442168936, -0.0102610337, -0.0401581526, 0.0546369664, 0.0238512494, 0.0167067759, -0.0507193282, -0.0231159292, -0.0235836152, 0.0078754947, -0.0262681339, 0.0169481672, 0.02042941, -0.0507918112, -0.0140914004, 0.0404013395, -0.0046911221, 0.0096593527, -0.0027055647, -0.0605211034, 0.006995399, -0.0176021978, 0.0399319902, 0.0322980881, 0.0364598222, 0.0266834255, 0.0082266638, 0.0207196046, -0.0181027688, -0.0291717686, -0.0178375468, -0.0517956726, -0.0314335823, -0.0750520974, 0.0212282855, 0.0583780967, 0.0141079957, 0.0011075268, -0.0463500991, 0.0488058813, -0.0254549775, -0.0376732722, 0.0133730434, -0.0221681166, -0.0325444676, 0.02797186, -0.0364421271, -0.0825760812, 0.0341805369, -0.0387991555, 0.0224933829, 0.000009713, -0.0612348691, 0.0088827601, -0.0317854173, 0.0391512476, 0.0369497016, 0.0323611982, -0.0036518325, 0.0300112367, 0.0250574816, 0.0304193888, -0.038552396, -0.0140391197, 0.0060818223, 0.0162967518, 0.0197030287, 0.0045691309, -0.0350789987, -0.062396653, -0.0170973353, -0.0092078447, 0.0541686118, -0.0486758463, 0.0073290402, -0.0121545214, 0.0156298224, 0.0088996235, -0.0390347503, -0.0010544675, -0.0384509638, 0.0243019499, 0.0215599593, -0.0323209018, 0.0278219953, -0.0521202348, 0.0118973926, 0.0399996825, 0.0009743196, -0.0276158396, -0.011784574, -0.0249216482, 0.0138075538, 0.0093058133, 0.0005262239, 0.0341569856, 0.0777354166, -0.0209297519, 0.0174489692, 0.1118760407, 0.0212489292, 0.0037352743, -0.0077942731, -0.0032904965, 0.0337964371, 0.0109387469, -0.0072940518, -0.0935214683, -0.008739586, 0.0016124189, 0.0470617898, -0.0688342527, -0.0119668292, -0.0246083103, -0.1108273566, 0.0546030626, 0.0058028889, -0.0131183518, -0.012107214, -0.0265465733, -0.0178779438, 0.0049344678, 0.0125811435, -0.0138033871, -0.0387859866, 0.0160465166, 0.0133838579, -0.0399492756, -0.0921265632, -0.0039324239, -0.0424578972, 0.0222695414, 0.0440990403, 0.0227455255, -0.0055377143, -0.0542965904, 0.0309250448, -0.0262267292, 0.0077578006, -0.0118258651, 0.0468649305, -0.0108114397, -0.0007531781, -0.052726537, 0.0388520733, -0.0323186703, 0.0051375255, -0.0388518907, 0.0410794616, 0.0117943808, -0.050982561, -0.0088009704, -0.0002840692, 0.0431858003, 0.0640116557, 0.0618648119, -0.0379992649, -0.0437698402, -0.033030577, 0.032365609, -0.0300928187, -0.0641975626, -0.0152830891, 0.0046280511, -0.0527046025, -0.0472527817, -0.0259783845, -0.0105637582, -0.0193684436, -0.0089487517, 0.0248107947, -0.0079826964, -0.034734901, -0.0294471867, 0.0074334564, -0.0538167357, 0.0600135624, -0.053071253, 0.0001382836, -0.0082697412, -0.014300948, 0.018764127, 0.0276649781, -0.0935084671, -0.0829042196, 0.0158264246, 0.0206611231, 0.0236889496, 0.0059140529, -0.0197546482, -0.0679691732, 0.0448695011, 0.0488365814, -0.0239088107, 0.045502618, 0.0182501003, -0.0052966583, 0.0041889604, 0.0180523973, -0.002547998, 0.0177524164, 0.0309880264, -0.0294366218, -0.0216651745, -0.0052864049, 0.0175558478, 0.0379811823, 0.0260892175, -0.0108965589, -0.0060677784, -0.0219900329, -0.0644586459, -0.0283506345, -0.000217932, -0.0450547859, -0.0088649644, 0.0295862909, 0.0194913726, -0.0039400049, 0.0154886553, -0.0643485636, 0.0582509711, -0.0583109222, 0.0321568102, -0.0335430056, 0.0647489354, 0.0295636375, -0.0095246453, -0.0073803961, 0.0380062833, -0.0115725193, 0.0219641197, 0.0190112535, 0.0382348076, -0.0388751067, -0.0076337517, 0.0328493193, 0.0467924736, -0.021137407, -0.0140204486, -0.0585461035, 0.0232739393, 0.0531865805, -0.0196381547, 0.1014972851, 0.0019079613, 0.0120815588, -0.0202818234, 0.0134371649, -0.0139550306, -0.0386958048, -0.0096248882, -0.0206536762, -0.0465839058, 0.031847585, 0.0816241577, 0.0310317539, 0.0491254479, 0.0054813232, -0.0132358745, -0.0108890804, -0.0195294674, 0.0176290944, 0.0076414747, -0.0075768121, 0.009746558, -0.0488728546, -0.0694387481, 0.0109168207, -0.0337155201, 0.0067859306, -0.0023822433, -0.0195806585, 0.0725045353, -0.0221612286, 0.0037163678, -0.0646370873, -0.0202108473, 0.0079633472, -0.0422051847, 0.0049092863, -7.075581147e-33, 0.0318074226, -0.0656777844, -0.0196266286, -0.0327433124, -0.0009681582, 0.0000630288, -0.0123992749, -0.0042021987, -0.0083974823, 0.0022089346, -0.0147778168, 0.0295802951, 0.0058468264, 0.0071335933, -0.0211536288, 0.0304717775, 0.0492298342, -0.0165675543, -0.0372755975, 0.0494489074, 0.0515241213, -0.018980097, 0.0070372182, -0.0933161378, 0.0350864753, -0.0669070333, -0.0278296266, 0.004240375, -0.0077777812, 0.0480264574, -0.0215938352, 0.0606140681, 0.0375059955, -0.0157114249, -0.0144764287, 0.0502186343, 0.0090833437, 0.0045920648, 0.0189866461, -0.0220614597, 0.0098608537, -0.090508379, 0.0434337407, -0.0413250402, -0.0538388789, 0.0333341807, 0.0174581893, 0.0042259693, 0.0479282029, -0.0435542092, -0.0494149886, 0.0367994234, -0.0055265208, -0.0124653392, 0.0095745428, 0.0722367167, 0.0170114301, 0.0204753876, -0.0620847829, 0.048367694, 0.0256789736, -0.0110664684, 0.071631521, -0.026672665, -0.0183697883, 0.0199350622, 0.0170857869, 0.0280593336, -0.0105639482, 0.0125829158, -0.0208496191, 0.0366937555, 0.0536114573, 0.0942416415, 0.0134491399, -0.0632677972, -0.0410059839, 0.0155746276, 0.0007735353, 0.002892277, 0.0504330322, -0.0237939879, -0.0463430956, -0.0254206695, -0.0140450662, -0.0634381995, -0.0368109792, -0.0323571339, 0.0188203938, 0.0019601765, -0.0413122438, -0.0252706576, 0.0429558791, -0.0608644336, -0.0052509117, 0.0174671076, -0.0616262406, 0.0294684861, 0.0152564822, 0.045410376, -0.0503848195, -0.0221232325, 0.0559884012, -0.0104043484, 0.0146327168, -0.0005499961, 0.0197492465, 0.0317297466, -0.0505256392, 0.0192916896, -0.0029785288, 0.0596626066, -0.0108464267, -0.0098933782, 0.0255122427, -0.0130900564, -0.0086139608, -0.0069494951, 0.0447643325, -0.0072024236, -0.0289684888, 0.0002417214, -0.0401183926, 0.008108723, -0.0306216609, 0.0157349817, -0.0351877697, 0.0216531325, 0.0375689492, -0.0477244072, -0.0054972935, -0.0321472362, 2.959e-7, 0.0005271136, 0.0803425536, 0.0008996273, -0.014047889, 0.0541681163, 0.036987029, 0.0278130956, 0.0158476103, 0.0039959257, -0.0325175337, -0.038397897, 0.0253076591, 0.080673866, 0.0018393911, -0.0380883217, 0.0239017028, -0.0629510358, 0.0123404348, -0.0685692206, 0.0007685002, 0.0363378748, 0.0406400934, 0.0083462168, -0.0313265473, -0.014340179, -0.0735974759, 0.0583312213, -0.0076295668, -0.02254113, 0.0367453098, 0.0797832906, -0.0011076081, 0.0208447296, 0.0025324542, -0.018803997, 0.025733579, 0.0029666268, 0.0677963197, 0.0363379531, 0.0516709872, 0.0138487034, -0.0658079982, 0.0069225607, -0.0553078875, 0.0664011315, -0.0150322979, -0.0478643961, -0.0672271773, -0.0321631469, -0.0536561869, 0.0139492592, -0.0413343683, -0.0327013768, 0.030368289, -0.002795218, 0.0202885307, -0.0285128802, -0.0162381511, -0.0146847917, -0.0607236437, -0.0358473882, -0.0516301915, -0.0560480915, -0.0006511126, 0.044555176, 0.0009283327, 0.0695631281, 2.685429472e-34, 0.0217798501, -0.0262993984, 0.039009247, -0.0050659063, 0.0422273129, 0.0435930677, 0.0084310379, 0.0218863804, -0.0035498876, -0.0610003769, 0.0265820697 ]
http://arxiv.org/abs/2406.17962v2
http://arxiv.org/pdf/2406.17962v2
SimsChat: A Customisable Persona-Driven Role-Playing Agent
2024-06-25T22:44:17
2024-06-30T21:15:47
Bohao Yang, Dong Liu, Chen Tang, Chenghao Xiao, Kun Zhao, Chao Li, Lin Yuan, Guang Yang, Lanxiao Huang, Chenghua Lin
cs.CL
cs.CL
Large Language Models (LLMs) possess the remarkable capability to understand human instructions and generate high-quality text, enabling them to act as agents that simulate human behaviours. This capability allows LLMs to emulate human beings in a more advanced manner, beyond merely replicating simple human behaviours. However, there is a lack of exploring into leveraging LLMs to craft characters from several aspects. In this work, we introduce the Customisable Conversation Agent Framework, which employs LLMs to simulate real-world characters that can be freely customised according to different user preferences. The customisable framework is helpful for designing customisable characters and role-playing agents according to human's preferences. We first propose the SimsConv dataset, which comprises 68 different customised characters, 1,360 multi-turn role-playing dialogues, and encompasses 13,971 interaction dialogues in total. The characters are created from several real-world elements, such as career, aspiration, trait, and skill. Building on these foundations, we present SimsChat, a freely customisable role-playing agent. It incorporates different real-world scenes and topic-specific character interaction dialogues, simulating characters' life experiences in various scenarios and topic-specific interactions with specific emotions. Experimental results show that our proposed framework achieves desirable performance and provides helpful guideline for building better simulacra of human beings in the future. Our data and code are available at https://github.com/Bernard-Yang/SimsChat.
SimsChat: A Customisable Persona-Driven Role-Playing Agent: Large Language Models (LLMs) possess the remarkable capability to understand human instructions and generate high-quality text, enabling them to act as agents that simulate human behaviours. This capability allows LLMs to emulate human beings in a more advanced manner, beyond merely replicating simple human behaviours. However, there is a lack of exploring into leveraging LLMs to craft characters from several aspects. In this work, we introduce the Customisable Conversation Agent Framework, which employs LLMs to simulate real-world characters that can be freely customised according to different user preferences. The customisable framework is helpful for designing customisable characters and role-playing agents according to human's preferences. We first propose the SimsConv dataset, which comprises 68 different customised characters, 1,360 multi-turn role-playing dialogues, and encompasses 13,971 interaction dialogues in total. The characters are created from several real-world elements, such as career, aspiration, trait, and skill. Building on these foundations, we present SimsChat, a freely customisable role-playing agent. It incorporates different real-world scenes and topic-specific character interaction dialogues, simulating characters' life experiences in various scenarios and topic-specific interactions with specific emotions. Experimental results show that our proposed framework achieves desirable performance and provides helpful guideline for building better simulacra of human beings in the future. Our data and code are available at https://github.com/Bernard-Yang/SimsChat.
[ 0.0605549701, 0.0405835584, -0.0435984954, 0.024680756, -0.0062871003, 0.0060907239, 0.0193379633, -0.0141471596, 0.0425119996, 0.0177973025, -0.0253438968, -0.0476705208, -0.0138226599, 0.0553234331, 0.0061233742, -0.0685062781, 0.0190761294, -0.0299715642, 0.0341091938, -0.0008900288, 0.0080448827, 0.0492918603, -0.0149660585, 0.0613056272, -0.009716033, -0.03199194, -0.0219664648, 0.0084649436, 0.0396439619, -0.0217297506, 0.0329408683, 0.0057150209, 0.0227127839, 0.0047044419, 0.000002078, -0.0203564409, 0.0158260707, -0.0306511503, 0.001758443, -0.0111453682, 0.0833933502, 0.0631205812, -0.0172105152, 0.0374593809, -0.0435849428, 0.0006949027, 0.0482754521, 0.0446971767, 0.1018838286, 0.0762471035, -0.0204117745, -0.0314195678, -0.0539549924, -0.0165028777, 0.0248320866, -0.032440491, -0.0009778868, 0.045287136, 0.0561832003, -0.0106945047, -0.0049061934, 0.0131195318, -0.0143762911, -0.0241498947, -0.0171556845, 0.0436742567, 0.0134590315, -0.0159067027, 0.0192844588, 0.0700075701, 0.0083830971, -0.0040075867, 0.0313660391, 0.0136088226, 0.0423049144, 0.0144898519, -0.0160410982, 0.0350863263, -0.0097538391, -0.0076608979, -0.0161715224, 0.0359499753, -0.0022897171, -0.0141745135, -0.0117462734, -0.0037312317, 0.0089520719, -0.0130843911, -0.0561574772, -0.0247723311, 0.0213966072, -0.0716009364, -0.012091035, 0.009571637, -0.0520498119, 0.0393304043, -0.0695887581, 0.0546337217, -0.0093418565, -0.0214598235, 0.0258452799, 0.0458676666, 0.0053911707, 0.024355853, 0.0445639491, 0.0144536132, 0.014328056, -0.0214762352, -0.0217577629, -0.0043045538, -0.0337535962, -0.0335354097, -0.0332732312, 0.0384489782, -0.0368617177, -0.0320464224, -0.0388095006, 0.0719148889, -0.0282158777, 0.0247268807, 0.0084360344, 0.0173885636, -0.0351930819, 0.0189476926, -0.056007009, 0.0000701096, -0.0372171141, -0.013670207, 0.0099924775, -0.0555845983, 0.0372516215, -0.016196942, 0.0065401108, 0.0101475595, 0.0320814401, 0.0907070711, -0.0059257764, -0.0120752482, 0.0387945622, -0.0268639363, 0.0287597757, -0.0566010438, -0.0411453806, -0.0428413935, 0.0506534353, -0.0300042238, 0.0207885113, -0.0446323752, 0.0103220521, 0.022465203, 0.0136606032, 0.0762322545, 0.0027052287, 0.0317078754, 0.0317610241, 0.0301278587, -0.0062821633, 0.0266080927, 0.0426631048, 0.0284580868, 0.0060873758, -0.0070247068, 0.0532953814, 0.044739984, -0.0592023358, -0.0237772092, 0.0178830102, -0.0044102981, 0.0035956209, 0.0596736223, 0.0096274139, 0.0347144306, -0.0606191643, -0.0238404498, 0.0748976991, 0.0616992079, 0.0147865489, 0.0361689292, 0.0420511626, 0.0728142709, 0.0319511592, -0.0384429842, -0.0017081538, 0.0442707539, 0.0396065004, -0.0230729114, -0.0731835887, 0.033426594, -0.0203426667, -0.0214417484, 0.0448033474, -0.0238678511, 0.0295106694, 0.0184243098, -0.0093214856, -0.0129488092, -0.0423377715, 0.0278732609, 0.0276970249, -0.0880405232, -0.0157852937, -0.0539190546, 0.0327332392, 0.1489654332, 0.0429341383, -0.0356905647, -0.0502507202, -0.0200995598, -0.0213639103, 0.0426726267, 0.0143613797, -0.027977908, 0.0208284892, -0.0398391932, 0.0287381578, 0.0020318134, -0.0182710495, 0.0571186133, -0.0002021938, -0.0071500274, -0.0174981207, 0.0607828423, 0.0211008471, 0.0066516208, -0.0805476829, -0.0012188834, -0.010240254, 0.0087377522, 0.0104581369, 0.0034787208, -0.0106198015, -0.0407998227, 0.0126424283, -0.0115882251, -0.0119390804, 0.0811494216, -0.0412500501, 0.0022043989, 0.0489017069, -0.0544339158, -0.0162026919, 0.0337955914, 0.0093789417, -0.0054922444, 0.0135884723, -0.0404304005, 0.0437269993, -0.0043100724, 0.0273271501, -0.0173616335, 0.0570695996, -0.004003061, 0.0292463563, 0.0199334826, -0.0381634682, 0.0091788182, 0.0254953168, -0.0213974919, -0.0647957772, 0.0113077667, -0.0010236428, 0.0011887469, 0.0207717568, 0.0203003027, 0.0018105201, -0.0548549183, -0.0543533005, -0.0914560407, 0.0066955923, 0.0390101671, -0.0323530622, 0.0337708332, 0.0311570857, -0.055315353, -0.0140881445, -0.0688097849, 0.0042198254, 0.0133816302, -0.0264044851, 0.0391683914, -0.0074931681, -0.0098314574, -0.0547402501, 0.0067485739, -0.0059695025, 0.0255477764, -0.0269048102, 0.0355505906, -0.0154501507, 0.0025203156, -0.0358122624, 0.0361062773, -0.031456314, -0.0645240918, 0.0369422659, -0.0323248804, 0.0358735621, 0.0279931836, 0.0394406244, 0.0417520814, 0.0185924862, -0.0315593146, -0.0489708818, 0.024236802, 0.0063416492, 0.0180715881, -0.0513222143, -0.0593222268, 0.0280326027, 0.0456957929, -0.0118907355, 0.0338466987, -0.0768667459, 0.0474701673, 0.0471461266, -0.0389173217, 0.0503772758, 0.0018656716, -0.018149877, -0.0179281924, -0.0074356166, 0.0193289742, 0.0393553898, -0.0063358513, -0.0281285495, -0.0043270946, -0.005290173, -0.0132035874, 0.0801776201, -0.0047382587, 0.0147070726, 0.0050042165, -0.017782893, -0.0155631592, -0.0273622759, 0.0161706973, -0.0484742895, 0.0160371196, -0.0565144904, -0.0268414468, 0.0359355584, -0.0080752224, -0.0355063565, -0.0481398292, -0.0400363952, -0.0407900102, 0.0284496434, -0.0417769291, -0.0422073342, -0.01968343, -0.0036362505, 0.0130692497, -0.0011142142, -0.0117762927, -0.0304052383, -0.0267357323, 0.0841982514, 0.0255288072, -0.0218080264, -0.0199296437, 0.0894201025, 0.0267462749, -0.074181132, -0.0245728679, -0.001961031, -0.0082347086, 0.0244894642, -0.0340835713, -0.0002826766, 0.0084304186, 0.0302751102, -0.0414967611, -0.0081360172, 0.12496759, 0.0082520293, -0.0082433373, 0.0264567267, -0.0076581291, 0.033637166, 0.0221195165, 0.0481842011, -0.0636265203, -0.007388161, -0.0046212785, 0.0205603987, -0.0249682982, -0.0439071395, -0.0174216386, -0.090874739, 0.0237727836, -0.0000868369, -0.0597524978, -0.0078375554, 0.0104608294, 0.0036754452, -0.0193536021, -0.0063204113, -0.0183348749, 0.0122595411, -0.0136050535, -0.0225888602, -0.0584147051, -0.0238607991, 0.0604930259, -0.046500776, 0.1188957393, 0.0286459941, 0.0978531688, 0.0279907007, -0.0157655589, -0.0553719886, 0.0661952868, -0.0018958374, 0.0137536414, 0.0448725447, -0.0189956129, -0.0363768339, -0.0640350357, 0.02271308, -0.0057139397, -0.0082701221, 0.021402508, 0.0240149517, -0.0170267671, -0.0243509244, -0.0325689651, 0.0144757722, 0.0343215056, 0.0453820974, -0.0391050279, -0.0905622616, -0.0174445771, -0.0170747899, 0.0070557711, -0.0457327403, -0.0890569314, -0.0190595686, 0.0064879181, -0.0156522933, -0.0627221912, -0.0610625073, 0.0156124625, 0.0267756451, -0.0050563398, -0.0332504734, 0.0043675322, 0.0314356014, -0.0049178656, -0.023238631, -0.0678155348, 0.0220543947, -0.0138578508, -0.0669208542, 0.0430462956, -0.0935238302, 0.0216058828, -0.021248119, -0.0424923003, -0.0293241944, 0.0192833152, 0.0394270644, 0.0175479576, 0.0433462448, -0.0057377825, -0.0536317565, 0.0041922065, -0.0000931905, -0.0687586516, 0.0024104009, 0.0043299343, 0.0570657179, 0.0043156836, 0.0028586921, -0.0228990726, 0.0366532169, -0.0395056531, 0.0160754584, -0.0476007946, -0.0629059076, 0.0074306014, 0.0227651168, -0.0563810505, -0.0189246442, -0.0181381758, -0.0065735555, -0.000530423, -0.0756257698, -0.0051898924, -0.0303568132, -0.0278847776, 0.0258956291, -0.0026342641, 0.0107661793, -0.0441157483, -0.0401202217, -0.0097612524, 0.00083789, 0.0008697563, 0.0108440146, 0.0073658726, 0.0106144501, 0.0182207227, 0.0067012357, 0.0260355715, -0.0338642076, 0.0408947729, 0.0105209295, 0.0274963677, -0.0347558744, -0.0611290671, 0.0035765595, 0.0592516847, -0.0380783118, -0.0441589504, 0.0289209075, 0.0278385989, 0.0263804477, 0.0111807361, 0.0373818278, -0.0363520607, 0.019292457, -0.0063475715, 0.011876245, 0.0064662327, -0.0484154969, -0.0333987251, 0.0280262064, -0.0220485702, -0.0327156484, 0.0203111507, 0.0036722329, 0.0406013764, -0.0105945114, -0.0412072688, -0.0664574876, -0.0042139115, 0.0233909842, 0.0146935331, -0.0342013426, -0.011635025, -0.0886713862, -0.0313309878, 0.0050643482, 0.064501524, -0.0115338368, -0.0390408151, 0.0173274633, -0.0053858822, -0.0049580289, 0.0325516872, -0.0643025637, 0.0314572006, -0.037295524, 0.0122151766, 0.0003012656, -5.294537493e-33, -0.0293096006, 0.0006492891, 0.0395832807, 0.0220877808, -0.0551408455, -0.0273096059, -0.037639413, -0.0054121893, 0.0112219378, -0.0030149829, -0.0104520191, 0.0060774814, -0.0024496661, 0.0049033272, 0.0005049513, -0.0129209217, 0.0612947308, 0.0184412543, -0.0187031981, 0.0373468436, 0.020636864, 0.0007347291, 0.0171153583, -0.1024905816, -0.0221239496, -0.0371604413, -0.0317977183, 0.0010784407, 0.0403084569, 0.0592821427, -0.0030346066, 0.0038045885, -0.0021742061, 0.0001654367, -0.0042418158, 0.0229436178, -0.0550749227, -0.0091250017, -0.0037933146, 0.0170875713, -0.0018897168, -0.0487645492, -0.0137919765, -0.011749899, 0.0165629908, -0.0184987001, 0.0504662618, -0.0206264723, 0.039744433, 0.0234057046, -0.0736382306, -0.0068558473, -0.0493660644, -0.0533675216, -0.0063967588, 0.0468633138, -0.006960223, -0.01262201, -0.0499481335, 0.0482969247, -0.0183775015, -0.01651242, -0.0010121389, 0.05059205, 0.0451568961, 0.0784237087, 0.0598801747, 0.0422901437, -0.0312214531, 0.0055375085, 0.0332776383, 0.0634146705, 0.0147542935, 0.0706528649, -0.0246579479, -0.0441123769, -0.0005184321, 0.0274507497, 0.0702795759, 0.0031547511, 0.0370045528, 0.034484271, -0.0383309238, 0.0167563502, -0.017306149, 0.0404492877, 0.004618099, -0.0546459667, 0.0126446737, -0.0114633795, -0.0467614159, -0.020053912, 0.0106112137, -0.0499848127, -0.0006009773, -0.0017013358, -0.0457023829, 0.0113842553, -0.0103676068, 0.0105188927, -0.0086028203, -0.0324055664, 0.0477664396, -0.0153966239, 0.0388012789, 0.060519129, -0.0306636263, 0.062472228, -0.0431646705, -0.0222162809, -0.0123754134, 0.0395272858, -0.0212540179, 0.0232228599, -0.0077774059, -0.0130956788, -0.0440779664, -0.0254591815, -0.000958245, -0.0099506313, -0.0201977734, 0.0080698542, -0.0045924811, -0.0140567441, -0.0269584842, 0.0097411955, -0.089062959, 0.0437423065, 0.0339614302, -0.0349195115, -0.0079926187, -0.0167671163, 2.825e-7, -0.0497314259, 0.0176372044, 0.0037755782, 0.0426507816, 0.0316147916, 0.0585747845, 0.0008840626, 0.0296872985, 0.0155478455, -0.0287562069, 0.0362307541, 0.053689234, 0.0374762006, 0.0110433325, -0.0981372148, 0.0108843567, -0.0013665397, -0.0106982654, -0.0260655172, 0.0076548546, 0.0791282058, 0.0714581758, 0.0557733588, -0.006925425, 0.030626975, 0.0452129059, 0.0286942814, -0.0579956397, -0.0242998227, 0.0333643481, -0.0189574379, -0.004812574, -0.0229264442, -0.00036194, 0.0097221956, 0.0113965217, 0.0091272769, 0.0534198992, -0.0303474776, 0.0117525253, 0.00515889, 0.0017094023, -0.0248589125, -0.006916035, 0.0315709561, -0.0085755903, -0.0347914547, -0.0452253483, -0.0267458204, 0.0022929909, -0.0186291635, -0.0268243849, -0.0580986701, 0.0040386454, -0.0091108959, -0.0205513835, 0.0079941116, -0.0677260309, -0.0012261404, -0.0224774424, -0.0596641041, -0.0340041704, -0.0291298572, 0.0051259245, 0.014508618, -0.0204918236, 0.04749671, 1.976410539e-34, -0.0310869552, 0.0575834848, -0.0181390792, 0.0168098751, 0.0500247106, 0.0218377691, 0.0555998906, 0.0404518694, 0.052374769, -0.046554286, -0.0132380333 ]
http://arxiv.org/abs/2406.17961v1
http://arxiv.org/pdf/2406.17961v1
NormTab: Improving Symbolic Reasoning in LLMs Through Tabular Data Normalization
2024-06-25T22:40:03
2024-06-25T22:40:03
Md Mahadi Hasan Nahid, Davood Rafiei
cs.CL
cs.CL, cs.AI, cs.DB, cs.IR
In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in parsing textual data and generating code. However, their performance in tasks involving tabular data, especially those requiring symbolic reasoning, faces challenges due to the structural variance and inconsistency in table cell values often found in web tables. In this paper, we introduce NormTab, a novel framework aimed at enhancing the symbolic reasoning performance of LLMs by normalizing web tables. We study table normalization as a stand-alone, one-time preprocessing step using LLMs to support symbolic reasoning on tabular data. Our experimental evaluation, conducted on challenging web table datasets such as WikiTableQuestion and TabFact, demonstrates that leveraging NormTab significantly improves symbolic reasoning performance, showcasing the importance and effectiveness of web table normalization for enhancing LLM-based symbolic reasoning tasks.
NormTab: Improving Symbolic Reasoning in LLMs Through Tabular Data Normalization: In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in parsing textual data and generating code. However, their performance in tasks involving tabular data, especially those requiring symbolic reasoning, faces challenges due to the structural variance and inconsistency in table cell values often found in web tables. In this paper, we introduce NormTab, a novel framework aimed at enhancing the symbolic reasoning performance of LLMs by normalizing web tables. We study table normalization as a stand-alone, one-time preprocessing step using LLMs to support symbolic reasoning on tabular data. Our experimental evaluation, conducted on challenging web table datasets such as WikiTableQuestion and TabFact, demonstrates that leveraging NormTab significantly improves symbolic reasoning performance, showcasing the importance and effectiveness of web table normalization for enhancing LLM-based symbolic reasoning tasks.
[ 0.0261214729, 0.055632174, -0.0390479304, 0.014897326, -0.0535171665, 0.0439601243, 0.0019634061, 0.0320287682, 0.0226896163, -0.0304871239, 0.0117782066, -0.0266424585, -0.0126315095, 0.0312022083, 0.0116162067, 0.0175897274, 0.0508345775, -0.0196599551, -0.014311539, 0.0262827557, -0.0102652358, 0.0024281964, -0.014847571, 0.0093600564, 0.0057474924, -0.0101810414, -0.0019706092, 0.0051256712, -0.0018084047, -0.0108970851, -0.0262541436, 0.0547695532, 0.0033259213, 0.0061851065, 0.0000021086, -0.025068013, -0.0516708456, -0.0206824988, -0.010189238, 0.0112791397, 0.0175777078, -0.031106934, -0.0231341571, 0.0364688076, -0.0368778892, 0.0078412155, 0.0059278915, 0.0097314846, 0.0318454988, 0.0508663319, -0.0012497806, -0.0365049727, -0.0108464658, 0.0539572053, 0.0520358235, -0.0039168582, -0.0179997683, 0.0490621142, 0.0128541337, -0.0571512803, 0.0406083651, 0.0407193489, 0.0394948944, -0.0223562643, -0.021290794, 0.0318823345, -0.0402773209, 0.010233731, 0.0000453431, 0.0611593127, 0.0171822477, -0.0271273497, 0.0366047435, 0.0160236862, 0.0127292778, 0.0103742871, -0.0505835451, 0.0497919507, -0.0250463728, 0.001969611, 0.0326739214, -0.0193102602, 0.016653236, -0.0069241235, -0.0658473447, 0.0984633341, -0.004647492, -0.0222676992, -0.0349082537, -0.0573639199, -0.017704891, -0.0606710054, 0.0431754179, 0.0893474743, -0.008187321, 0.0395573974, -0.0160114486, -0.0204593278, 0.0400629193, -0.0565716587, 0.0402923934, 0.0733581185, 0.0097629977, 0.0233336575, -0.0355131589, 0.0491832122, 0.0047677793, 0.0558783747, -0.0542176329, -0.0183631442, -0.062191315, -0.0158051867, -0.0264236573, 0.037795227, -0.0385890603, 0.012783289, 0.0271807816, 0.0378452204, -0.0057004127, 0.0589404367, 0.0187851954, -0.0207678583, 0.0208324473, -0.0090244627, -0.0149378423, 0.0105424374, -0.0300426707, -0.029431073, -0.032283701, -0.0661820024, 0.0057582171, -0.0226781927, 0.0105987415, -0.0122035164, 0.0438566469, 0.0988170207, -0.0156475492, 0.0360312425, -0.0621682815, -0.0497330055, 0.0108613484, -0.0485912226, -0.0128763607, -0.0706671327, 0.027641397, 0.0590318739, -0.0256482698, -0.0496806614, 0.015573347, 0.0442068502, 0.0065271067, 0.0545966513, -0.0108804721, 0.0288222842, -0.0004297076, -0.0062500094, -0.0351005942, 0.0549460053, 0.0539144613, 0.0100116152, -0.0181848239, -0.0158247706, -0.0180128254, 0.0434128344, 0.0038088902, -0.0166444667, -0.0740788579, 0.0416880883, 0.0412143879, 0.0307979751, -0.0148350112, 0.0417276546, -0.0027653009, 0.0245167892, 0.0182531811, 0.0831784979, -0.00417034, 0.0419668965, -0.0018739138, -0.017069811, 0.0362314172, -0.0084490813, -0.0499520488, 0.0295240916, -0.0445788912, -0.004551583, -0.0318484418, 0.0581261665, 0.0038584396, -0.0605754778, -0.0415982045, 0.0077221915, 0.018253509, -0.0103378156, 0.0301526804, -0.0197820365, -0.0081032971, 0.0508686379, 0.0220511835, -0.0807047933, -0.0207244381, -0.0114861643, 0.0411413275, 0.0661915168, 0.0614308305, -0.0501562208, -0.0458687283, -0.0124307731, 0.0005652469, -0.0167613495, 0.0296814311, -0.0396720134, -0.0445606187, -0.0154634137, -0.0289660152, 0.0420369618, 0.0076797912, -0.0074179638, -0.005601217, 0.0019567327, 0.024018731, -0.095055528, -0.0350623876, 0.0271283947, 0.0176466629, 0.0172543731, -0.0242587831, 0.020815691, 0.0038862179, 0.0277854744, 0.0056462279, 0.0375759788, 0.0090508498, -0.0486545861, 0.0008137823, -0.0072953478, -0.0276709255, -0.0105588073, 0.0758353099, 0.0020041657, -0.0324410163, 0.0194472838, -0.0310790744, -0.0249145776, 0.0405037738, -0.0078716353, -0.0123358164, -0.0060405438, 0.0127560291, -0.0175579675, -0.0299509391, -0.0097263902, 0.0580290854, -0.0196991023, 0.0112902848, 0.0236247722, -0.0579362884, 0.0232909843, 0.0324786827, 0.0399581529, 0.0389454886, -0.0179386418, 0.0242883042, -0.0180348009, 0.027490098, -0.0510790609, -0.0063930689, -0.0834673196, 0.0303312186, -0.0226476602, 0.0463108644, 0.0468416587, 0.0223481394, 0.0018033665, -0.0108740218, -0.0656997263, -0.0079457201, 0.0038117301, -0.0078426404, 0.0653132573, -0.0196102541, -0.0624578856, -0.0097404001, -0.0052835271, -0.0650773346, 0.0193027221, -0.0087046847, 0.0031887041, 0.000293526, -0.0013520446, -0.0199904796, 0.0064085112, 0.007384121, -0.0218746271, 0.0008804657, -0.0451713502, 0.0209821668, 0.0437821485, -0.0305715352, 0.0171423629, 0.0333041884, -0.0165713895, -0.0738107637, -0.0045400052, 0.0608315356, 0.0242005475, -0.0079868743, -0.0866655558, 0.0102468655, 0.0840604082, -0.0310195722, 0.0054442687, -0.0122283781, 0.0185657088, -0.0076518091, 0.0394892804, -0.0682566017, 0.0372856185, -0.0078000445, 0.042670399, -0.0175488032, -0.0409814641, 0.0486589596, -0.0359201916, 0.0126761962, -0.0402105004, -0.005258684, -0.0417531021, 0.0950826705, -0.026813969, 0.012108881, -0.022410946, -0.0542135797, -0.0423508212, 0.002948449, 0.0183659997, -0.049453672, 0.1123668179, -0.0618336387, 0.0163165312, -0.0244692061, -0.009109593, -0.0300163347, -0.0049434998, -0.0190770943, 0.0158764794, 0.0128536737, -0.0054903049, -0.0050141118, -0.0030774337, 0.0315114483, 0.0369235128, 0.0027392188, -0.0112712374, -0.0269290041, 0.0122306962, 0.0011801822, -0.0230114348, -0.008470661, 0.01365826, 0.0180741772, 0.0482715815, 0.0221209116, -0.0419009551, -0.0070887748, 0.0255651176, 0.0175848491, -0.0483018644, -0.01198838, 0.0607000776, -0.0219825879, -0.0639658645, 0.0121772913, 0.1125045121, 0.0672201589, 0.007266087, 0.0097444635, -0.0325210616, 0.0079278341, 0.029163558, 0.0484960675, -0.1008514389, -0.0264094435, 0.0180199724, -0.0092717819, 0.0182921626, -0.0506622195, -0.0166553147, -0.0744473562, 0.0033415784, 0.0134703731, -0.0399605781, -0.0527043939, -0.0213907547, -0.0058692247, -0.0010486142, 0.0348851345, 0.1008149534, -0.0227342136, 0.0095442748, 0.0128048612, -0.064508304, -0.0438430123, -0.0032366239, -0.0459770113, 0.0347517468, 0.0227361042, 0.0204017367, 0.001484688, -0.0361934379, 0.0238241926, 0.0319471322, 0.0129279662, -0.0496930405, -0.006454011, 0.0415205099, -0.0015891626, -0.0266902652, -0.0012274787, -0.0659148395, -0.005904675, -0.000242954, 0.0932439342, -0.0415203422, -0.0013959067, -0.0075176777, -0.0634871125, -0.0346376039, 0.0632722452, 0.0315510668, 0.0125114815, -0.063283056, -0.0607278123, -0.0923530832, -0.0297944918, 0.024991041, 0.0169290453, 0.0201779995, -0.0213030111, -0.0610661246, -0.0137365041, 0.073182866, 0.0373832621, -0.0005848366, 0.0071651218, -0.0196948424, -0.018709885, -0.0771856681, -0.0156761296, -0.062242046, 0.0373803228, 0.0153325973, 0.0037899311, -0.0082745021, -0.0444508418, -0.0196119696, -0.0025377965, -0.0257788543, -0.0369216166, -0.0332890786, 0.0170409307, 0.0285014454, 0.003511535, -0.001828976, -0.0281843469, -0.0357575789, 0.0067914044, -0.0644034892, -0.0556026325, -0.0013269583, 0.0101709338, 0.0406852663, 0.031631723, -0.0238692779, -0.0069321292, -0.0271335468, 0.0436740555, -0.0022283497, 0.0599089451, 0.03904384, 0.0733405724, -0.049132511, 0.0440348238, -0.0588036962, 0.0114031024, 0.0376462899, -0.0655908659, -0.0564560257, -0.006278045, 0.0544772074, 0.0012885814, 0.0036788564, -0.008145405, -0.0313232578, -0.0034877216, 0.0374739915, -0.0064174975, 0.0488173962, 0.0336971991, 0.0172301866, 0.053238105, -0.0535580665, -0.0086338082, -0.0156178484, 0.0102843139, 0.0218969658, -0.0431901775, 0.0132783195, 0.0117072472, -0.018607175, -0.0629900396, 0.0200340897, -0.0384437814, -0.0682863444, 0.0392879732, -0.0095033636, 0.009445006, 0.0418550856, 0.0270475484, -0.0210459027, 0.0023804908, 0.026104968, -0.002621057, 0.032684043, 0.0093058841, 0.0380596519, 0.0118034035, -0.0096893664, -0.0658653527, 0.0723451823, -0.0277327336, 0.0224166512, -0.0305547137, -0.0246612541, -0.0442166813, 0.019416621, 0.0352529772, -0.0091035422, -0.043280527, -0.029174177, -0.0440102555, -0.0564263947, 0.0268246811, 0.0548022427, -0.0120824547, -0.0084772455, 0.0038122309, 0.0738793015, 0.058327239, -0.0095479283, -0.0171272513, 0.0243513659, -0.0270228144, -0.0396047048, -0.0185681991, -6.342103529e-33, -0.033798907, 0.0007745799, 0.0257584602, -0.0383411944, 0.0161953699, -0.0466977805, -0.0367006324, -0.0129502472, 0.0131356129, -0.0128391264, -0.0255741421, 0.0405589268, 0.0177510139, -0.0275625773, 0.0440695286, -0.0705368891, 0.0123786908, 0.0061407443, -0.0249411874, -0.0148159126, 0.0016489585, 0.0247333143, 0.0533287376, 0.0009193568, 0.0224788375, -0.0034003123, -0.0036182182, -0.0145357335, 0.0055724075, 0.0229328573, -0.0057355049, 0.018932268, 0.0230515245, -0.007137306, -0.0267695468, 0.0129816104, -0.0915192887, -0.0422798395, 0.0351649262, 0.0223063398, 0.0044425083, -0.0409226753, 0.054701522, -0.0184919592, -0.0147538688, -0.0090878801, 0.0165689643, -0.0450815037, 0.024449762, -0.0262960736, -0.0155892624, 0.030171955, -0.020822417, 0.0361297801, 0.0310217254, 0.0024453648, -0.0020501777, -0.0276986808, -0.0024202629, 0.0103743812, 0.0322757214, 0.0315740742, 0.0789079294, -0.0275751762, 0.0064608627, 0.0172910299, 0.0198104903, -0.0205455739, -0.0469677001, 0.0367481709, 0.0034260326, 0.0249485206, 0.0417477787, 0.0026567918, 0.040881332, -0.0073961406, -0.0618560985, 0.0680703744, 0.0325142108, 0.0263032503, -0.0200836845, 0.0055150772, -0.0284456369, -0.0089023579, -0.0231885985, 0.0069783134, -0.0392700657, -0.0304346625, 0.0213759411, -0.0437882915, 0.002846627, -0.0263952725, 0.0037335893, -0.0229579601, -0.0548480451, 0.0561437197, 0.0216648448, 0.0118690897, 0.001691154, 0.0068674474, -0.0007093535, -0.0375546403, -0.0056317723, -0.0344476253, 0.0516470671, 0.0672859401, 0.0076765888, 0.0382886045, -0.0748257488, -0.01094779, -0.0670128018, 0.0081117526, 0.0233765878, 0.0124828918, 0.0347108804, -0.001687945, 0.0121231768, 0.0181274805, 0.0157389957, -0.016350532, -0.0407623947, 0.0380325727, -0.0150288381, 0.0114354296, 0.0030406869, -0.0109776268, -0.0465926416, 0.0255467463, 0.0389960445, -0.0283109862, -0.0027450891, 0.0164294448, 2.78e-7, 0.0630563051, 0.0384395495, 0.0239178482, 0.0535012633, -0.0100903166, -0.0017078611, -0.0219680928, -0.0336125158, -0.0191384405, -0.046188239, 0.0299539026, -0.0012517627, 0.0379038826, 0.0533090495, -0.0247516669, 0.0055878139, -0.0453942455, -0.0380967781, -0.0457732715, 0.0463034697, 0.0617955327, 0.0386973359, -0.0195474047, -0.0328903496, -0.0035579232, -0.0302343722, -0.0197715145, -0.0564916432, 0.0564075448, 0.0243938118, 0.0289320126, -0.045405671, 0.0281869136, 0.040551085, 0.0034169238, 0.0147067485, -0.0142648909, 0.0595966391, -0.0159429237, 0.0713055581, -0.0062231836, -0.1333393604, 0.0125023052, -0.0788890645, 0.1120581701, -0.0064080311, 0.0104258945, 0.0624051653, -0.0082079507, -0.0173821542, -0.0298228655, 0.0070517845, 0.0153711969, 0.0578597672, 0.0067405328, -0.0251907855, 0.029177608, -0.0542685539, -0.0422249697, 0.0258588064, -0.0087019177, -0.0163243432, -0.0211631469, -0.0093426537, 0.0069427034, -0.0482133217, -0.0252190866, 3.068848178e-34, 0.0462592654, -0.0225286689, 0.0392224304, 0.0590400361, 0.0635938272, -0.0048000477, -0.0317766108, 0.0047386196, 0.0054363208, -0.0935609043, -0.009672639 ]
http://arxiv.org/abs/2406.17947v1
http://arxiv.org/pdf/2406.17947v1
Do they mean 'us'? Interpreting Referring Expressions in Intergroup Bias
2024-06-25T21:47:53
2024-06-25T21:47:53
Venkata S Govindarajan, Matianyu Zang, Kyle Mahowald, David Beaver, Junyi Jessy Li
cs.CL
cs.CL
The variations between in-group and out-group speech (intergroup bias) are subtle and could underlie many social phenomena like stereotype perpetuation and implicit bias. In this paper, we model the intergroup bias as a tagging task on English sports comments from forums dedicated to fandom for NFL teams. We curate a unique dataset of over 6 million game-time comments from opposing perspectives (the teams in the game), each comment grounded in a non-linguistic description of the events that precipitated these comments (live win probabilities for each team). Expert and crowd annotations justify modeling the bias through tagging of implicit and explicit referring expressions and reveal the rich, contextual understanding of language and the world required for this task. For large-scale analysis of intergroup variation, we use LLMs for automated tagging, and discover that some LLMs perform best when prompted with linguistic descriptions of the win probability at the time of the comment, rather than numerical probability. Further, large-scale tagging of comments using LLMs uncovers linear variations in the form of referent across win probabilities that distinguish in-group and out-group utterances. Code and data are available at https://github.com/venkatasg/intergroup-nfl .
Do they mean 'us'? Interpreting Referring Expressions in Intergroup Bias: The variations between in-group and out-group speech (intergroup bias) are subtle and could underlie many social phenomena like stereotype perpetuation and implicit bias. In this paper, we model the intergroup bias as a tagging task on English sports comments from forums dedicated to fandom for NFL teams. We curate a unique dataset of over 6 million game-time comments from opposing perspectives (the teams in the game), each comment grounded in a non-linguistic description of the events that precipitated these comments (live win probabilities for each team). Expert and crowd annotations justify modeling the bias through tagging of implicit and explicit referring expressions and reveal the rich, contextual understanding of language and the world required for this task. For large-scale analysis of intergroup variation, we use LLMs for automated tagging, and discover that some LLMs perform best when prompted with linguistic descriptions of the win probability at the time of the comment, rather than numerical probability. Further, large-scale tagging of comments using LLMs uncovers linear variations in the form of referent across win probabilities that distinguish in-group and out-group utterances. Code and data are available at https://github.com/venkatasg/intergroup-nfl .
[ 0.036218863, 0.0376335084, -0.0231784917, 0.0131481942, 0.0133059481, 0.0442698523, -0.0232793652, -0.0116231963, 0.0038414977, -0.097073257, -0.0371561833, -0.024290707, 0.0385512933, -0.0276990738, 0.0507676005, -0.0378563665, 0.0212653968, 0.013976803, 0.0627827868, -0.0529590845, 0.0350790396, 0.0094431257, -0.0177984983, 0.0545456484, -0.0797432289, 0.0239688922, 0.0546194054, -0.012682857, -0.0199972652, -0.0332504287, 0.0306548532, 0.0318740606, -0.0300733093, -0.0280717537, 0.0000018626, -0.0527144708, 0.0186715703, -0.0248182975, 0.0478950329, 0.0128063709, 0.0353149138, 0.0265694242, 0.0416452847, -0.0239157584, -0.0374822505, -0.000406534, 0.066110082, -0.0247619487, 0.020246122, 0.0103458017, -0.0085837059, -0.0011244421, -0.1032507569, -0.0093467515, 0.0130505124, -0.0563071668, 0.0587258339, -0.008590579, -0.0869506747, -0.0673996434, -0.004104184, 0.0157873537, -0.0096314466, 0.0212811157, 0.0071038781, 0.0032592933, -0.0265668742, 0.0134892529, 0.0016570989, 0.0001939535, 0.0691675916, -0.018446248, -0.0264504924, 0.0221161209, -0.0055213938, 0.0109394994, 0.0039295447, 0.0262742061, -0.0211913157, 0.0076118801, -0.0328785703, 0.0572377034, -0.0109967282, 0.0139413355, -0.0302442387, 0.0886634886, 0.0220273342, 0.0249955263, -0.0648750216, 0.0796256885, -0.0280879997, -0.0231001414, 0.0317910388, 0.011232188, -0.0283941366, 0.0092329085, 0.0005631371, -0.0241239108, 0.0029352636, -0.0389196835, 0.0379421189, -0.0038646748, -0.0277787969, 0.0203485005, -0.0215743091, -0.0432488285, 0.013833886, -0.0447578877, 0.0287016835, -0.0579332598, -0.0114858346, 0.0031625074, 0.0166169759, 0.0009603598, -0.0279582944, 0.0519973226, -0.043200843, -0.0246909577, -0.0577265248, -0.0053607356, -0.0170966629, -0.0203936622, -0.0171616916, -0.0195758007, -0.0759775788, 0.0262250099, -0.0648956597, 0.0222672448, -0.0505083352, -0.1064751148, 0.0225463193, 0.0245290827, -0.0225031804, -0.0299389083, -0.0073110973, 0.023124408, 0.0436580926, 0.0328213274, -0.0188111253, -0.0219051167, -0.0208877772, -0.0389544144, 0.0007145126, -0.0510222241, 0.0237191133, 0.0155754127, 0.0127832144, -0.0633412302, 0.0214450639, 0.03352575, -0.0238337126, 0.0676615238, -0.0382697023, 0.0341156274, 0.0357527249, 0.0138497949, -0.0216169413, 0.03867203, 0.0102471821, -0.0216886364, 0.0521181636, -0.0093666576, -0.0103595546, 0.0392753892, 0.0014053927, -0.0265515987, -0.040038038, 0.008529257, 0.0156999193, -0.0379186422, 0.0263298228, 0.0576128401, -0.0022070715, -0.0381697863, 0.0691844821, 0.0852426738, -0.0533152595, -0.0111524509, 0.0501310267, 0.0258207638, -0.0312931389, -0.0216125473, -0.0587951206, 0.0298210792, 0.0113179693, -0.0258548614, 0.0467733964, 0.0194139928, -0.0128694139, -0.0503349639, -0.0066759884, 0.0018646835, 0.0340312012, -0.0066388501, 0.0411411934, -0.012380749, 0.0356737077, 0.0636688843, 0.0481841229, -0.001079759, -0.0128801903, -0.0562703535, 0.0213913489, 0.0195874125, -0.022511363, -0.0220615249, -0.0132914707, -0.0197910219, -0.0089622363, 0.0417616405, 0.035582561, -0.0014979072, -0.0154521102, -0.0015173397, -0.0099733369, 0.0365759768, 0.0338907577, 0.0545354337, 0.0245535821, 0.0130391233, -0.0008344025, -0.0092944866, -0.0249571558, 0.0170071293, 0.0017245017, -0.0142996283, -0.0016220066, -0.0039359266, 0.0120010767, 0.0513243936, -0.0254797172, 0.0028146021, -0.015450702, -0.0147692692, -0.008788242, 0.0468198694, -0.014439051, 0.0538763031, 0.0257979035, -0.085064888, -0.0735867769, -0.0154286418, 0.0116929337, -0.0070940689, 0.0381852873, -0.0043678498, 0.0344942212, -0.02425473, -0.058109168, -0.0297994427, 0.0279999003, -0.001079726, 0.0188551396, 0.0145344343, -0.012698357, 0.0117006283, -0.062802352, 0.0624391995, -0.0028205428, -0.0474860296, 0.0048980652, 0.0145948362, -0.0063086017, 0.039538566, -0.0300601125, -0.0707919747, 0.0405096449, -0.1014101133, -0.0030451091, -0.0046306825, 0.0030451869, 0.0158240478, 0.0287026837, -0.0300052278, -0.0288923811, 0.0278042536, -0.0564974546, -0.0304838791, 0.0429252684, 0.0194337554, -0.0020773078, 0.0343199186, -0.0262469631, -0.0160285663, -0.0370603167, 0.0243636612, 0.0255994629, -0.0483570248, -0.0708967447, -0.0690821037, -0.0678276196, -0.027103439, 0.0020753746, 0.0036419407, 0.0492903963, 0.0296611879, 0.035300605, 0.063726604, 0.0318116136, 0.0239225477, 0.010098272, -0.0718296021, -0.0158275105, -0.0241256915, -0.0287745651, -0.0409322716, -0.0365088843, -0.0732246786, 0.0166352019, 0.0834097192, 0.0339601785, -0.0031873572, -0.0595912859, 0.0523788668, -0.0563515946, -0.0331490226, -0.0026255515, -0.0396966338, -0.0250360817, 0.0210630484, 0.001925687, -0.0198238455, 0.0269061383, 0.0014616663, 0.0205798056, -0.00328257, -0.0261123534, -0.0122220237, 0.0654186979, 0.0139815081, 0.0182891283, 0.0274799876, -0.0265632551, 0.0183373783, -0.0324064158, 0.0488376617, -0.0092086308, 0.0355376266, -0.0010925485, 0.0081030251, 0.0096193915, 0.0144097134, -0.008792093, -0.0362663604, -0.0781583488, -0.0452600643, 0.0375105999, -0.0186048411, 0.0169470515, 0.0442990251, 0.0238222219, -0.0928791091, 0.0057788729, -0.0184654538, 0.0190361962, -0.0243800841, 0.0201874673, 0.0174963772, 0.0163807906, -0.0370762087, -0.0453784093, 0.050435096, 0.0311938375, 0.0280315708, -0.0222819764, -0.0103448378, 0.0088393679, -0.0125109674, -0.0070153023, -0.0022566991, 0.0876843929, -0.0275716819, 0.0493195951, 0.0874443799, 0.0174430199, -0.0335448869, 0.0069983089, 0.0101320799, 0.0196374673, 0.036385335, -0.0031487735, -0.0988788456, -0.0167007465, -0.0073540797, 0.022087181, -0.0114386873, -0.0668098405, -0.0179848522, -0.0852129236, -0.013339798, 0.0110271126, 0.0289932992, -0.0041288594, -0.012670964, 0.0542808995, 0.0045267739, -0.0170200989, -0.0070848935, 0.0182836074, 0.0457236171, 0.0304535273, -0.0644969195, 0.0047388659, 0.0294715129, -0.0443593413, 0.0860300586, 0.0142616499, 0.0328900702, -0.0001830649, -0.013470605, 0.0272451025, -0.022498304, -0.0185099021, -0.0373837501, 0.0294534117, -0.0396721214, 0.013706293, -0.0052118874, 0.0022075053, -0.037314523, 0.0265516695, 0.0197376404, 0.0203460082, 0.0159844644, -0.0193251539, -0.0245218333, 0.0205245614, 0.0188596956, 0.05121319, 0.0349675529, -0.0298599657, 0.0252489354, -0.0667982101, 0.0205878988, 0.0169649962, -0.0867207646, 0.0034928406, -0.0249864012, -0.0451997668, -0.0079936162, -0.031624537, 0.0296728425, -0.0070327972, 0.0111182528, 0.0234026611, 0.0203440003, -0.0438590646, -0.0358901024, 0.0427537672, -0.0530066937, 0.0302141625, 0.0554336123, -0.0399518237, 0.0571758114, 0.0332541279, -0.0448983684, 0.0741002336, 0.0079216342, -0.020326063, 0.0492899455, 0.0620425195, 0.0184926391, 0.0031042364, 0.0227282401, -0.038476333, -0.0018377298, -0.0048027462, 0.0340138152, 0.0305438451, -0.0182609688, 0.0342994407, -0.006037741, 0.0370109379, -0.0111985113, 0.0810034201, 0.0077120187, -0.0053055952, -0.0546766296, -0.0528848916, -0.0070247068, -0.0283221025, 0.0536013618, -0.0103800436, -0.0234240256, 0.0084974049, -0.0258498788, -0.0157281756, -0.0203323737, 0.0305111781, -0.0581247285, -0.0245857593, 0.006067941, -0.0515119061, -0.0226477943, -0.0454274081, 0.0569525175, -0.0067969211, 0.0840637386, 0.0694495738, 0.0294984877, 0.000069257, -0.01135178, 0.0175459776, 0.0356506594, 0.01936909, 0.0512260795, 0.0223966092, -0.0133587467, 0.0303892344, -0.0138201294, 0.0118139582, -0.0360415541, 0.0069511179, -0.0103532188, 0.0293196533, -0.0457933098, 0.0256546959, -0.0323347822, 0.034195818, 0.0005205778, -0.0085989144, 0.0247222744, 0.0173862092, 0.0046719764, 0.0004022948, 0.0093179904, -0.0039132144, -0.0168641526, 0.0461227745, 0.0550971739, -0.0023332748, 0.0242930055, 0.0042428905, -0.0831912383, -0.040758729, -0.0173026677, 0.0111867534, -0.0053457445, 0.0006953177, 0.0341993421, -0.0781227127, -0.0340066999, 0.0190741122, -0.0255951509, 0.0207745153, 0.0030103028, 0.0200358219, 0.0359155983, 0.0277507734, -0.0038181951, -0.0114607718, 0.0463538468, -0.0532047339, -0.0136761591, -0.0369598232, -6.310298325e-33, -0.0315401405, -0.1017540619, -0.0050076754, -0.0036506215, -0.0364423096, -0.0059108911, -0.0386095755, 0.0048513371, 0.0107891057, 0.0362047292, 0.0047085881, 0.0150445551, 0.0135658551, -0.0135848448, 0.060638018, -0.0476743355, 0.0277655497, -0.0210793689, 0.00468142, -0.0053893486, 0.0284732152, 0.0188852381, 0.0286051091, -0.1159260944, 0.0789307952, -0.0110329371, -0.0370686315, 0.0128013976, 0.0839480311, -0.0067476057, -0.0134199532, 0.0366695821, 0.0276760124, -0.0495389551, 0.0058460054, 0.0721859708, -0.0092984904, -0.0203193687, 0.0269384943, 0.0213265121, 0.0571827665, 0.0273960121, 0.038537167, 0.0115196956, -0.048796244, -0.0010394554, 0.0592699274, -0.0292171426, 0.0706932992, -0.0173233375, -0.0371845551, 0.0375638492, -0.0113294655, 0.0618718751, 0.0274679903, 0.0041298042, -0.0104344757, 0.0487907827, -0.0513075553, 0.0509540997, -0.0546581559, -0.0036303424, 0.0224763155, 0.0439853854, 0.0417107865, 0.0178785659, 0.0613962039, 0.0601239055, -0.0203443337, -0.0398647934, 0.0038093741, -0.0187017769, 0.0055608763, 0.0654092953, 0.0189928412, 0.0176984202, 0.0135140186, -0.0286833104, 0.0556027852, -0.000123696, 0.0322238617, 0.0048505529, -0.0871559232, 0.0015982563, -0.0151421335, -0.0697284192, 0.0056498582, -0.0848369747, -0.0061457311, -0.0507951491, -0.0475706905, 0.0099248337, 0.0474085584, -0.0115948711, -0.0551363975, -0.0241205972, -0.011217894, 0.0006394929, -0.0081909578, 0.013723298, -0.0517440513, 0.0227617491, -0.0252184868, 0.0115574906, 0.042738717, 0.0264939759, 0.0018885202, -0.0024372865, -0.0747431964, -0.0082419328, 0.0230779499, -0.004836238, -0.0463121161, -0.0125098825, 0.0269299187, 0.0489001274, 0.0047590267, -0.0269562695, 0.0275392905, -0.0005306525, 0.0262333676, 0.0057001305, -0.0683663934, 0.0187531933, -0.0363541357, -0.0207150374, 0.0018732765, -0.0065002646, 0.0197542999, -0.0222915336, -0.0212704651, -0.0609407686, 2.542e-7, 0.066335991, 0.0949637741, -0.0262059756, 0.0169175938, 0.0338906087, 0.0776726231, 0.0069076917, -0.0085779261, 0.0358818285, -0.0330091305, 0.0066608819, 0.0428018756, 0.0171710942, 0.0541442744, -0.0592534877, -0.0087401588, -0.0334254317, 0.025530044, 0.005587121, 0.0114342635, 0.0107529992, 0.0154553512, 0.0607430451, -0.0043193838, -0.0520222858, -0.0495009571, 0.0145642404, -0.0814161599, -0.0204728488, 0.0533228777, 0.0147653036, -0.0144106708, -0.0250875149, 0.045313701, -0.0159804709, -0.002091988, 0.0339919329, -0.0499064438, -0.0175112803, -0.0020163185, -0.0457777642, -0.0112131443, 0.0144629441, -0.0180047825, -0.0034062697, -0.0345195234, -0.0373716056, -0.0496195108, -0.0469978191, -0.0379389711, 0.0040148459, -0.0195241719, -0.0175045822, -0.0195886642, 0.0218039788, 0.0335663296, 0.0073119863, -0.0531642102, -0.0047499067, 0.0032586155, -0.0273131132, -0.027270684, 0.0001995905, 0.0361889303, -0.0086713405, 0.000823734, 0.0360819176, 1.980962825e-34, 0.0326328203, -0.0016335041, -0.0096134348, 0.0292671435, 0.0641723946, -0.0070811417, -0.0554972254, -0.0201255903, 0.0742040947, -0.0004826417, -0.0232907441 ]
http://arxiv.org/abs/2406.17935v1
http://arxiv.org/pdf/2406.17935v1
Sequential Editing for Lifelong Training of Speech Recognition Models
2024-06-25T20:52:09
2024-06-25T20:52:09
Devang Kulshreshtha, Saket Dingliwal, Brady Houston, Nikolaos Pappas, Srikanth Ronanki
cs.CL
cs.CL, cs.SD, eess.AS
Automatic Speech Recognition (ASR) traditionally assumes known domains, but adding data from a new domain raises concerns about computational inefficiencies linked to retraining models on both existing and new domains. Fine-tuning solely on new domain risks Catastrophic Forgetting (CF). To address this, Lifelong Learning (LLL) algorithms have been proposed for ASR. Prior research has explored techniques such as Elastic Weight Consolidation, Knowledge Distillation, and Replay, all of which necessitate either additional parameters or access to prior domain data. We propose Sequential Model Editing as a novel method to continually learn new domains in ASR systems. Different than previous methods, our approach does not necessitate access to prior datasets or the introduction of extra parameters. Our study demonstrates up to 15% Word Error Rate Reduction (WERR) over fine-tuning baseline, and superior efficiency over other LLL techniques on CommonVoice English multi-accent dataset.
Sequential Editing for Lifelong Training of Speech Recognition Models: Automatic Speech Recognition (ASR) traditionally assumes known domains, but adding data from a new domain raises concerns about computational inefficiencies linked to retraining models on both existing and new domains. Fine-tuning solely on new domain risks Catastrophic Forgetting (CF). To address this, Lifelong Learning (LLL) algorithms have been proposed for ASR. Prior research has explored techniques such as Elastic Weight Consolidation, Knowledge Distillation, and Replay, all of which necessitate either additional parameters or access to prior domain data. We propose Sequential Model Editing as a novel method to continually learn new domains in ASR systems. Different than previous methods, our approach does not necessitate access to prior datasets or the introduction of extra parameters. Our study demonstrates up to 15% Word Error Rate Reduction (WERR) over fine-tuning baseline, and superior efficiency over other LLL techniques on CommonVoice English multi-accent dataset.
[ -0.0038060404, 0.0754957348, -0.0040951339, -0.0128230955, -0.0063646464, 0.0147690373, 0.0058790124, -0.0461251102, -0.0068528377, -0.00675327, -0.0530883074, -0.0473468304, 0.0326144174, 0.0195848402, -0.0010532343, -0.0659690499, 0.0322353654, -0.0364763625, -0.027481569, -0.0237209089, -0.0152735021, -0.0432790816, -0.0283734798, 0.0166184101, 0.0114684189, 0.009937631, -0.035660252, -0.0034672213, 0.0302138645, -0.0195339806, 0.001818085, 0.0323929563, 0.0370677114, -0.0098623754, 0.0000023612, -0.0418973677, -0.0424729176, 0.0082348781, -0.0671636835, -0.0152095277, 0.0253525879, 0.0244912449, -0.0301359445, 0.0222402737, -0.0226442404, -0.0191798825, 0.0633772612, 0.0303424615, 0.0535869673, 0.0429745913, -0.0002461515, -0.0649994537, 0.0234608091, -0.0378941931, -0.0323786177, -0.0273734368, 0.0403918698, 0.0015410879, -0.0695433542, -0.0465459004, -0.0379533693, 0.0391624421, 0.0046793632, -0.0100062592, 0.002335489, 0.0062471754, 0.014083609, -0.0156134581, -0.0048219161, -0.0030593472, -0.0079988455, 0.0001713982, 0.0100909621, 0.0298088193, -0.0171102397, -0.0455299094, -0.0159427524, 0.0014978454, -0.0076226825, 0.0007665341, -0.0191785637, -0.0528125316, 0.0176971294, -0.0243733134, -0.0538636819, 0.012088811, 0.013236993, -0.0500719845, 0.0097196829, -0.0330374986, -0.0462521017, 0.0015985582, 0.0335453153, 0.0146541819, -0.0016147582, 0.0297246743, -0.0466605388, -0.0275751129, -0.0176772363, -0.0102228783, 0.0297767185, 0.0436208546, -0.0195568614, 0.0100811543, -0.0115577225, 0.0321538709, 0.0548836067, -0.0507941581, -0.0183279514, 0.0385812037, -0.0954967141, 0.0002706095, -0.0518698916, 0.0748095363, 0.0024517633, 0.0170123354, -0.0376181118, 0.020044202, -0.0150757339, -0.0077460222, 0.0143363709, -0.0278454665, -0.0620957129, -0.0033981388, 0.0193855874, 0.0439416729, 0.0281181093, 0.053975895, 0.0050288034, -0.0097949374, -0.0252801143, -0.0247448999, 0.0374187194, 0.0012306498, 0.0178372841, 0.0757233202, -0.0122831743, -0.0339139439, -0.0644036606, 0.0039659478, -0.0224497337, -0.0371214375, -0.0175469648, -0.0229581706, 0.0302018207, -0.0458627939, -0.009872702, -0.093669191, -0.006890378, -0.0280372705, 0.0064108572, 0.0921148509, -0.0202139765, 0.022443153, 0.0564865805, -0.0083788782, -0.0168871842, 0.0631094575, 0.0013642274, -0.0372726843, 0.0101804333, -0.0184415821, -0.0109058209, 0.0349960029, 0.009628606, 0.0231770147, -0.0312747508, -0.0348831937, 0.0260812696, 0.0236294437, 0.0360628068, -0.016974166, -0.0469943918, 0.0407404676, 0.0684554875, 0.0085041169, 0.0446107052, 0.0326625742, 0.0154320616, -0.0245209616, 0.0319104865, 0.009969742, -0.0560831763, -0.0294041615, 0.0122813303, 0.037386097, -0.0131231109, -0.0064493068, -0.0176218208, -0.0335588939, 0.0059554838, -0.0267939288, 0.0039933082, 0.0086313961, 0.0549761318, 0.0288674813, -0.0325410366, 0.0557266064, 0.0071785874, -0.0638871938, 0.0251758676, -0.0285171233, -0.0341592021, 0.0675686523, 0.0167329088, -0.0555068031, -0.0662584603, -0.0446321703, 0.0184713192, -0.0211225599, -0.0058669676, -0.0130878408, 0.0771942139, -0.0660547391, -0.0361968465, 0.0517302044, 0.0034705366, 0.0319346748, 0.011142035, 0.0223303959, 0.004875802, -0.0169181656, -0.0244281329, -0.0013622598, -0.0546141937, -0.0358801, 0.0471190475, 0.0188016426, 0.0318702199, 0.0088642174, -0.0518854894, -0.027238505, 0.0134876445, -0.0146923773, -0.0065703481, 0.0289641544, -0.0693455786, -0.0006288352, 0.0790139735, -0.0011900529, 0.0194191467, 0.0000519856, 0.0030173853, -0.0794105083, -0.0061723506, -0.0198848434, 0.0552611575, -0.0353102237, 0.0769281909, -0.0003703001, -0.0231708866, -0.0191576965, 0.0060256333, -0.0008275772, -0.0344625153, 0.0096121682, 0.0109750936, -0.0120545961, -0.079012163, -0.0131430849, 0.0015243986, -0.0160013661, 0.0215407964, -0.0000430476, -0.0051114867, -0.0288645905, -0.0366294086, -0.0715230405, -0.010462366, -0.0276634116, 0.0326785035, -0.0012687881, 0.0241008885, -0.0170254894, -0.0365139507, 0.0135908416, -0.0025541938, -0.058728274, 0.0146141909, -0.0283302795, 0.0071393633, -0.0193026084, -0.0035146333, 0.0050237509, 0.04509544, 0.0061605568, -0.0580657832, -0.0233342163, -0.077391088, 0.0232888646, -0.0040256749, 0.0353912525, 0.0037323933, -0.0209731013, -0.0577860773, 0.0141676692, 0.0578289665, 0.0457562134, 0.0035839861, 0.0119186789, 0.0434676781, -0.0309594348, -0.0018603038, 0.0074805315, -0.0232557803, -0.012999434, 0.0159490407, 0.0343123078, -0.0351375602, 0.0219059046, 0.009350813, 0.0107984459, -0.0175830629, -0.0290439501, -0.0281715598, -0.0152647961, -0.0474520996, -0.0285429973, -0.0475363284, -0.014842446, 0.0003796015, -0.0288351048, 0.0491927229, -0.0127013484, 0.0041055926, -0.0721472129, -0.0556181818, 0.001095084, 0.0143385716, -0.0210060086, 0.0146187069, 0.0451781899, -0.0592335314, -0.0315592624, 0.0041570715, 0.0565660372, -0.0336284786, 0.0758590326, -0.0571896918, -0.0032024782, -0.0128326425, -0.0430909321, -0.0504950061, 0.0314181671, -0.0148909725, 0.0176976118, -0.0279411525, -0.0114012212, -0.0478139371, -0.0014856166, 0.049338568, 0.089986071, -0.0027870943, -0.01597731, -0.0053227912, 0.032326676, 0.0284702573, 0.06066126, -0.0180057697, -0.0108869793, 0.0609566867, 0.0034531117, 0.0561922714, -0.0423066616, -0.025909204, 0.0106293084, 0.0102645848, -0.0323939994, -0.0026266098, 0.0065228343, 0.0104605258, -0.0548017435, -0.0158621781, 0.072477214, 0.0230717026, -0.025359327, -0.0068341852, -0.0199704375, 0.0438503362, 0.0312910266, -0.0060715778, -0.0453738756, -0.0417462252, 0.0786361247, 0.0729275271, 0.0122208809, -0.0068212994, -0.0121094957, -0.0809386671, 0.0282952711, 0.0362313949, -0.0347890556, 0.0000844276, -0.0439651683, -0.0237060171, -0.0430615507, 0.0349083208, 0.0417311378, -0.0342805944, -0.0227920711, 0.0442015491, -0.0531583391, -0.0728794485, -0.0090520009, -0.0144464001, 0.0561932996, 0.0471589491, -0.0124071874, -0.003228962, -0.0191577077, 0.0722113028, 0.0052524116, 0.0366576314, -0.0162409917, 0.0495221503, 0.0398575589, -0.0083655939, 0.0304450244, 0.0298847426, 0.0149691505, -0.035807278, 0.0656598136, 0.058025118, -0.0264154244, 0.0409591906, 0.0395758338, -0.0397088341, -0.0053329584, 0.0037365935, 0.0008214417, -0.0435549878, 0.017694829, -0.0031586958, 0.0203597918, -0.0236497056, 0.0114762252, 0.0147914141, -0.0769185275, -0.0255928002, -0.0227113906, -0.0024817523, 0.0145363333, 0.043674726, 0.0545427352, 0.0077166455, 0.0006135794, 0.0339592099, -0.0716538131, -0.0133930957, -0.0316446684, 0.0078316787, -0.0815931335, 0.0053044637, 0.0689409822, -0.0551726259, -0.0107614584, -0.0300466381, -0.0252184123, -0.0640394986, 0.0547175892, 0.0285869967, 0.0134998169, 0.0962646008, 0.0051527279, -0.0416729823, 0.0739583373, 0.0296696872, -0.0203610174, 0.0541674607, 0.0764405653, -0.0319537148, -0.0155573273, 0.0181629304, -0.0456479602, 0.0219588336, 0.0009595891, 0.0165414084, 0.0157945119, -0.0246607494, 0.0113969184, 0.0285552386, 0.0456971601, -0.0108502116, -0.0035974709, -0.0421386436, 0.080889374, 0.0355215222, 0.0337850749, -0.0227011852, 0.0246363711, -0.0013302439, 0.0370599739, -0.0014037529, 0.0233599916, 0.0045950529, -0.0300723761, 0.0015462305, 0.0321775228, -0.0019371051, 0.0476051494, 0.0417267866, 0.0282953847, -0.0243985187, 0.0371117927, 0.040994022, 0.0095834043, 0.0036846162, 0.0276845936, 0.0090541197, 0.0192510821, -0.0486256853, 0.0259031728, -0.0361168385, -0.0331315994, -0.0399228297, 0.0097473776, 0.0659490451, 0.0082392925, 0.0166503266, -0.003812857, 0.007983339, -0.0431070998, 0.011298405, -0.0120930141, 0.0184579063, -0.0007576222, 0.0014872774, -0.0058191861, 0.0551525578, 0.062295571, 0.0022045767, 0.0678426176, -0.0360908508, 0.0379014611, -0.0466219559, -0.0219999347, -0.0143887298, 0.0029361192, -0.0267451201, -0.0318682864, 0.0046333591, -0.069444105, 0.0464599356, -0.0337095372, 0.0119492551, -0.0202687457, 0.0360968187, -0.0117491111, -0.0397476889, 0.0225949846, 0.0183553472, -0.011970927, 0.0315957293, 0.0072942092, 0.0361605883, -6.623223737e-33, 0.0023430106, 0.0302606057, 0.0382935181, 0.0282505285, -0.0713542625, 0.0043961159, -0.0236267876, 0.0279914625, 0.0363571867, -0.0154925315, -0.0275749099, 0.0012016097, 0.0292191096, 0.001970462, 0.0257435273, -0.0039676991, 0.1045231223, -0.0006922213, 0.0015045896, -0.0152648669, 0.0227948111, 0.0664871484, 0.0358401686, -0.0580850206, -0.0420182943, 0.0794253722, 0.0017808005, 0.0143458946, 0.0307010803, 0.0480159968, 0.0182021465, 0.029855717, -0.025085859, -0.071115844, -0.0664060786, -0.0333215073, -0.1071897671, -0.0048042382, -0.0254103784, 0.0306788385, 0.005360357, -0.0422320142, 0.0282859895, 0.0197136048, -0.0326032527, -0.0166647453, 0.0158952959, -0.0705501735, -0.0523580201, -0.0292552765, -0.0148388669, 0.0373467393, 0.0161287263, -0.0278324094, 0.0485907495, -0.0069265538, -0.0111467456, -0.0745299309, -0.0873370394, -0.0042479974, 0.0581081323, 0.0458451062, 0.0057636439, 0.0164017994, -0.0010087269, 0.0168302506, 0.0273885261, -0.021188695, 0.0392528661, -0.0286048483, 0.0249151886, 0.0015992246, -0.0172859021, 0.0033578565, 0.0462274104, -0.0734418035, -0.0194691792, -0.0042603123, -0.03357777, 0.061745014, -0.0174744241, 0.0367828123, 0.0211951602, -0.0087149525, -0.0672033951, 0.0394271724, -0.0371898115, -0.0028854981, 0.0609018989, 0.0488514379, -0.0171032716, -0.0251074079, 0.041004546, -0.0185324512, 0.0214876607, 0.0099252248, -0.0008756748, 0.0017571293, -0.0016884743, -0.003743316, -0.0075350329, -0.028178867, 0.0135057233, 0.0391289145, 0.0039358153, -0.046328295, 0.0562596507, 0.1119722128, -0.0493553393, -0.0033486071, 0.0208220072, 0.029494483, 0.0404809937, 0.0440779217, 0.03544816, 0.0202638991, 0.013392604, 0.0483403131, 0.0519646853, -0.0224101692, -0.0495456159, 0.0747946873, 0.0545673482, 0.0535741113, -0.006707788, 0.0005505683, 0.0077556432, -0.0363697559, 0.0121338405, -0.0665823221, -0.011782499, 0.0088519622, 3.139e-7, -0.0502694175, 0.0208249856, 0.0342626013, -0.0658673942, 0.0433561914, 0.039558921, -0.0349521972, -0.0269785374, -0.0151862744, 0.0673796609, 0.0376159735, -0.016442338, 0.0525201522, 0.0026149144, -0.1041223183, -0.0263886936, -0.004782937, 0.0265355632, -0.016014671, -0.0013917788, -0.0260769054, 0.0745787472, -0.007728131, -0.0251417328, 0.0037508239, -0.0240806695, 0.0545058958, 0.0242391303, 0.0399204493, -0.0312261172, -0.0335992388, 0.0424221568, -0.016507227, 0.0296307001, -0.0578233451, 0.0181753803, -0.0213969667, 0.0508393683, -0.044466652, -0.0288655721, 0.036762882, 0.0485913567, -0.0204233807, 0.0317929238, 0.0516564101, -0.0539454073, 0.0118640605, -0.0190692768, -0.0464281663, -0.0099954195, 0.0078817364, 0.0081243524, 0.0305312071, 0.0170664582, -0.0018565907, 0.0547017157, -0.0304643437, -0.0626762882, 0.0328896269, 0.0267900918, -0.020268796, -0.0218275413, -0.0391002037, -0.0124940835, 0.0292108841, -0.0652379915, 0.0218095314, 2.749872965e-34, 0.0105550736, -0.0571407489, 0.0411260128, 0.0554312952, -0.0128157241, -0.0504334755, -0.0048785284, 0.022570027, 0.0150480429, -0.0703709275, -0.0459950827 ]
http://arxiv.org/abs/2406.17926v1
http://arxiv.org/pdf/2406.17926v1
FASA: a Flexible and Automatic Speech Aligner for Extracting High-quality Aligned Children Speech Data
2024-06-25T20:37:16
2024-06-25T20:37:16
Dancheng Liu, Jinjun Xiong
cs.CL
cs.CL, cs.SD, eess.AS
Automatic Speech Recognition (ASR) for adults' speeches has made significant progress by employing deep neural network (DNN) models recently, but improvement in children's speech is still unsatisfactory due to children's speech's distinct characteristics. DNN models pre-trained on adult data often struggle in generalizing children's speeches with fine tuning because of the lack of high-quality aligned children's speeches. When generating datasets, human annotations are not scalable, and existing forced-alignment tools are not usable as they make impractical assumptions about the quality of the input transcriptions. To address these challenges, we propose a new forced-alignment tool, FASA, as a flexible and automatic speech aligner to extract high-quality aligned children's speech data from many of the existing noisy children's speech data. We demonstrate its usage on the CHILDES dataset and show that FASA can improve data quality by 13.6$\times$ over human annotations.
FASA: a Flexible and Automatic Speech Aligner for Extracting High-quality Aligned Children Speech Data: Automatic Speech Recognition (ASR) for adults' speeches has made significant progress by employing deep neural network (DNN) models recently, but improvement in children's speech is still unsatisfactory due to children's speech's distinct characteristics. DNN models pre-trained on adult data often struggle in generalizing children's speeches with fine tuning because of the lack of high-quality aligned children's speeches. When generating datasets, human annotations are not scalable, and existing forced-alignment tools are not usable as they make impractical assumptions about the quality of the input transcriptions. To address these challenges, we propose a new forced-alignment tool, FASA, as a flexible and automatic speech aligner to extract high-quality aligned children's speech data from many of the existing noisy children's speech data. We demonstrate its usage on the CHILDES dataset and show that FASA can improve data quality by 13.6$\times$ over human annotations.
[ -0.0487929955, 0.0918983147, 0.0295730364, 0.0281372461, 0.0056420099, -0.0091560967, 0.0321530066, -0.0465188809, -0.0684890971, 0.0548242219, -0.0614322238, -0.0502848104, 0.0762623399, -0.0226770844, -0.0071440665, -0.0695271268, 0.0268419851, -0.0233333986, 0.0245112795, 0.0071112346, 0.019421, 0.0185328666, -0.0335989147, 0.0255329497, -0.0526862107, -0.0365913622, -0.0141400946, 0.0328656919, 0.0475026965, 0.0156084774, 0.0125666391, 0.0007459568, -0.0106707783, 0.0708453804, 0.0000016286, -0.0134808663, -0.0022606223, 0.0216676369, -0.0753996223, 0.0203163754, -0.0003468038, 0.0096408119, -0.0472389273, 0.0328873917, -0.0063861185, -0.0160987489, 0.0773068741, 0.0073336242, 0.0409200825, 0.0537635274, -0.0059780711, -0.0134779895, 0.0257218704, 0.0002007826, -0.0464719534, -0.0145981992, -0.0026594154, 0.0157108456, -0.0104244286, -0.0675824285, 0.0317806005, 0.038054511, -0.000897716, -0.0391282439, 0.0206985977, -0.0418303274, -0.06502112, -0.0524782091, -0.016712714, 0.0911961645, 0.0008780507, 0.0042435853, 0.0180390459, -0.0325945653, -0.0232924782, -0.0185860936, -0.0262028407, 0.0254475679, 0.013703973, 0.022007063, -0.011796074, 0.0034393058, 0.0389485806, -0.0367854536, 0.0156855173, 0.0263539683, -0.0006473006, -0.0349667557, 0.0049579949, 0.0064946478, 0.0200818013, -0.0175525919, 0.0288365502, -0.0066600554, -0.0461664908, -0.0074144891, -0.1069415659, -0.0537047908, -0.0300096199, 0.0051963646, -0.0104178414, 0.0085121393, -0.0521942861, 0.0220327601, 0.0518204495, 0.0553387441, -0.0423761234, 0.0273945369, -0.012177282, -0.0278544016, 0.0397712663, 0.0275173429, 0.007482904, -0.006346934, -0.0029124569, -0.0106917731, -0.0622860566, 0.0177018046, -0.0402535796, 0.0543564223, 0.0462645553, 0.0252925139, -0.0468891338, -0.0367945544, 0.0243607704, 0.0708948523, 0.0046759355, 0.0357860625, 0.0395137072, -0.0479194559, -0.0085716862, 0.0374264047, 0.0765775666, 0.0118962191, 0.000513471, 0.0208940711, 0.0114025334, 0.0413919427, -0.017945081, 0.041034054, -0.0294033699, -0.046232529, 0.0092923269, -0.0479848124, 0.0346056521, -0.0243263934, -0.0283672344, -0.0671969429, -0.0163514167, -0.0383018926, 0.0230092872, 0.014815392, -0.0545026623, 0.0306533445, 0.0081215883, -0.0354879647, 0.013062302, 0.0305761471, 0.0773182064, -0.0450853147, 0.0147161316, -0.0326478072, -0.0402409621, 0.0970781296, 0.0014545521, -0.0077173379, -0.0071248286, -0.0211930275, -0.0198246054, 0.0616700314, 0.0634513348, 0.0244252738, 0.0169743616, -0.0024813511, 0.0721408576, 0.0315275341, 0.0182061028, 0.0310010742, -0.0097754067, 0.0282653458, 0.0355955474, 0.0078084138, -0.0315800309, 0.0272964127, 0.0379243866, 0.0034374839, 0.0384161249, 0.0524222404, -0.0066091483, -0.0259199031, -0.0309677403, -0.0302362479, -0.0184039343, 0.0249455143, -0.0018545588, -0.0269479081, -0.0362762958, -0.0101445327, -0.0014658211, -0.0606496222, 0.0203738343, -0.0233188868, 0.015758669, 0.040216621, -0.0163277928, -0.0451749153, -0.074905023, -0.0164265148, -0.0436575152, 0.0425680988, -0.0253348257, 0.0120665729, 0.0575389452, -0.0740178972, -0.0361916758, 0.0608093478, -0.0069567231, 0.0004448902, -0.0153196836, -0.0256793946, 0.0028130221, 0.0162865855, -0.0005687426, 0.0162277669, -0.0259286407, 0.0078159533, 0.0203992128, 0.0674765036, 0.0572946556, 0.0580189042, -0.0091007371, -0.0275941212, 0.0457259454, 0.0024637128, 0.0656665191, 0.0149416598, -0.0370256566, 0.0244307537, 0.0598287098, -0.0547019765, -0.0553670973, 0.0744194314, -0.0034980502, -0.0736456364, 0.0003138329, -0.0310282949, 0.0026748204, -0.0086451424, 0.0506971292, 0.0035802221, -0.0127076302, 0.0117132179, -0.0025198769, 0.0286221821, -0.0318288617, 0.015914198, -0.0280641038, -0.0246777069, -0.064614594, -0.0062695355, -0.0082343305, -0.0319810398, 0.0700713173, -0.0000946973, -0.0314763226, -0.0526902676, 0.0134845376, -0.0510906242, 0.0518999845, -0.0206016935, 0.0139817446, 0.0271270052, 0.0197623875, 0.0052923788, -0.0050898218, -0.0589406118, 0.0064897267, 0.0011768388, -0.0298850387, 0.0188228395, -0.036165636, -0.0239994489, -0.070258528, 0.012734591, -0.0071974713, -0.0314303674, -0.0195976254, -0.0299582519, 0.0059263133, -0.006198131, -0.0304337163, 0.0279528406, 0.0111494251, -0.026267305, 0.0019835255, 0.0410501175, 0.0105599789, 0.0715928674, -0.028880883, -0.0388599373, -0.0172633547, -0.0020938767, -0.0510324948, 0.0111904833, 0.026231356, 0.0547064058, -0.0022407649, 0.0282212608, -0.019228952, 0.0266375765, 0.0218362212, 0.0387928449, -0.0371180698, 0.0207423829, 0.0215867702, 0.0378597677, 0.0179875735, -0.0518033728, -0.0480704904, 0.005489965, -0.038505055, -0.0460101031, -0.0039267503, -0.0547410548, -0.0549601093, -0.012936973, -0.0334869772, -0.0059250649, -0.0147060119, -0.0412168391, -0.00189861, -0.0238647051, -0.0305886958, -0.0677745715, 0.0018907845, 0.0365188643, -0.0615837649, 0.0755926594, -0.0121928379, 0.0122474311, -0.0309971999, -0.0433897562, -0.0536561571, -0.0100497324, -0.0367397889, 0.0109372744, -0.010232388, -0.0866370872, -0.0406210497, -0.042886544, 0.0192526318, 0.0546954535, -0.0101364339, 0.0613182224, -0.0125537496, 0.0558902621, 0.0492554195, 0.0281409435, -0.0024154156, -0.0165441707, 0.0063012452, 0.0132694272, 0.0280751102, 0.0301394612, -0.0613515377, -0.0471318699, 0.0241284259, 0.0388668366, -0.0122860996, -0.0061649675, 0.067478545, -0.0382949002, 0.0565799028, 0.0414353944, -0.0137427589, 0.0111997155, -0.0285426192, 0.0125521729, 0.0484261811, -0.0097592026, -0.0633716732, -0.0476058349, -0.0266478974, 0.0772471428, 0.0366645306, 0.0528759435, 0.024132438, -0.0083704852, -0.0437378772, 0.0155064603, 0.0368614085, -0.0407241248, -0.0559704751, -0.0165521521, 0.0020896769, -0.0392567925, 0.1111878082, 0.0225190967, -0.003498283, -0.0304766595, 0.0362995863, -0.0398846418, -0.0547844283, -0.0278094709, -0.0622802041, 0.1344997883, 0.0459377654, 0.0127568562, -0.0042637447, -0.0394485779, 0.0240831561, 0.0061182082, -0.0548374802, 0.0385362804, -0.0050330344, -0.0068951808, -0.0477083474, -0.0102325091, 0.0607875884, 0.0048912559, 0.0353144109, 0.0104977377, 0.0359780416, 0.0083080865, 0.0423336625, 0.0012948384, 0.0272631627, 0.0092725363, 0.0638804436, 0.0114042154, -0.0218222197, 0.0295572281, 0.0260391515, 0.0307958834, -0.0291252229, 0.0011095698, 0.0643772557, -0.0203757193, -0.036357142, -0.0693763793, 0.0203389451, -0.0020235076, -0.0129594207, 0.0247548167, -0.0623997673, 0.0175004117, 0.0003550062, -0.0226113126, 0.0587133765, -0.00464838, 0.0022016449, -0.062915124, -0.0146921985, -0.015091571, -0.0288534928, 0.0053315251, 0.0300249662, -0.0865540579, -0.0641407073, 0.0347619317, -0.0330237001, 0.018524671, 0.0657574087, -0.0040639164, -0.0125854872, 0.0267804414, 0.0234967414, -0.0101415291, -0.050395973, 0.046921324, -0.0254234523, 0.0117609669, 0.0280803107, -0.0643288344, 0.0039738477, -0.0389463753, 0.0043616355, -0.0346236303, -0.0683351904, 0.0032984787, 0.0276339259, 0.0689690858, 0.0088640777, -0.0178478733, -0.0575524569, 0.0191706792, -0.022725923, 0.0211916789, -0.0293283872, -0.0287448112, 0.0164177027, 0.0307721868, 0.0097261025, 0.0253140181, -0.0273135956, -0.0223417748, -0.0245422795, 0.056274958, -0.0191202331, 0.0142464666, -0.0035251521, 0.0298459847, -0.0050562979, 0.0461020991, 0.0303265415, 0.0520353653, -0.0042949044, -0.0200415254, -0.0011941121, 0.0284162499, -0.0207067132, 0.0175675601, 0.0027831104, -0.0079221306, -0.0303620026, -0.0088011222, 0.0311764944, -0.0123483166, 0.0579076, 0.0208241791, 0.0178606361, -0.020882491, 0.0124448026, -0.0186201837, 0.0243192371, 0.0017589303, -0.0180534329, -0.0405629501, 0.0276713334, 0.0108778886, -0.0247054566, -0.0102536296, -0.046271693, 0.05908297, -0.0370302312, -0.0194983836, 0.0055178045, -0.0378117785, -0.0478323326, -0.0609167404, 0.0347940959, -0.0345899239, -0.0191283543, 0.0315312482, -0.0135999219, -0.0232024342, 0.0284052137, -0.0180334263, -0.0908877403, 0.0051898067, 0.0433273315, -0.061490383, -0.0040516127, -0.0494479872, -0.0039113034, -6.112127977e-33, 0.0090801604, 0.0151291313, -0.0016604307, 0.0030935479, -0.0499848835, 0.0377477519, -0.0091630658, 0.0259897001, -0.0226941146, 0.0035682283, -0.0039652614, -0.0060295081, 0.0404066779, 0.0294022653, -0.0254050326, -0.0406213403, 0.0792637989, -0.0281649996, -0.0192734152, -0.033202894, -0.0019229685, -0.0088431127, 0.0106260199, -0.058561936, -0.0114750946, -0.0137356473, -0.0284942836, 0.0081457421, 0.0288923569, 0.0705896392, 0.0404422022, 0.010377489, -0.0152441217, 0.0190894641, 0.0115344534, 0.0463851467, -0.0937939584, 0.0276039857, 0.0126845399, 0.0482806005, -0.0352531448, -0.0228014197, 0.0766762719, 0.0005938698, -0.0087662758, -0.0436848551, 0.0520917252, -0.0263282973, 0.005041956, -0.0319290757, -0.0138389925, 0.0169943795, -0.0387681872, -0.0214494113, 0.0661778599, 0.0601862259, -0.0117999269, -0.0158694666, -0.0729724988, 0.0186868794, 0.0496985875, 0.0640166402, 0.0548903085, 0.0334571898, -0.013151085, 0.0421416648, 0.0575327873, -0.0547024794, 0.0119951991, -0.0202694479, 0.0063261716, 0.0202033576, 0.0149997473, 0.0217538029, 0.0375265814, 0.0096946117, -0.0569229983, 0.0091311447, 0.0087995222, 0.0049887616, -0.0354129896, 0.0200913362, -0.0440165587, 0.036040321, -0.0141545758, -0.0203378797, -0.017265439, 0.0056081866, 0.0077662277, -0.0209155697, -0.0377944857, 0.0009651934, 0.0263446271, -0.0079185152, 0.0461913832, 0.0075741853, 0.0284863189, -0.030166911, -0.0190438684, 0.0010647837, -0.0370116606, -0.0604505613, 0.0324191637, 0.0538707227, 0.043337509, 0.0045473147, 0.0398891121, 0.0413500555, -0.0667970181, 0.0100075835, 0.0062947432, 0.0489098839, -0.0438122563, -0.0006910271, 0.066424489, -0.0328910798, 0.015058618, 0.0554322079, 0.0053936443, 0.059365958, 0.0025939643, -0.0069815139, 0.0236090198, -0.0362329632, -0.0393280499, 0.0255039837, 0.0089976341, -0.0087693203, 0.0224591717, -0.0261957347, 0.0160729401, -0.0356843136, 2.289e-7, -0.0462581776, 0.0365614481, 0.0689742267, -0.0265101604, 0.0425062329, 0.0146664092, -0.0055599967, -0.0422785506, -0.0027731371, 0.0162262991, 0.0499926209, -0.0429197922, 0.1033926606, 0.0352809206, -0.0665413216, -0.0265944041, -0.0856194124, 0.0011569503, 0.008141811, 0.0180255119, -0.0010864375, 0.0845322385, 0.0084249265, -0.0572121516, 0.0248831715, 0.0172348227, -0.0459604412, -0.007633172, 0.0093147438, -0.0206720792, 0.001984603, 0.0163663737, -0.0086556301, -0.0092884526, -0.0268094763, 0.0683218464, -0.003134778, 0.0578056797, -0.0338254459, -0.0406723469, -0.0205374677, 0.0143784927, -0.0204018466, 0.0020182584, 0.0460871831, 0.0036494697, -0.0011958189, -0.0194499623, -0.0720232502, 0.0178810004, -0.0013651833, 0.0083610741, 0.0167542305, 0.0250033513, 0.0281655844, -0.0287583321, -0.0382889584, -0.0020123064, -0.0321474485, -0.0038907116, -0.0429885574, -0.0128683615, 0.0036317168, -0.0192517117, 0.0092699192, -0.0006863973, 0.0404278524, 1.356145677e-34, 0.0054440284, -0.046290163, 0.0788374543, -0.0121735232, -0.0191319976, -0.0439183712, 0.0232585054, 0.0148948561, 0.007510649, 0.0147560872, -0.0215520654 ]
http://arxiv.org/abs/2406.17923v1
http://arxiv.org/pdf/2406.17923v1
PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning
2024-06-25T20:11:37
2024-06-25T20:11:37
Shiva Kumar Pentyala, Zhichao Wang, Bin Bi, Kiran Ramnath, Xiang-Bo Mao, Regunathan Radhakrishnan, Sitaram Asur, Na, Cheng
cs.CL
cs.CL
Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pipeline leads to alignment tax that degrades the LLM performance. This paper introduces PAFT, a new PArallel training paradigm for effective LLM Fine-Tuning, which independently performs SFT and preference alignment (e.g., DPO and ORPO, etc.) with the same pre-trained model on respective datasets. The model produced by SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications. This work reveals important findings that preference alignment like DPO naturally results in a sparse model while SFT leads to a natural dense model which needs to be sparsified for effective model merging. This paper introduces an effective interference resolution which reduces the redundancy by sparsifying the delta parameters. The LLM resulted from the new training paradigm achieved Rank #1 on the HuggingFace Open LLM Leaderboard. Comprehensive evaluation shows the effectiveness of the parallel training paradigm.
PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning: Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pipeline leads to alignment tax that degrades the LLM performance. This paper introduces PAFT, a new PArallel training paradigm for effective LLM Fine-Tuning, which independently performs SFT and preference alignment (e.g., DPO and ORPO, etc.) with the same pre-trained model on respective datasets. The model produced by SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications. This work reveals important findings that preference alignment like DPO naturally results in a sparse model while SFT leads to a natural dense model which needs to be sparsified for effective model merging. This paper introduces an effective interference resolution which reduces the redundancy by sparsifying the delta parameters. The LLM resulted from the new training paradigm achieved Rank #1 on the HuggingFace Open LLM Leaderboard. Comprehensive evaluation shows the effectiveness of the parallel training paradigm.
[ 0.0488101542, 0.0250965487, 0.0106262639, 0.022329133, -0.0385594331, 0.0245424043, -0.0273460411, -0.014608386, -0.0161372852, -0.0584242493, -0.056008555, -0.0193923749, 0.0420058109, -0.0077188294, 0.0111773666, -0.0409964211, 0.0575883873, -0.0023462169, 0.0007667802, -0.0002799721, 0.0227778107, 0.0332440697, 0.0024456477, -0.0037975598, 0.0378585197, -0.013189204, 0.0073822546, 0.0245597027, 0.0353117883, 0.0096805627, 0.0267194025, 0.0653642341, 0.0099993823, 0.0475839861, 0.0000020418, -0.0627609417, -0.034015391, 0.0200814661, -0.005214815, 0.0238223411, 0.0406183638, 0.0189314857, -0.0049730754, 0.0319929942, -0.0266948324, 0.0010140586, 0.0415021665, 0.0699818805, 0.0495666228, 0.0860074759, -0.0074410476, -0.0557974763, 0.0364532396, -0.0274825152, -0.0841122642, -0.0249321274, 0.0372222662, -0.0255678855, -0.0136216255, -0.0073626158, 0.0064674658, 0.0077647921, -0.0406152159, 0.056325227, 0.023923425, 0.0039142966, -0.0393882208, -0.0412660576, -0.0407707356, 0.0603329986, -0.0970495641, -0.0072932858, 0.0155858705, -0.0330073945, -0.0247206148, -0.0081022177, -0.0250701103, -0.0188694168, 0.0012435097, 0.0668160841, 0.051679261, -0.0468604751, 0.026035469, 0.0280735586, -0.0539322272, 0.0305568203, 0.0307244211, -0.0394837894, 0.0131658837, -0.0364823192, -0.0782086328, -0.0206541438, 0.0031298061, -0.0208491143, -0.0545137674, 0.0208327696, -0.0569663122, -0.0237397999, 0.0532024577, -0.0915495604, 0.0637378693, 0.0490325354, 0.0058354228, 0.0271429364, -0.0267784782, 0.0414633825, -0.0148495417, 0.0286271125, -0.0601825565, 0.0166949295, -0.0396837555, 0.0048268312, -0.0029387095, 0.0441493765, -0.0705359429, -0.0403656624, -0.0447105728, 0.0165897515, 0.0166900717, -0.00452821, 0.0610538386, -0.0178247299, -0.0143927587, -0.0160040241, -0.0246610139, 0.0710852742, 0.012687658, 0.0252942964, -0.0524130017, -0.035407234, -0.001414791, 0.0068647182, -0.0065217912, -0.0286721699, 0.0202661399, 0.034642186, 0.0025287853, -0.026994722, -0.0511443205, 0.0205528643, -0.0084929559, -0.0763862506, -0.028342193, -0.0488007888, 0.0436927825, -0.0194854476, 0.0294383056, -0.0280498322, 0.0188948456, 0.0007980803, -0.0491028763, 0.072940737, -0.0420760848, 0.0171116535, 0.0152374292, -0.0202159323, 0.0446619503, 0.0956159905, 0.0209514387, 0.0089380695, 0.0343264602, -0.0230415296, -0.0639968738, 0.0374096446, -0.0178612955, 0.005939519, -0.0447936282, 0.0018840842, 0.0387408361, 0.0085763475, 0.0319753699, 0.0538571328, -0.0105616786, -0.0053708847, 0.0458657779, 0.0154143125, 0.0643359348, -0.0026363782, -0.0165741052, 0.0286353882, -0.02054804, 0.0433749668, -0.0272935815, -0.0183793884, -0.0563575774, 0.0377307497, 0.0587081909, 0.0848836526, -0.0055827005, -0.0177912749, -0.0341919251, -0.0005926106, -0.0456833057, 0.0018836788, 0.0456352159, -0.0190230515, -0.0762778893, 0.0360028818, 0.0473018214, -0.0057357959, -0.0332471319, -0.0251756459, -0.0235039238, 0.078303583, 0.023268275, -0.0283784438, -0.0554844327, -0.0071203504, -0.0403521284, 0.0346825607, 0.0431613699, 0.0228630174, 0.0171411633, -0.0649140924, -0.0576779023, 0.0154250544, 0.0025875005, 0.0026677137, 0.0434418134, -0.0132520888, 0.014154885, 0.0224207062, -0.0536522493, 0.0486172885, -0.0025385779, -0.0265628844, 0.0223072432, 0.0052949805, 0.0676868483, 0.0234932564, -0.0035800531, -0.0077567776, 0.0616710186, -0.0044843443, 0.0022131496, -0.0293146223, -0.0144174295, -0.0103345728, 0.0724674463, 0.0304153059, -0.009626084, 0.0509035848, -0.0592728108, -0.0545187034, -0.0476924889, -0.0053419289, 0.0453339666, -0.0331872851, 0.0659221187, -0.0072900183, 0.0215302054, -0.077237919, 0.0558093227, 0.0000105432, 0.0343971141, -0.0057216147, 0.0143984361, 0.0184800643, -0.0294567589, -0.0354020596, 0.0214028899, -0.0069785458, 0.0157784689, -0.0029096929, -0.0011647551, -0.0551204197, -0.0468136668, -0.0713018253, 0.0673218593, -0.0107007883, 0.0282436926, 0.035124518, 0.0300814901, -0.0075064464, -0.00520512, 0.0130747464, 0.0105578033, 0.0127051948, 0.0088191181, 0.0212743748, 0.0059948042, -0.0467958562, -0.0407432877, -0.0088062445, 0.0132304672, 0.0058153137, -0.025068054, -0.028765399, -0.061750751, 0.0467313677, -0.0125915296, 0.0516883321, 0.0063254819, -0.0173378233, -0.0179293696, -0.014536567, 0.0303593036, 0.0568798669, 0.019075077, -0.0147323338, 0.0365444757, -0.0019057781, -0.0320100859, 0.0066701351, -0.0424016714, -0.0061433222, -0.0541774109, 0.0054225479, 0.0023417075, 0.0644244775, -0.0108924313, -0.0022544989, -0.0429321863, -0.0335827135, 0.0081676077, -0.0011193111, 0.0208584629, 0.0028700125, -0.0255864244, 0.0062160096, -0.0314056315, -0.0688884258, 0.0553615279, -0.0440456159, 0.017050406, -0.0744398162, -0.0214820225, -0.0569027551, 0.0145689156, -0.0135768764, 0.0055369139, -0.0349292569, -0.026644852, -0.034123078, -0.0231826808, 0.0055510844, -0.0207252484, 0.0357349217, -0.0312334541, 0.0274127293, -0.0475865602, -0.0247417632, -0.0319772772, -0.0690810308, -0.0217518322, 0.0138409352, 0.0338538177, -0.0321257226, -0.0376533531, 0.0311578009, -0.0241528377, 0.0226190481, -0.01228402, 0.0163497608, -0.0463131256, 0.0140904216, 0.0670661256, 0.0138175329, 0.0134707615, 0.0113060232, 0.0006708018, 0.0119573502, 0.0376989059, -0.0268367399, -0.0301701985, -0.0104809841, 0.0488385595, -0.0750218481, -0.0234878957, -0.0121864509, 0.0064378562, -0.053328868, 0.0444827974, 0.1204554364, 0.0359320529, -0.0058423979, -0.0076826909, 0.0198567659, 0.062066026, 0.032701198, 0.0228652339, -0.0538119413, -0.036569912, -0.0083653657, 0.0120523954, -0.0016709839, 0.0366433151, -0.0359760858, -0.121402815, -0.0028881209, 0.0439311266, -0.0200209431, -0.0072063394, -0.0595495477, 0.0057896846, -0.0327958427, 0.0379303619, 0.0098020853, -0.0339166783, -0.0084982775, 0.0101083545, 0.0237932857, -0.0517930761, -0.0168811325, -0.0307166353, 0.0537409969, 0.0433699116, 0.0067839674, 0.0189538803, -0.0327290781, 0.0634151176, -0.0179552417, 0.0121510094, -0.0653660074, -0.0015904298, -0.0391601957, 0.0046954723, -0.0191750284, 0.0075850729, -0.0235282071, -0.01202399, -0.0198424477, 0.0696180239, -0.0102304043, 0.0198076069, 0.0155168669, -0.0032176855, -0.03392848, 0.0414697714, -0.0361599624, -0.0252739936, -0.0275094695, 0.0315205194, -0.0309760142, -0.0123867188, 0.0184648838, 0.0219158679, -0.0585426837, -0.0328964405, -0.0209287573, 0.0315967314, 0.0376749896, 0.0171016902, 0.0304051358, 0.0615963042, -0.0139267175, -0.0047136848, -0.0344590172, 0.0371032991, 0.0001513154, 0.0035125811, -0.0734866485, 0.0113187432, 0.0597171485, 0.0029266474, -0.0071966993, 0.0118701207, -0.065826878, -0.0928675011, 0.002944947, 0.0293257814, 0.0165851377, 0.0576520264, 0.0074257343, -0.0543720759, -0.0088232178, 0.0607449189, -0.0390359908, 0.0200483259, 0.0051503438, -0.0340832509, -0.0062217857, 0.0086517232, -0.0091706729, -0.0099465996, 0.0086318152, 0.0195135847, -0.01062652, -0.0935209319, 0.0125748087, 0.0328311808, 0.0639614761, -0.0237634387, -0.0088596642, -0.0513111502, 0.0551321059, 0.0245444514, -0.0046095331, -0.0168445893, -0.0185883623, 0.0223928224, -0.0107175102, -0.0040594698, 0.0232927296, -0.0178776868, 0.0014588536, -0.0165088847, 0.0161977988, -0.0212753229, -0.0412360206, 0.0155544085, 0.018420618, -0.0147246867, 0.051092688, 0.0339293629, 0.0167311542, -0.0343485735, 0.0275545418, 0.0181349069, 0.0496816188, -0.051662419, 0.0346366465, -0.0417302251, -0.0605214685, -0.0399362221, -0.017549675, 0.078560032, -0.025175713, 0.0828139335, 0.0110498955, -0.0217183698, -0.0163510591, 0.0157611351, -0.0262373369, 0.0363842882, 0.0249756146, 0.0106149381, -0.0444401689, 0.022395296, 0.0650156364, 0.0008202499, 0.0335798822, -0.0024419392, -0.0096623786, -0.0461587608, -0.0443980955, 0.0214212164, 0.0647088885, -0.0449380726, -0.0104492689, -0.003517793, -0.026227938, -0.0000931183, -0.0224476326, -0.0071172002, -0.0205516331, 0.0358747877, 0.0179465041, -0.0248461962, -0.008978501, -0.0220948998, -0.040948946, 0.0241845697, -0.0218235049, 0.0060837702, -6.887474133e-33, -0.0468988493, 0.0095565291, 0.0000814059, -0.0106574316, -0.063291952, -0.0234631095, 0.0372545011, -0.0220385585, -0.0193136372, -0.0367775001, -0.0249363817, -0.0009905621, 0.0279280785, 0.0067372224, -0.0074182996, 0.0267379116, 0.0748299062, -0.0011878876, 0.0192656126, -0.0276841242, 0.0233726762, 0.052300781, 0.0358649567, -0.0446119979, -0.0280793849, 0.0215383824, -0.0204021651, 0.0170752574, -0.0399285406, 0.0375889204, -0.0129332701, 0.0082480218, 0.0236038454, -0.0147961667, -0.0232447032, 0.0174268298, -0.060452044, 0.0119014801, 0.0778970718, 0.0258640964, 0.001632079, -0.0096185245, 0.0561232641, 0.0072666751, -0.0090595083, -0.0295577329, 0.0256718379, -0.0484227017, -0.0318165757, 0.0332858413, -0.0562576465, 0.0433763079, 0.0117827402, 0.008726852, 0.064250946, -0.0303006824, -0.0039755115, 0.0418865606, -0.0264484584, 0.0095359664, 0.0373367891, 0.076866284, 0.0584282316, 0.0492265411, 0.0085741775, 0.0395636223, -0.0060921083, -0.0336436182, 0.0496573672, -0.0302628502, 0.0183898956, 0.080724977, 0.0085809473, 0.0264024269, 0.0444977731, -0.0347943045, -0.0647360533, 0.0222476609, 0.0043032612, 0.0096368585, 0.0219686292, 0.0576547049, 0.0045451615, -0.0185968168, -0.0440505631, 0.0019430631, -0.0452896617, -0.0027612248, 0.0157697126, 0.0349263623, -0.0062910141, 0.0265521836, 0.006934287, -0.0618059412, 0.0221508909, -0.0131013431, -0.0174533557, 0.0136339292, -0.0165476706, 0.0055787526, -0.0862689391, -0.0816387832, -0.0058751334, -0.0026576584, -0.0097606303, -0.0238104109, 0.0280501917, -0.0163358133, -0.0801641345, 0.0338140391, -0.003885437, -0.0012930601, 0.0136831775, 0.0711199045, 0.0434631109, -0.0216123499, -0.0014360319, 0.0535096936, 0.0232626181, -0.0006833444, -0.0236329045, -0.0117195491, 0.0029721241, 0.0280888062, 0.0086151157, 0.0351312831, 0.0098935645, 0.0062154932, 0.0111309718, -0.0245700292, -0.0304995291, 0.0283804182, 2.837e-7, 0.0280751698, 0.0468679965, 0.0227152798, -0.0109961787, -0.0087144151, 0.0183964763, -0.0070536397, 0.013769675, 0.0107788807, 0.0248669218, 0.0235539041, 0.0031925812, 0.0334992036, -0.019371653, -0.0890473202, -0.0118696159, -0.0382314399, 0.0388551615, -0.0625109002, 0.0033825596, 0.0350501314, 0.1087398753, 0.0139166173, -0.0146325743, -0.005144591, 0.0249710232, 0.0356769934, -0.0282025877, -0.0070031635, 0.0570965596, -0.0196586158, 0.0032857107, -0.0361485891, 0.026964182, -0.0378036425, 0.0539337136, -0.0290707983, 0.1036699116, -0.0267581977, -0.02087388, 0.0385037027, 0.0149213793, -0.0230368469, -0.0644438639, 0.075370416, -0.0191806909, -0.0490577035, -0.0600327067, -0.0136487642, -0.0192297827, 0.046433188, -0.0185429715, 0.0064451192, 0.0295321196, 0.029898379, -0.0071128299, -0.0726207346, -0.0688439682, -0.020518627, 0.0243469123, -0.0348394699, -0.0332419313, -0.011368067, 0.0180626977, 0.0402133875, -0.0353845693, -0.0029623264, 3.095333765e-34, 0.0322080553, -0.0107728792, 0.0299658086, 0.0292716846, -0.0097444961, -0.0213100314, -0.0201153643, 0.0550421663, -0.017724121, -0.0942324772, 0.0016246967 ]
http://arxiv.org/abs/2406.17911v2
http://arxiv.org/pdf/2406.17911v2
X-ray Made Simple: Radiology Report Generation and Evaluation with Layman's Terms
2024-06-25T19:52:01
2024-06-30T21:53:56
Kun Zhao, Chenghao Xiao, Chen Tang, Bohao Yang, Kai Ye, Noura Al Moubayed, Liang Zhan, Chenghua Lin
cs.CL
cs.CL
Radiology Report Generation (RRG) has achieved significant progress with the advancements of multimodal generative models. However, the evaluation in the domain suffers from a lack of fair and robust metrics. We reveal that, high performance on RRG with existing lexical-based metrics (e.g. BLEU) might be more of a mirage - a model can get a high BLEU only by learning the template of reports. This has become an urgent problem for RRG due to the highly patternized nature of these reports. In this work, we un-intuitively approach this problem by proposing the Layman's RRG framework, a layman's terms-based dataset, evaluation and training framework that systematically improves RRG with day-to-day language. We first contribute the translated Layman's terms dataset. Building upon the dataset, we then propose a semantics-based evaluation method, which is proved to mitigate the inflated numbers of BLEU and provides fairer evaluation. Last, we show that training on the layman's terms dataset encourages models to focus on the semantics of the reports, as opposed to overfitting to learning the report templates. We reveal a promising scaling law between the number of training examples and semantics gain provided by our dataset, compared to the inverse pattern brought by the original formats. Our code is available at \url{https://github.com/hegehongcha/LaymanRRG}.
X-ray Made Simple: Radiology Report Generation and Evaluation with Layman's Terms: Radiology Report Generation (RRG) has achieved significant progress with the advancements of multimodal generative models. However, the evaluation in the domain suffers from a lack of fair and robust metrics. We reveal that, high performance on RRG with existing lexical-based metrics (e.g. BLEU) might be more of a mirage - a model can get a high BLEU only by learning the template of reports. This has become an urgent problem for RRG due to the highly patternized nature of these reports. In this work, we un-intuitively approach this problem by proposing the Layman's RRG framework, a layman's terms-based dataset, evaluation and training framework that systematically improves RRG with day-to-day language. We first contribute the translated Layman's terms dataset. Building upon the dataset, we then propose a semantics-based evaluation method, which is proved to mitigate the inflated numbers of BLEU and provides fairer evaluation. Last, we show that training on the layman's terms dataset encourages models to focus on the semantics of the reports, as opposed to overfitting to learning the report templates. We reveal a promising scaling law between the number of training examples and semantics gain provided by our dataset, compared to the inverse pattern brought by the original formats. Our code is available at \url{https://github.com/hegehongcha/LaymanRRG}.
[ 0.0326596126, 0.0697698668, 0.0403918587, -0.0080308467, -0.0059831962, 0.0165301654, 0.0300583728, 0.0449650027, -0.0607360639, -0.0212373547, -0.0389189832, -0.0534482822, 0.0179488026, 0.0552174225, 0.0476547852, -0.1057389826, 0.0117679862, 0.0049078222, -0.0065812096, -0.0065234238, 0.0157306548, 0.0250960179, -0.009090892, 0.0328540951, -0.0711000189, -0.0062678647, 0.024743462, 0.0209767651, 0.0234123468, -0.0376103222, -0.036531277, 0.0519200563, 0.0217153747, 0.0116172172, 0.0000019898, -0.020379195, -0.0051795784, 0.0353621021, 0.0420645252, -0.0050501991, 0.0891580805, -0.0814995244, -0.0122551294, -0.0100933453, -0.0154539794, -0.0507804677, 0.0464903601, 0.0475498028, 0.0576882176, 0.0545061566, -0.0279189087, -0.0162619855, -0.0092548691, -0.0505453385, 0.0482215211, -0.024717221, -0.0102070281, 0.0239418373, -0.0161303058, 0.017873941, -0.0441663787, 0.0414984226, 0.0109527558, 0.0103394873, 0.0237873886, 0.0713847652, 0.0239762627, -0.019491462, 0.0054188739, 0.0751729161, -0.002859765, 0.018308308, -0.0086202091, -0.0142463818, -0.0061240634, 0.0041199396, -0.0644726157, 0.0518641621, -0.0103330752, -0.0069099688, -0.0278090406, 0.0081090452, 0.0076817414, 0.0268214662, 0.0235243663, -0.042268049, 0.0182069913, 0.0161173902, -0.0077872495, 0.0102848615, -0.0044799605, -0.0510888547, 0.0001519622, 0.0265023578, -0.0384015329, 0.0183927771, -0.0583513677, 0.0110215573, 0.066848807, -0.0139148664, 0.013336611, -0.0110491822, -0.0576294623, 0.045455303, -0.0099301869, 0.0013440162, -0.0096936384, 0.0500112362, -0.0319028385, 0.0058272765, -0.0269341152, -0.0118450141, -0.0364740752, -0.000221158, 0.0068817837, -0.0191086512, -0.0876914859, -0.0153953964, 0.0028386188, 0.0140217645, 0.0336702801, 0.0779837444, -0.1156737879, 0.0200706981, -0.0260861814, 0.0353918411, -0.0272715818, -0.005663699, 0.041801963, -0.0245076455, 0.0106685869, 0.0378971137, -0.0071052518, 0.0035085483, -0.0183406249, 0.0400986932, -0.0258247815, -0.0025410522, 0.0341387466, -0.0790545195, -0.0289109778, -0.0439592637, 0.0446162298, -0.0875725225, 0.0340748355, -0.0042150379, 0.0149545893, -0.0076293573, 0.0038470011, 0.006606156, -0.0124818543, 0.0648498684, -0.0052479897, -0.0133484825, 0.0335883386, 0.0357722491, 0.0237324666, 0.0295647737, 0.0205423739, 0.0330371112, 0.0125644179, 0.001970636, 0.0097710202, 0.0098279342, 0.0102910511, -0.0080596013, -0.0325505398, -0.0264098551, -0.0197879095, 0.0899359062, 0.063288115, 0.0535698906, -0.0457929485, -0.0204698667, 0.0201333947, 0.0111605795, 0.0161749702, 0.0192264635, 0.0653691441, -0.0253441501, -0.0305964854, 0.0191927589, 0.0019767811, 0.0530083515, 0.0169351175, 0.0185522251, 0.0044847783, 0.0110087255, -0.0532657541, -0.0244063344, 0.0017195046, 0.0032933869, -0.0063768821, -0.0302047953, 0.0422699489, 0.0314026102, -0.0506465174, 0.0442138128, 0.0536453538, -0.0459060967, 0.0052323495, -0.0649146885, -0.0109608574, 0.0358411409, 0.0997849107, 0.0158700515, -0.0315391161, 0.0127125392, 0.0329892114, 0.0696505383, 0.0256773885, -0.0150722964, 0.0281909779, -0.0527207814, -0.0436354727, -0.0039319531, 0.0109173516, 0.030123122, -0.0028435972, 0.0095787738, 0.0102286423, -0.0550979488, -0.0356623717, 0.0321641974, -0.030480044, -0.024687564, 0.0077253585, -0.0015611504, 0.0136776399, -0.005279908, -0.031431146, 0.0172454044, -0.0005588004, -0.0207976718, 0.0078074089, 0.0132870823, -0.0527298786, -0.0307710394, -0.0001255147, -0.0807981342, 0.0087707983, 0.0017379171, -0.0558654033, -0.00927113, -0.0167258121, -0.0483859815, 0.0117115863, -0.045650024, -0.0139551461, 0.0879335329, 0.0047790883, -0.0037891776, 0.0547918193, -0.0288757663, 0.011377505, 0.025994027, 0.0270593874, -0.0010450721, -0.0446722507, -0.0032208958, 0.0205256473, 0.0042192298, 0.0139055755, 0.0018617355, 0.0361710079, -0.0297581665, 0.0014894797, -0.0641607791, -0.0215456113, 0.022837352, -0.0316363126, 0.0421299301, 0.0078545865, 0.0255660191, -0.0193478968, -0.0528881997, -0.0016395493, 0.0521524064, 0.0169767197, 0.0523375161, 0.0226017889, 0.0156400818, -0.0262590721, -0.0266027339, -0.0001068052, 0.0137180509, 0.0371488519, 0.0012845455, -0.0151659874, 0.0286828298, -0.0614107922, -0.0067233061, -0.002454001, -0.073277697, 0.0476669446, -0.0422228836, -0.0078230435, 0.0336823016, 0.0381031483, 0.0287832022, 0.023145292, -0.0119290566, -0.0297603998, -0.0141342189, -0.0540263988, -0.017450003, 0.0445222557, -0.0818722174, -0.0083278408, 0.0386196487, 0.0246475916, 0.0374531373, -0.0297650956, 0.0004250513, -0.0054225116, -0.0480100997, -0.02977084, 0.0356542468, -0.0151440166, 0.019652281, -0.0075688669, -0.0387417823, 0.0610251427, -0.0050274464, 0.0040110219, -0.0342217535, 0.0185851436, -0.0041463021, 0.0489026606, -0.0000929117, 0.0540505797, 0.0262772106, -0.0143679539, -0.0567617752, 0.0004726148, 0.0568711832, -0.0425808132, -0.0139995571, -0.0164177064, -0.0050297854, -0.0267831236, -0.0254274495, -0.0331643857, -0.0259486698, 0.0523788519, -0.0004945554, 0.0021754, -0.0192459058, -0.0596188754, 0.0373344608, 0.0070185019, 0.0239184666, -0.0121340808, -0.0388200954, 0.0091412803, 0.0754652992, 0.047575634, 0.0200112648, -0.040872179, -0.0466067903, -0.0338311642, 0.0708560124, -0.0112053063, 0.032092344, -0.0189839806, -0.0026235045, -0.0178173613, 0.0488238409, 0.0126127554, 0.0272186492, 0.038500756, -0.0343757272, -0.0399208702, 0.0669189841, 0.0204137061, -0.0062238527, -0.0019100815, 0.0439305194, 0.0284488574, -0.0054787048, -0.0050018192, -0.0362307653, -0.0089443047, -0.027144013, 0.0445540026, -0.011660343, 0.0076512182, -0.0022187175, -0.0560057424, 0.0588055514, 0.0122677572, -0.0648723766, -0.0362517498, -0.0156208444, -0.0181573108, -0.0053086914, -0.0361426324, -0.006905769, -0.0164563414, 0.0036205521, 0.0378657766, -0.0318874791, -0.0899583772, 0.0147926118, -0.0445189364, -0.0121545214, 0.0692969188, 0.0367992446, -0.0194679573, -0.0387190692, -0.0131680919, 0.053725794, 0.004246939, -0.000820316, 0.0791086629, 0.0042704409, -0.0218666568, -0.0341929123, 0.0520462133, 0.0124132112, -0.0015030315, -0.0426787026, 0.0112009523, -0.0140956333, -0.0022670641, -0.0821725428, 0.0591771305, 0.0254915226, 0.0444763713, 0.0065086209, 0.0003069002, -0.0289584547, -0.0142255612, 0.0857664496, -0.0736360475, -0.0404476337, 0.0260631628, -0.0639772788, -0.0773215592, -0.0834090784, -0.0448250696, 0.0129503505, 0.0646844581, -0.0414588004, -0.0330590047, -0.0525661036, -0.0151640326, -0.0338522829, 0.0414735712, 0.013513797, -0.0123778107, -0.0344222672, -0.0672798306, 0.0375222415, -0.0089944731, -0.027977312, 0.0687617883, -0.0645189211, 0.0080365352, -0.0190579277, 0.058343932, 0.0223134607, 0.009597715, 0.0024855104, -0.0173213277, 0.0890026242, 0.0883928016, -0.0003641233, 0.00880061, 0.0129812369, -0.0143369585, -0.0250814036, 0.0061807111, -0.0404447578, 0.0296483785, -0.0010551006, 0.0460669175, -0.0206435826, -0.0199340414, 0.0366201177, 0.0147347022, 0.0165752433, -0.0872292891, 0.0070102615, -0.0357646681, -0.0408938229, -0.0308673158, -0.0050107972, -0.0555012561, 0.0160854068, 0.0337316357, -0.0058030104, -0.0037812823, 0.0234996099, -0.0383636691, 0.0342730656, -0.0512624681, 0.004049364, 0.0257971324, -0.0491344258, 0.0122383358, 0.0052425708, -0.0275653489, -0.0386851653, -0.0249951761, 0.0008988017, -0.0099000223, 0.0593087934, 0.0394243188, 0.0520984307, -0.0138760805, 0.0285851099, 0.0057513523, -0.023739256, -0.0730499178, 0.0220923461, 0.0344996937, 0.022742752, 0.0190065093, -0.054551281, -0.0167220384, -0.0218015723, 0.0568927936, -0.0078587858, -0.0165253319, 0.0475889817, 0.0021741868, 0.0475235097, -0.0321679786, 0.1156670004, -0.0035620029, 0.0246661622, -0.0309461225, -0.0460138507, -0.0570167117, -0.0266614761, 0.0249739811, -0.0108848149, 0.0021775498, 0.0168196224, -0.0602719709, -0.0606548637, -0.037051104, 0.0164757352, -0.0137338722, 0.0063799089, 0.0836929753, 0.0688548982, -0.0022909101, -0.0009300351, -0.0197661091, 0.0400470458, -0.0123150367, -0.0303677488, 0.0273075812, -5.890095461e-33, -0.0208832752, -0.0279828366, 0.0287729148, -0.0537797362, -0.0017802058, 0.0399895534, -0.0293314625, -0.0059413291, 0.0429552905, 0.0223674066, -0.0051221158, -0.0279699396, 0.0193370264, 0.0469821244, 0.0163120795, 0.0341902301, 0.0392194092, -0.0236015581, -0.0437734909, 0.0418399796, 0.0071021044, 0.0133838924, 0.0608067773, -0.0843214542, -0.0276089981, 0.0226694159, -0.0372061618, -0.0319472775, 0.0294249468, 0.0594986826, -0.0506711788, 0.0128429495, 0.0122805284, -0.0288359243, -0.0232398212, 0.0012175123, -0.0843753144, -0.0016143316, 0.0020401222, -0.0028487006, -0.0970471501, -0.0004129265, 0.0373792201, -0.0347171575, 0.0087427078, -0.0417411551, 0.0398667119, -0.0107522635, -0.023194911, -0.0194094926, -0.0484234132, 0.0429819562, 0.0079279067, -0.034703996, -0.0206759721, 0.0060550645, -0.0271884333, 0.0150659615, -0.0226209257, 0.0219855346, 0.0493340939, 0.044035241, 0.0516275242, 0.0042296518, -0.028453799, 0.0426568314, 0.0329115465, -0.019076556, 0.015043675, 0.031482961, -0.0232841298, 0.059801504, -0.011030484, 0.0678119361, 0.0596594028, 0.0056084273, -0.0585301891, 0.0695841014, -0.0122488318, 0.0663748011, 0.0296793543, -0.0201696474, -0.0276667234, 0.0043757688, 0.0209966786, -0.0068663051, -0.0174011532, -0.0457250252, 0.0299281869, 0.0387456305, -0.0219487604, 0.0256858692, 0.0299334358, 0.0108646601, -0.0099731041, -0.0158942081, -0.0027391794, 0.0259364489, -0.0368129425, -0.0319990851, -0.0372036174, -0.012608761, 0.0579276867, 0.0100087021, 0.0320814699, 0.0218433514, 0.0409208238, 0.031940598, -0.0449843928, 0.0153443655, -0.0245431643, -0.0192316212, -0.0293153618, 0.0167239588, 0.0278348066, -0.0185981859, -0.0519836284, 0.0019302557, -0.0258922391, 0.001080435, -0.0220607668, -0.0815738887, 0.0058104303, 0.0218103118, -0.061289411, -0.000639236, -0.0332904682, -0.0626904964, 0.0929208994, -0.0103429798, -0.0414802432, 0.019899385, 2.629e-7, 0.0415118746, 0.0135627259, 0.0153076816, 0.0442494862, 0.0094491197, 0.0180165377, -0.0172584876, 0.0529555045, -0.0254128613, -0.0260426067, 0.0212230459, -0.0041793422, 0.0063605644, 0.0455699265, -0.0352128856, -0.0026970543, -0.0601326562, 0.0173157509, -0.002818438, -0.0441204533, 0.0539884008, 0.0668671504, -0.0308713317, -0.0673664883, 0.0258016828, -0.0105952015, 0.0665925369, -0.005488574, -0.0190442763, -0.0008930973, 0.0275030434, 0.0365536772, 0.0081613474, 0.008025324, -0.0041070334, -0.0602515265, 0.0580833852, 0.0435129851, -0.0241334308, -0.0226589348, 0.0294641722, 0.0021041995, 0.0032824681, -0.0162343234, 0.0182597525, -0.0138279516, -0.0087055657, -0.0142535679, -0.0234485753, 0.0406161807, 0.0345826671, -0.063874647, -0.0748023242, 0.0245781448, 0.0344711095, 0.0182889774, -0.0396420062, -0.0598933324, -0.0087285209, 0.0224074982, -0.0306117591, -0.0601006374, 0.0079459734, 0.0020226594, -0.0159054734, -0.0555042811, -0.0450379178, 2.340498792e-34, 0.0399244055, 0.0202186834, -0.0204838011, -0.0272339135, 0.0124089867, -0.0038151974, -0.0054661306, -0.0060143853, 0.0127997659, -0.0325973183, -0.0457595252 ]
http://arxiv.org/abs/2406.17903v1
http://arxiv.org/pdf/2406.17903v1
Mapping the Past: Geographically Linking an Early 20th Century Swedish Encyclopedia with Wikidata
2024-06-25T19:34:00
2024-06-25T19:34:00
Axel Ahlin, Alfred Myrne, Pierre Nugues
cs.CL
cs.CL, cs.DL
In this paper, we describe the extraction of all the location entries from a prominent Swedish encyclopedia from the early 20th century, the \textit{Nordisk Familjebok} `Nordic Family Book.' We focused on the second edition called \textit{Uggleupplagan}, which comprises 38 volumes and over 182,000 articles. This makes it one of the most extensive Swedish encyclopedias. Using a classifier, we first determined the category of the entries. We found that approximately 22 percent of them were locations. We applied a named entity recognition to these entries and we linked them to Wikidata. Wikidata enabled us to extract their precise geographic locations resulting in almost 18,000 valid coordinates. We then analyzed the distribution of these locations and the entry selection process. It showed a higher density within Sweden, Germany, and the United Kingdom. The paper sheds light on the selection and representation of geographic information in the \textit{Nordisk Familjebok}, providing insights into historical and societal perspectives. It also paves the way for future investigations into entry selection in different time periods and comparative analyses among various encyclopedias.
Mapping the Past: Geographically Linking an Early 20th Century Swedish Encyclopedia with Wikidata: In this paper, we describe the extraction of all the location entries from a prominent Swedish encyclopedia from the early 20th century, the \textit{Nordisk Familjebok} `Nordic Family Book.' We focused on the second edition called \textit{Uggleupplagan}, which comprises 38 volumes and over 182,000 articles. This makes it one of the most extensive Swedish encyclopedias. Using a classifier, we first determined the category of the entries. We found that approximately 22 percent of them were locations. We applied a named entity recognition to these entries and we linked them to Wikidata. Wikidata enabled us to extract their precise geographic locations resulting in almost 18,000 valid coordinates. We then analyzed the distribution of these locations and the entry selection process. It showed a higher density within Sweden, Germany, and the United Kingdom. The paper sheds light on the selection and representation of geographic information in the \textit{Nordisk Familjebok}, providing insights into historical and societal perspectives. It also paves the way for future investigations into entry selection in different time periods and comparative analyses among various encyclopedias.
[ 0.0591724664, 0.0051165889, -0.0146976141, -0.0019435985, -0.0237176586, 0.0152775152, 0.0243367739, -0.0066446541, 0.0446034037, -0.0588803217, 0.0279927496, 0.0093387254, 0.0501387641, -0.0422385931, -0.0170748439, -0.0373417363, 0.0136391316, -0.0089686289, -0.0435534604, -0.0143202068, -0.0490040295, 0.049573984, -0.0328537151, 0.0290863365, 0.057112962, 0.0265788306, 0.0034146749, 0.0109280711, 0.0215153247, -0.0327285752, 0.0168323871, -0.0058322535, 0.007996548, 0.0178657435, 0.0000017788, -0.022205798, -0.0133998236, -0.0315536894, -0.0081419451, 0.0333383195, 0.0030031018, 0.0014393949, -0.0082709827, -0.0815456659, 0.0363776088, -0.0268428531, 0.0251474902, -0.0085550928, -0.0268905498, 0.0132404538, 0.012110753, -0.0517607406, 0.0407899916, 0.0473405123, 0.0216983724, 0.0100543983, 0.0057421653, 0.0576406009, 0.0563375317, 0.0393606462, 0.0000919712, 0.0508176051, -0.0434091389, 0.0117775248, 0.0124524916, 0.0485386923, -0.039458368, -0.0014770456, 0.0182605833, -0.0184469298, 0.1108962893, -0.0319136269, 0.0429294519, 0.0343072601, -0.0341621116, 0.0562147461, -0.0083520338, -0.0394839272, -0.0356653929, 0.0050281677, 0.0315228105, -0.0184691045, 0.004771193, 0.0375945456, -0.0129760038, 0.0861075819, -0.0192382243, -0.0020159509, 0.0069552595, -0.0067331041, -0.0051061003, -0.0771214962, -0.0042352253, 0.0201961845, 0.051908154, -0.0097607421, 0.0184811112, 0.0049284543, 0.0093791299, -0.0224671271, -0.0648418218, -0.0289897192, -0.0099171242, 0.0495029539, 0.0592181906, -0.0125099933, 0.0189744849, -0.0275990218, -0.0201823376, 0.006086492, -0.0530153438, -0.0104760295, 0.0035322704, -0.0009582225, 0.0190606024, -0.0145148989, -0.0092555285, -0.022363754, 0.0404239595, 0.0368752964, -0.0498236492, -0.0178614482, 0.010509437, -0.0123629905, -0.0534498058, 0.0014474085, -0.0370989367, 0.0167399719, -0.0814953297, 0.0808106437, 0.0793038309, 0.01076716, 0.0620340742, -0.0162617564, 0.007039255, -0.004504798, 0.0373566672, -0.0002894225, -0.0537088327, -0.0112317773, 0.0231289715, -0.0108646071, 0.0310555864, -0.0428552032, 0.0094885975, 0.0353485756, -0.0162080508, -0.0531633012, -0.0116750086, -0.0121047888, -0.0863194019, 0.0133353546, -0.0356493816, 0.0184362512, 0.0986382589, -0.0080994861, 0.027891187, 0.0162611548, 0.0127059454, 0.0107802786, 0.0633966774, 0.0420938693, 0.0348712057, 0.0144247301, -0.0044670627, 0.0058506746, -0.0018844582, 0.0184005331, -0.0582293831, -0.0165593605, 0.0069679013, -0.0153004555, -0.0233432129, 0.0266776122, 0.0435655862, 0.0639067888, 0.0403758846, 0.004561007, 0.0072782631, 0.0345950201, -0.0344298035, -0.0280737597, -0.0267018285, -0.0354643203, -0.0042911363, 0.0149691282, 0.0065328842, -0.0046313838, 0.0379424505, -0.0134903528, -0.0018979213, 0.0457510985, -0.0006394265, 0.0079640783, 0.0389243364, -0.0395779423, 0.0069454159, 0.0885064229, 0.0082779368, 0.0202107523, 0.0433309413, 0.0698144361, 0.025388401, 0.0707712844, 0.0452266745, 0.0114892228, 0.0257742498, -0.005851984, 0.0261212196, 0.0173391942, 0.0350175872, -0.0068716528, -0.0190984197, -0.0076922285, -0.023914203, 0.0515750609, -0.00292053, 0.0000747, 0.009714704, 0.0864883661, 0.0119564757, -0.0702692121, 0.0213833079, -0.0152684376, -0.0105899209, -0.0072420142, -0.0457061566, 0.0203709379, -0.0447817743, 0.0514715016, -0.0120561915, 0.0093357852, 0.0304026771, -0.0440978706, 0.012543804, 0.0063041644, 0.081656903, 0.058912389, 0.0074594985, -0.0543911047, -0.0080700237, -0.0440912433, 0.0320705809, -0.0217058565, -0.0025562095, -0.00004192, 0.0076382528, 0.0434126258, 0.0099772057, 0.0041862074, 0.0038189178, -0.0287201405, 0.0711458102, -0.0154957678, 0.0209791809, -0.0160639472, -0.0720884874, 0.0135590313, -0.0090075294, 0.0113674626, -0.0010999693, -0.013236315, 0.0141936941, 0.0300705172, -0.0619888976, -0.0011692924, -0.0070846593, -0.0408012569, 0.0071075424, -0.0213811342, 0.0298093241, -0.0069355322, -0.0187990107, 0.0037584309, 0.027125366, 0.001653546, 0.0510458574, -0.0323939584, 0.0003755626, -0.0017114754, -0.0038610508, 0.0076461351, -0.0181553904, -0.0112029137, -0.0142690176, 0.0311926641, 0.030216774, -0.0478039905, -0.0613837615, -0.0037129752, -0.0597940758, 0.0017462492, -0.0196610633, 0.0035465092, -0.0010184225, -0.0044434532, -0.0087105474, 0.0100737195, -0.0281515662, -0.0528783053, 0.0039677811, -0.0464928858, -0.0297552645, 0.0085102702, -0.0214661229, 0.089025788, -0.019882137, -0.056038782, 0.0143025322, -0.0662309453, -0.0132195763, 0.023018999, 0.0250237081, 0.0189856589, -0.0490810461, 0.03796288, -0.0842703208, -0.0538627654, 0.0184958484, 0.1469365805, 0.0129862418, -0.003523573, -0.0321450457, -0.0033707051, 0.0554016121, -0.0447697155, -0.0137445806, 0.0127441538, 0.0935602114, 0.0108591188, 0.0334751196, -0.0205921978, -0.0379888117, 0.0063280747, -0.0172118694, 0.0237316862, -0.0174328368, 0.0215551239, 0.0146854399, -0.013357739, -0.0019383503, -0.0186579432, 0.0019069744, -0.0308182351, -0.0421720333, -0.047505457, -0.0015378582, -0.0363183543, 0.0022555131, 0.0491416752, 0.0286725648, 0.0349162892, 0.0302302819, 0.0243134741, -0.0038438807, 0.0248346906, 0.0478777438, 0.0274692141, -0.0254396014, 0.0563447066, -0.0852998942, -0.0168561041, -0.022996515, 0.0196027812, 0.0787365213, -0.015542658, 0.0289102346, -0.0637461767, -0.0963201523, -0.0478446148, 0.0124547249, -0.045224756, 0.0161850583, 0.0343051888, 0.0574776456, -0.0295516849, 0.0003483458, 0.0129485698, -0.0569409765, 0.0125454441, 0.0220364481, -0.0490897186, 0.0724383965, 0.0063273017, 0.0036921075, -0.0532893911, -0.0322399065, 0.0076237987, -0.0486020558, 0.0137055507, 0.0052173673, -0.0544821434, -0.0348925553, -0.0165435057, -0.0481414385, -0.0245559346, -0.0587041825, -0.0036948642, -0.006291287, 0.0169658214, -0.0358921699, -0.0968062654, -0.0007631873, 0.0238159876, -0.0431069396, 0.0287798643, -0.0038074506, 0.084514901, -0.0284075327, -0.0489213541, -0.0098649142, -0.0125473188, -0.0332490355, -0.0144951055, 0.094335705, 0.0052762148, 0.0138845677, -0.0295978244, 0.0124175688, 0.0251644794, -0.0247673485, 0.0252309497, 0.0133128054, 0.0309453625, 0.0200066473, 0.0250241682, 0.0151581196, 0.0734773502, -0.0177616794, -0.0190492868, 0.0085784439, 0.0054616039, -0.0483907983, 0.046295274, 0.0190635826, -0.0302050114, 0.0392520577, -0.0157379191, -0.0806401893, -0.0704666227, -0.0499925353, -0.0225334503, 0.0787707716, 0.0558599047, 0.0094265565, 0.0331962518, -0.0260935277, -0.0507998765, -0.0139829423, -0.0055796984, -0.0166130923, 0.0349331647, 0.0043282192, 0.0491467044, -0.0026973167, -0.016964158, 0.002835935, 0.0387553722, -0.0415870398, 0.0289933868, -0.0015629806, 0.0251961965, 0.029302299, -0.0276057534, -0.0300661419, 0.040598426, 0.0068517858, -0.0200568903, 0.0083287908, -0.0246355217, 0.0635768473, -0.0086578634, 0.00038035, -0.035910625, 0.0282286108, -0.0255701654, -0.0485900566, -0.0246195942, 0.0131204054, -0.0337043703, 0.0021967161, 0.0486480445, -0.0473556705, -0.0002229352, -0.0065554227, -0.0095143309, 0.0197924078, 0.0030278906, 0.0334250294, -0.0345841646, 0.0162563175, -0.0261473823, -0.0554152839, 0.0534440354, -0.0137578342, 0.049462717, -0.0594188459, 0.0122487471, 0.0569034107, 0.0649541691, 0.0009450682, 0.0089661758, -0.0223818645, 0.0273375083, 0.0249004178, 0.0793787986, -0.036217235, -0.0383216888, 0.0043538818, -0.0366010703, -0.0490056351, -0.037593361, -0.0598258637, -0.0000265176, 0.0362375006, -0.0117845377, 0.0397271216, -0.0294423588, 0.0661099404, 0.0082053877, -0.024598496, 0.0228153057, -0.0015687109, 0.0171444248, -0.037075188, -0.0072430223, 0.0132691199, -0.024250241, 0.069869861, -0.0385842845, 0.0005214026, 0.0256481711, 0.0602429993, -0.0516175553, -0.0208542049, 0.0149697205, 0.0201033019, -0.0230999775, -0.0158218537, -0.0041536447, -0.0664577335, -0.0066872663, -0.0169686899, 0.0133325709, -0.0381318629, 0.0168242902, -0.0038724202, 0.0147327557, 0.0266175196, -0.0302128941, -0.0588219203, 0.0471303836, -0.1471533179, -0.0311244875, 0.0010320242, -5.874101759e-33, 0.0147966035, -0.0547721162, 0.0420462638, -0.0016551663, -0.0314439908, -0.0467905514, 0.010943695, -0.0069084424, -0.0398361497, -0.0084278863, -0.0260155872, -0.0348033756, 0.0181115977, -0.0417716391, -0.0216054339, 0.0010310692, -0.0171181764, 0.0281669013, -0.0119089978, -0.0171874948, 0.0409288034, -0.0172569733, 0.076366432, -0.0874282494, 0.0383785255, -0.0066709672, -0.0236281734, -0.0116434125, 0.1176823154, 0.0248859897, -0.020143427, -0.0180080924, 0.004621462, -0.0199835524, 0.0368130617, -0.0093943644, -0.0679851323, -0.0161182173, 0.019361157, 0.0397983342, 0.0872881114, -0.0410667956, 0.0425403193, 0.0122781554, 0.0021558029, 0.0345980562, 0.0442378931, 0.0192389674, -0.0669060871, 0.0322624147, -0.0307170618, -0.0032152203, -0.0178067163, 0.0392824821, 0.0325360037, 0.0255874265, 0.0290038381, -0.0431593172, -0.0801891014, 0.0272573084, 0.0468755811, 0.0007240924, 0.0211792327, -0.0527540632, 0.010563774, 0.0318569392, 0.0040636659, 0.0224902313, -0.0672060326, -0.0079311905, 0.0446993224, -0.0082987826, -0.0312986597, 0.0638510436, -0.0154615641, -0.030620046, -0.0461017713, 0.0356971137, -0.0338843055, -0.000579647, -0.00624375, -0.0289588384, -0.0571949966, -0.0294879153, -0.0097208321, -0.046571549, -0.0005503976, -0.0279842597, 0.0065016197, -0.0326648727, -0.0486485586, 0.0509826168, -0.0299462527, -0.0458063371, -0.104140453, 0.0363979526, -0.0291065071, 0.0034786463, -0.0336219519, -0.0212681461, -0.0201789215, -0.0224990454, -0.0261277258, 0.0589838326, 0.0105052982, 0.0415234379, -0.0169558283, 0.0250942521, -0.0318352543, -0.0073773963, 0.0277332123, -0.0291579328, -0.01956545, -0.0614839047, 0.0406847522, -0.0120932441, 0.0322134048, -0.0471519455, 0.0723896474, 0.0221413635, 0.0216613673, -0.0412159301, 0.0110309348, 0.0180715844, -0.0047274195, 0.003520122, 0.0152403796, 0.0559026934, 0.0251238886, 0.0254176054, 0.0127194487, -0.0419657826, 2.601e-7, 0.0385088995, 0.0674271733, -0.0087180939, -0.0210227612, -0.0136908665, -0.0267266445, 0.0370424576, 0.0109229293, -0.0411059186, -0.0101825222, 0.0536074601, -0.0057618967, -0.0115442593, 0.0871026367, -0.1131827086, -0.0186802279, 0.0115247387, 0.0079361591, -0.0763972774, 0.0009274062, 0.0263580158, 0.0192267466, 0.0146960113, -0.0053009177, -0.0093065351, -0.0185666792, -0.0222727712, -0.072269462, 0.0070476197, -0.0064697675, 0.0548385233, 0.0060317181, 0.0169896241, -0.0230409596, -0.0112977456, -0.0020612541, 0.0071677663, 0.0057373447, 0.0056663761, 0.0129411723, -0.0253876373, -0.0389390588, 0.0087549621, -0.0838002115, 0.0513436608, -0.0045106472, -0.0284794606, 0.0575672761, -0.0104216058, 0.0252963323, 0.0207780991, -0.007019416, -0.01801374, 0.0059376745, 0.002756886, 0.0196053106, 0.0217520762, -0.0269604623, 0.0023656485, 0.0308504831, 0.0146281412, -0.0199183021, -0.0261820238, -0.0129289236, -0.0333978795, 0.0143502587, 0.0313260145, 1.82551311e-34, -0.0224836711, -0.0251316465, -0.0698371157, -0.0062403642, 0.011105773, -0.0026393086, 0.0070715994, 0.0290585924, 0.0173150208, -0.0169268008, 0.0039116722 ]
http://arxiv.org/abs/2406.17901v1
http://arxiv.org/pdf/2406.17901v1
Script-Agnostic Language Identification
2024-06-25T19:23:42
2024-06-25T19:23:42
Milind Agarwal, Joshua Otten, Antonios Anastasopoulos
cs.CL
cs.CL
Language identification is used as the first step in many data collection and crawling efforts because it allows us to sort online text into language-specific buckets. However, many modern languages, such as Konkani, Kashmiri, Punjabi etc., are synchronically written in several scripts. Moreover, languages with different writing systems do not share significant lexical, semantic, and syntactic properties in neural representation spaces, which is a disadvantage for closely related languages and low-resource languages, especially those from the Indian Subcontinent. To counter this, we propose learning script-agnostic representations using several different experimental strategies (upscaling, flattening, and script mixing) focusing on four major Dravidian languages (Tamil, Telugu, Kannada, and Malayalam). We find that word-level script randomization and exposure to a language written in multiple scripts is extremely valuable for downstream script-agnostic language identification, while also maintaining competitive performance on naturally occurring text.
Script-Agnostic Language Identification: Language identification is used as the first step in many data collection and crawling efforts because it allows us to sort online text into language-specific buckets. However, many modern languages, such as Konkani, Kashmiri, Punjabi etc., are synchronically written in several scripts. Moreover, languages with different writing systems do not share significant lexical, semantic, and syntactic properties in neural representation spaces, which is a disadvantage for closely related languages and low-resource languages, especially those from the Indian Subcontinent. To counter this, we propose learning script-agnostic representations using several different experimental strategies (upscaling, flattening, and script mixing) focusing on four major Dravidian languages (Tamil, Telugu, Kannada, and Malayalam). We find that word-level script randomization and exposure to a language written in multiple scripts is extremely valuable for downstream script-agnostic language identification, while also maintaining competitive performance on naturally occurring text.
[ 0.0125886453, 0.067239292, -0.0227795485, 0.0449290834, -0.0620191097, 0.029164087, 0.0140951779, 0.0629899651, 0.0490732193, -0.0039109327, -0.0120156994, -0.0493550189, 0.0245351642, 0.0436066948, 0.0406960137, -0.0015852863, 0.0034534694, 0.0008126332, 0.0183574744, 0.0246229246, 0.0109805735, 0.0269646943, -0.0132283382, -0.003853878, -0.0284003168, -0.0074507133, -0.0072533982, 0.0229743347, -0.00444867, 0.0073311175, -0.0147718629, 0.0498419814, 0.0425235294, 0.1029859558, 0.0000012312, -0.0325564854, -0.0100996941, -0.0392885171, -0.0054197386, 0.0013721721, 0.0732858628, -0.0153160663, 0.0149607286, 0.0199781954, 0.0006500728, -0.0094710896, 0.1021235734, -0.0155164227, -0.0164624173, 0.0818453357, -0.0257604495, -0.0197753217, -0.027942149, -0.0122807166, -0.0100325905, 0.0231912397, 0.0065927166, -0.0226198826, 0.0051273634, -0.0437050462, -0.0383498557, 0.0077328859, -0.025273094, -0.0140751218, 0.0002871189, 0.0153226051, -0.0233234297, -0.0056639398, 0.0679637566, 0.0404772051, -0.0692716017, -0.0597712025, -0.0007605918, -0.0132549284, -0.0539945923, -0.0087085739, -0.0270952415, 0.0269849468, -0.0480221659, 0.0727126822, 0.0407132804, 0.0261740871, 0.0345058218, -0.0049301041, 0.04141175, 0.0293280482, 0.0188029353, 0.0010407275, -0.0575477071, 0.031677831, 0.0338158235, -0.0314829051, 0.0119115124, -0.0007612474, 0.0139245214, 0.0095193675, -0.0046304069, -0.0544311814, 0.0099593196, -0.0011075559, 0.0788964704, 0.0624227114, -0.0321396515, 0.0126980487, -0.0545312688, 0.0517836213, -0.0097895162, 0.0264044311, 0.0120593095, 0.0233868659, 0.0526572689, -0.0402096026, 0.023409402, -0.0143657755, -0.0192069188, 0.039360635, 0.0103154788, 0.0266110897, 0.0178749505, -0.027475819, -0.0170620978, -0.0547167622, -0.0443883613, -0.0019582175, 0.0205089599, 0.0085579995, -0.0177947599, 0.0326403752, -0.0007154312, -0.0949488878, -0.0455739386, 0.0700245798, 0.0084040239, 0.0183015652, -0.0095763365, 0.073838599, -0.0622387305, -0.0553084463, -0.054544013, 0.0442409478, 0.004783723, -0.0270234477, 0.0196326934, -0.078990601, 0.0674337372, 0.0206218362, -0.0015740504, -0.0580365956, -0.0002824913, 0.0146006411, -0.005144624, 0.0452376939, 0.0392708033, -0.0193245057, 0.0349397473, 0.006793844, 0.0578518026, 0.0447941236, 0.0020450221, 0.0229275394, 0.0292237569, -0.0220732521, 0.0193672646, 0.0321778655, -0.0323271379, 0.0290421862, -0.0319843665, -0.0238261186, 0.0562519431, 0.011646294, 0.0133335628, 0.0161667708, 0.0061927182, -0.0119760809, 0.0080542043, 0.0324864872, -0.02353031, 0.0100447657, 0.0053563998, 0.0080310088, 0.0085065588, 0.062724404, 0.0008423689, 0.0197889283, 0.0203377437, -0.0090489844, 0.0379526988, 0.0109273344, -0.0370943062, 0.0334450938, -0.0209477525, -0.0403849781, 0.0400950834, -0.035492748, 0.0550198741, 0.0295486487, -0.0331372619, -0.0527220145, 0.0991255641, -0.0387763977, 0.0560318232, -0.0353690386, 0.0358996801, 0.0088664955, 0.0226339791, -0.0548192076, 0.0203988552, 0.0190687031, -0.0012176617, 0.1026594713, -0.0360881351, -0.0162038058, 0.0489048362, -0.0384153239, 0.0308883078, 0.0372761525, -0.0039711567, 0.0471615121, -0.0209438726, 0.0108852275, 0.0024058474, -0.0310501363, -0.0214997996, 0.0490767807, -0.0584775433, -0.025090564, 0.0199835114, 0.0063149198, 0.0308894366, -0.0308721084, -0.042041678, -0.0164709035, -0.0075554494, -0.0399095044, -0.001932657, -0.0080369217, -0.0222822819, -0.039676372, -0.0013003867, -0.0448628627, 0.0140312184, 0.0454208925, -0.008663971, -0.0515925065, 0.0395992622, -0.0199877433, 0.0376842059, -0.0249026548, 0.0261729695, -0.053867057, -0.0173484217, -0.0474234261, 0.0477898307, 0.0213089865, 0.0348855369, -0.0076852725, -0.032506682, 0.072865583, -0.0432591215, -0.0207194425, 0.0128405103, -0.0054746023, 0.0369865373, 0.0023645307, -0.0172674526, -0.0241759829, -0.0587257631, -0.0695287362, -0.0501834005, -0.0232703276, 0.0107287187, -0.0072302371, 0.0246492233, 0.0102128917, -0.0492005534, -0.0490580052, -0.0135029769, 0.0677396208, 0.0221393332, 0.0341019332, -0.0206549931, 0.024285648, -0.0100108199, -0.066143468, -0.0355167761, 0.0946159959, 0.0002387664, -0.0343078896, -0.0571352653, -0.006108765, -0.0332007334, -0.003626494, -0.0283444654, -0.0466762036, -0.0109792473, -0.0282426625, -0.0117539, 0.0554016083, -0.018361954, -0.0392575487, 0.0449764803, 0.0001076681, 0.0223622024, -0.0670181662, 0.0030484335, 0.0166049991, -0.0043469192, -0.0363676734, -0.0186209995, 0.049331028, 0.0060269325, 0.0056485892, -0.0424250551, -0.0404495522, -0.019143546, 0.0464418791, -0.0065083755, 0.0146071957, -0.0437029786, 0.0569423102, -0.0157252327, -0.0629601777, 0.0817690194, -0.0380897075, -0.0230077934, -0.0670572966, 0.0029951229, -0.0250931643, -0.0126290191, 0.0331466645, 0.0175338835, -0.0486597456, -0.0540277362, -0.0688586906, 0.0414323807, 0.0204459261, -0.0155305183, 0.0331437401, -0.0093907621, 0.0199460667, -0.0443044528, -0.0482109524, -0.0390185155, -0.0922212303, 0.0376261435, 0.0134799378, 0.0226883423, -0.0413681231, 0.0156810433, 0.0323665291, 0.0072314036, 0.0655715466, -0.0041611046, 0.0243397262, -0.0156189576, 0.0390026793, 0.0125249019, -0.0084071476, -0.0028447025, 0.0088886255, 0.0080662444, -0.0235254616, 0.0567425787, 0.0005519647, 0.022350993, 0.0361059457, 0.0172313079, -0.0162165724, 0.0016599658, -0.0088293469, -0.005891134, 0.0494174547, -0.0044554751, 0.0326211452, 0.0155284209, -0.0676728636, -0.0100941164, -0.0580380745, 0.0329591818, 0.0228163693, 0.0221276507, -0.0493759289, 0.0143540641, 0.0137222707, -0.0008715764, 0.0104319509, -0.0202590805, -0.0153318048, -0.072945796, -0.0373359285, 0.0165294334, -0.0259289388, -0.0282647349, -0.0458827913, -0.0105577894, 0.000735239, 0.0003091365, 0.0611632019, -0.0436004512, -0.0240248572, 0.0338652059, -0.0067094276, -0.0319127552, -0.0318538919, -0.0239808615, -0.0193422474, -0.0128111541, -0.0285467114, -0.0676928684, -0.0008206514, -0.0158793721, -0.0518213995, -0.0293450579, 0.008395073, 0.0625158176, 0.0247176439, -0.046955619, -0.0215620063, 0.014810265, -0.0409174971, -0.0369842276, -0.0566939488, 0.041812636, 0.0095657855, -0.0282386951, 0.0484146774, -0.012305609, 0.0089505184, -0.0150476201, 0.0451466516, -0.0401926525, 0.0099980719, -0.0335157663, 0.040899571, 0.0264292173, -0.0260089319, 0.0524454154, -0.0015223161, -0.057550244, -0.0378655083, -0.0327422209, 0.0002937123, 0.0414094627, 0.0428466499, 0.0238572266, 0.0005928545, -0.0429574773, -0.0544293821, -0.0296084881, -0.0329998881, 0.0039043687, -0.0402065143, -0.0282827076, 0.0256963745, -0.031543687, -0.0130128004, -0.0199981146, -0.0759255439, -0.0905880406, 0.0383350998, -0.012351092, 0.0235018786, -0.0088668326, -0.0222469177, -0.0505723469, 0.0460569523, 0.0011004767, 0.0002585125, 0.0319548734, 0.0057784859, -0.0264386013, -0.0245697033, 0.0508199595, -0.0298759304, -0.0154658528, 0.0248598009, -0.0472370051, -0.0362820104, 0.0160323046, 0.009065534, 0.0487016253, 0.0284348894, -0.0610979423, 0.0178662799, -0.0598132499, -0.002447424, 0.013162503, 0.0072863181, 0.0080157919, -0.0437251627, 0.0310699623, 0.0530420728, -0.0022701682, 0.0522877797, 0.0211791378, 0.025537841, 0.0202150568, 0.0440109521, 0.0198973697, 0.0702780932, 0.0092705451, 0.0471519604, -0.0148414252, 0.0534369536, -0.0561106019, 0.0384671241, -0.0005329076, 0.0225243494, 0.018852694, 0.0219450668, 0.0077018202, 0.0243430305, 0.0304658357, -0.0194980409, -0.0146533297, -0.0211425442, 0.0451918878, -0.0633305088, 0.0269102789, -0.0214316193, 0.0257981643, -0.0455319993, 0.022325756, -0.0060305409, -0.0238075349, 0.0157229882, -0.0086525138, -0.0442607775, 0.0232780371, 0.1139028445, 0.0371311791, -0.0145842079, 0.0136444354, 0.0157420412, -0.0150121013, -0.0284143388, 0.0225043464, 0.037392199, 0.0328845419, -0.0046283323, -0.0220121406, -0.0712033734, -0.0045091165, -0.0530658029, 0.0371228494, -0.0299303569, 0.0139833838, 0.0395559967, 0.0430178195, 0.0527779944, -0.0158261582, 0.0343219675, 0.005206279, -0.0168541223, -0.0123856561, -5.115143831e-33, 0.0072113592, 0.000960681, -0.0725122988, 0.0286535863, 0.0373505838, 0.0109591288, -0.034744177, 0.0270650666, -0.0214542244, 0.0032129518, -0.0264004823, 0.0376354121, 0.0120504666, 0.0251782741, 0.0188299902, -0.0306295622, 0.0227676667, 0.0241980888, -0.0357562527, 0.0166768786, 0.0987806097, -0.026010979, 0.1023292392, -0.0351120867, -0.0115552843, -0.0785356238, 0.0030869152, 0.0144030452, -0.0065239943, 0.0079973675, -0.0453645326, 0.0310541205, 0.0399820097, 0.041968476, -0.0160183907, 0.016738141, -0.0185746308, -0.014143521, -0.0026889432, 0.0782342404, -0.0272081215, -0.0276482254, 0.039916385, -0.0591168813, -0.0383740924, 0.0167367067, 0.0101679983, 0.0278179441, -0.022497138, 0.0110290358, -0.0578281134, 0.0418615639, 0.0110105034, 0.0051140944, 0.0381269827, 0.0118159214, 0.0249925442, 0.0108016208, -0.0556283332, -0.0119500821, -0.0667924955, 0.0590129681, 0.0387503207, 0.0468268916, -0.057990741, 0.0497799926, 0.0047351653, -0.0234221723, 0.0010641147, -0.0179913305, 0.028415123, -0.0357409194, 0.0221385267, 0.0166797098, 0.0391397327, -0.0821524709, -0.0360265449, 0.0104900505, 0.0711107776, -0.0386074856, 0.0339880213, 0.0053818426, -0.0383900106, -0.049633041, 0.0041327295, -0.0558024831, -0.0294580273, -0.0157051422, 0.0182733182, -0.0235652551, 0.0188202858, 0.000195469, 0.0559586808, -0.0339899138, 0.0324285477, 0.0283290464, -0.0469285697, 0.0454745404, 0.0311330222, -0.010280801, -0.030076718, -0.0503616333, -0.013078968, 0.0032423977, 0.0130556952, -0.0085106092, 0.0583208203, 0.005937451, -0.0688009337, 0.0239437316, -0.0161634218, 0.0228585843, 0.0545125157, 0.0599531606, 0.0328087285, -0.0149806123, 0.013914872, -0.0303982534, 0.0803074688, 0.0213413686, 0.0108609423, 0.0239775181, -0.0414954051, 0.0160205886, -0.0198727399, 0.0025496532, -0.0430019088, 0.0245893858, 0.0169374123, -0.014024864, 0.0211329758, 0.0207031649, 2.087e-7, -0.0119504994, 0.0060001854, 0.0531806387, -0.0384884402, 0.0300149769, 0.0135239381, -0.0543212369, 0.0256321114, -0.0396243632, -0.0099389292, -0.0437944271, 0.0056870896, 0.0790551156, 0.0297347568, -0.0763279274, 0.0271252487, -0.1026995853, 0.0264312439, -0.0254849754, -0.0311947744, -0.051202938, 0.0958590955, 0.0157014579, -0.0169883482, -0.010940847, -0.0537552536, 0.0616714358, 0.0097793322, 0.0227393638, 0.0084333066, -0.0503843315, -0.0248887613, -0.0220720507, -0.003193455, -0.0233953167, -0.0081735794, 0.0472459719, 0.033080738, -0.0342856459, 0.0478722714, 0.0120694479, 0.0100179436, -0.0134295393, -0.0425028391, 0.0264015589, 0.0836032778, -0.0672332719, -0.007403519, 0.0200614352, 0.0016941606, 0.0493345372, 0.0189028811, -0.0310375839, 0.0337487496, 0.0169260669, -0.0334165618, -0.0072156214, -0.0232915115, -0.0545862429, 0.0317225605, -0.0385108814, -0.0414337441, -0.005131003, -0.0049638059, 0.0700995922, -0.0286115352, 0.0265639089, 1.95660851e-34, 0.0096403193, 0.0131272767, 0.0046610651, 0.0039756345, 0.0133946883, -0.0218465962, -0.0047543249, 0.0285217557, 0.0243552439, -0.0762094483, -0.0015471175 ]
http://arxiv.org/abs/2406.17888v1
http://arxiv.org/pdf/2406.17888v1
CTBench: A Comprehensive Benchmark for Evaluating Language Model Capabilities in Clinical Trial Design
2024-06-25T18:52:48
2024-06-25T18:52:48
Nafis Neehal, Bowen Wang, Shayom Debopadhaya, Soham Dan, Keerthiram Murugesan, Vibha Anand, Kristin P. Bennett
cs.CL
cs.CL, cs.AI, cs.LG
CTBench is introduced as a benchmark to assess language models (LMs) in aiding clinical study design. Given study-specific metadata, CTBench evaluates AI models' ability to determine the baseline features of a clinical trial (CT), which include demographic and relevant features collected at the trial's start from all participants. These baseline features, typically presented in CT publications (often as Table 1), are crucial for characterizing study cohorts and validating results. Baseline features, including confounders and covariates, are also necessary for accurate treatment effect estimation in studies involving observational data. CTBench consists of two datasets: "CT-Repo," containing baseline features from 1,690 clinical trials sourced from clinicaltrials.gov, and "CT-Pub," a subset of 100 trials with more comprehensive baseline features gathered from relevant publications. Two LM-based evaluation methods are developed to compare the actual baseline feature lists against LM-generated responses. "ListMatch-LM" and "ListMatch-BERT" use GPT-4o and BERT scores (at various thresholds), respectively, for evaluation. To establish baseline results, advanced prompt engineering techniques using LLaMa3-70B-Instruct and GPT-4o in zero-shot and three-shot learning settings are applied to generate potential baseline features. The performance of GPT-4o as an evaluator is validated through human-in-the-loop evaluations on the CT-Pub dataset, where clinical experts confirm matches between actual and LM-generated features. The results highlight a promising direction with significant potential for improvement, positioning CTBench as a useful tool for advancing research on AI in CT design and potentially enhancing the efficacy and robustness of CTs.
CTBench: A Comprehensive Benchmark for Evaluating Language Model Capabilities in Clinical Trial Design: CTBench is introduced as a benchmark to assess language models (LMs) in aiding clinical study design. Given study-specific metadata, CTBench evaluates AI models' ability to determine the baseline features of a clinical trial (CT), which include demographic and relevant features collected at the trial's start from all participants. These baseline features, typically presented in CT publications (often as Table 1), are crucial for characterizing study cohorts and validating results. Baseline features, including confounders and covariates, are also necessary for accurate treatment effect estimation in studies involving observational data. CTBench consists of two datasets: "CT-Repo," containing baseline features from 1,690 clinical trials sourced from clinicaltrials.gov, and "CT-Pub," a subset of 100 trials with more comprehensive baseline features gathered from relevant publications. Two LM-based evaluation methods are developed to compare the actual baseline feature lists against LM-generated responses. "ListMatch-LM" and "ListMatch-BERT" use GPT-4o and BERT scores (at various thresholds), respectively, for evaluation. To establish baseline results, advanced prompt engineering techniques using LLaMa3-70B-Instruct and GPT-4o in zero-shot and three-shot learning settings are applied to generate potential baseline features. The performance of GPT-4o as an evaluator is validated through human-in-the-loop evaluations on the CT-Pub dataset, where clinical experts confirm matches between actual and LM-generated features. The results highlight a promising direction with significant potential for improvement, positioning CTBench as a useful tool for advancing research on AI in CT design and potentially enhancing the efficacy and robustness of CTs.
[ 0.0341914967, 0.0756905377, -0.0192269739, -0.0242245998, -0.0243775304, 0.0208011065, 0.071753785, 0.0120807076, -0.011140896, 0.0038926164, -0.0147846732, -0.0505289957, -0.0056777154, 0.0771829039, 0.0144705046, -0.0309538767, 0.0131591083, 0.0019572678, 0.0563207082, 0.0140008926, -0.0389234349, 0.0025511163, -0.0222866684, 0.0511494279, 0.0047053206, -0.0197427701, 0.0430521443, -0.0079397177, -0.0176839568, -0.0628404841, 0.0042569963, 0.0680394247, -0.0148774376, 0.0306352973, 0.0000022159, -0.0905337185, -0.0238326192, 0.0547323637, -0.0360464752, -0.0155102508, 0.0427501835, -0.0020905447, 0.0230113957, 0.0073267017, -0.0024875994, 0.0003001198, 0.054423105, 0.0217079353, 0.0090541271, 0.0622982308, -0.0286918804, 0.0118649285, 0.0317558646, -0.0489570573, 0.0172999706, -0.0188161898, 0.0177911948, -0.0292122439, 0.0303079672, -0.0133354422, 0.0137721039, 0.0201764591, 0.0203935672, 0.0151913585, 0.0276607331, 0.0364012495, -0.0563330501, -0.0421245173, -0.0049047717, 0.0363325216, 0.038207233, 0.0253060348, -0.003829458, -0.0277357064, 0.0505188182, -0.003145199, -0.0419027582, -0.0174984876, -0.0268248674, -0.0285738278, 0.0098455567, -0.0230919383, 0.0008203477, -0.037046209, -0.0042560156, 0.0300340001, 0.0061176876, -0.0091308663, 0.0297157448, 0.0128258131, -0.0010675407, -0.0330138579, -0.0131121157, 0.0064102965, -0.0487550497, 0.0058370708, 0.0343899578, -0.044862479, 0.0949405432, -0.0560828857, 0.0462405197, 0.0061939424, -0.0482289158, 0.0035554781, -0.0240444802, 0.0763537362, 0.0032353913, 0.0477878526, -0.0383026972, 0.0480201356, -0.0162976403, 0.0065170415, -0.0124134906, 0.0227757208, 0.0028059569, 0.0062274965, -0.0632433221, 0.0036531377, -0.0039785253, -0.0005793672, 0.0035000546, 0.0110396361, -0.0352547467, 0.0295529347, 0.0077498937, -0.0174645502, -0.0123401731, 0.0293434914, -0.0030779187, -0.0987271816, 0.0205086004, -0.0103252809, -0.0103519494, -0.0558442585, 0.0313966833, 0.0536379144, 0.0111975381, -0.0169239715, 0.0030645821, -0.0054606623, -0.0385302752, -0.0598364659, 0.0032766187, -0.0941651762, 0.0411215946, 0.0087011326, 0.0120143993, -0.0554232858, -0.0360451825, -0.0061836513, -0.0432339795, 0.0336932056, -0.0424055606, -0.0084495079, 0.0114945751, -0.0022808809, -0.0268960688, 0.035697937, 0.0466952547, 0.005648735, -0.0059435768, -0.0248402562, -0.0083718942, 0.0142145166, 0.041222319, -0.0077486564, -0.03009299, 0.0128193153, 0.0270324573, 0.0557815209, -0.0480367765, 0.0299959555, -0.0050452584, 0.0108647598, 0.0432560667, 0.0455508232, 0.03560929, 0.0631712005, 0.0012603491, -0.0749338418, 0.0159760732, -0.0269900765, -0.027579397, -0.0030422898, -0.025535496, -0.0154132126, -0.0021778704, 0.0109326551, -0.0270922203, -0.0367941819, -0.0276353396, -0.0144672329, -0.0062330407, -0.0064842738, 0.0297724307, -0.0417864509, -0.0237118863, -0.0400949605, 0.0065094866, -0.0044937069, -0.0495747738, -0.0618125126, -0.0302423518, 0.0464180484, 0.0348661281, -0.0353996381, -0.0229005888, -0.0335837901, 0.0123057272, -0.0015011732, 0.0361298583, 0.0177206341, 0.0314262845, 0.0202323701, -0.028100865, -0.0118231485, 0.0136341397, 0.0170880277, -0.0206420012, 0.010192913, -0.0473299511, -0.0436324067, -0.04885244, 0.0602572188, -0.0547512844, -0.0503049195, 0.0188190602, -0.0303502902, 0.0491724014, 0.0043440433, -0.0307607986, 0.0343801975, 0.0035605102, -0.0275609996, -0.0110013047, -0.0094609959, 0.0293559674, -0.0431033634, 0.007804581, -0.034734305, 0.0088998629, -0.0080179032, -0.042045936, 0.0050070123, 0.0173088964, -0.0336006545, 0.0147885559, 0.0298131146, 0.0308394786, -0.0187671576, 0.0235275701, -0.0579841919, 0.0207700748, -0.0038989279, -0.0069883345, 0.0399638675, -0.0732381791, 0.0248551872, -0.0382074118, -0.0442322828, 0.031096885, -0.0200937819, -0.0193506461, -0.0019814519, 0.0258230381, -0.0257091075, -0.0356350578, -0.1033574566, 0.0326191038, 0.0189346671, 0.0266638231, -0.0135425786, -0.0200984795, -0.0184533857, -0.0165172145, -0.0323646143, -0.030847881, 0.0327148549, 0.0482349694, 0.0201402847, 0.0525330417, -0.0573512018, -0.0796044022, -0.0320135057, -0.0595776066, 0.064618215, 0.0512916967, -0.0407880098, 0.0396053344, 0.0088945776, -0.0085240211, -0.0060056681, 0.0080750724, -0.001019582, 0.034435153, 0.0176458508, 0.014769149, 0.0623742789, -0.02800004, -0.0043082237, 0.0431654193, -0.0137036201, -0.0313624255, 0.0013860653, -0.0262916125, -0.0636612326, 0.0123758065, -0.0741388574, 0.0274786223, 0.0729548559, 0.0199269522, 0.0152566889, -0.0522183888, -0.008197221, 0.0128726382, -0.0090255188, -0.0529496297, -0.0030896943, -0.0472121201, 0.0384674482, -0.0346123725, -0.0808364674, 0.1035536975, 0.0002450657, 0.0443494432, -0.0628963336, 0.0452455729, -0.0374187529, -0.020195879, 0.0119232051, -0.0068483148, 0.0525164306, 0.0165330805, 0.0103894742, -0.0143350372, 0.0101523744, -0.036706306, 0.0631595477, -0.0483478606, 0.0450502709, 0.0173866525, -0.0202427488, -0.0305760633, -0.0182378609, -0.0105839316, 0.0193493143, 0.0193896759, 0.0352931246, -0.0223300364, -0.0052200472, -0.0123137636, 0.0463409163, -0.0271056239, -0.0006254901, 0.0224355962, -0.0466905981, 0.0311738905, -0.0229706317, 0.0305348299, -0.033381097, 0.0521252416, 0.0001658151, 0.009361051, 0.0280123502, -0.0325393826, 0.0652260631, 0.0034656103, -0.0252065212, -0.0288178697, 0.0579075031, 0.0233681742, -0.0258229151, 0.0118532041, 0.1106526181, 0.0211081784, -0.0722102076, -0.0091026407, -0.0147434669, 0.0187582504, -0.0165532138, 0.0035468573, -0.0455975346, 0.0014794369, -0.0020367943, 0.0541228242, -0.043660298, -0.0300530717, -0.0132447788, -0.0304994583, 0.0319438651, 0.0041996245, -0.0770604685, 0.0154073006, -0.0690118223, 0.0185477659, -0.0430560485, -0.0254639834, 0.0014483858, 0.022851998, -0.0220092125, 0.0217810851, -0.088997595, -0.0824803934, 0.0368411168, -0.0273499768, 0.0878796354, 0.0671746507, 0.0049946653, -0.0139297023, 0.0031669752, 0.0173491612, -0.005021669, -0.0077113737, -0.0068503167, 0.095652245, 0.0318542235, 0.0342151821, -0.0309072621, 0.0406908393, -0.0361609943, -0.0067142751, -0.0064736982, 0.0983502716, -0.0457407385, -0.0156071009, -0.0203657392, 0.009038222, 0.0162936635, 0.0409341715, 0.0449333228, -0.0157973506, -0.0042780936, -0.0232685208, 0.0228318833, -0.0398058556, 0.0068755383, 0.0124187581, -0.0370259546, -0.0190034024, -0.0379893109, -0.0461801849, 0.0565834381, -0.0489615612, 0.0035964937, 0.0442168191, -0.0349595509, -0.0220090486, -0.0182007216, 0.0687083304, 0.0355310962, 0.032162603, -0.0726945847, -0.0313823298, -0.0002444135, -0.021236835, -0.0050863149, 0.0458461866, -0.034527909, -0.0271770805, 0.0031926865, 0.0754669011, 0.0202781204, -0.0232528225, 0.0082981726, -0.0212221332, 0.0406701975, -0.0089315558, 0.0307356194, -0.0398689136, -0.0317638628, -0.0282075629, -0.0156782642, -0.0207754001, -0.0307490397, 0.0322595835, -0.0083072521, 0.0089117279, 0.0106616495, -0.0222781599, 0.0553137995, 0.061606463, 0.0532696433, -0.0310325529, -0.0064472319, -0.0322667398, -0.0078242412, -0.0256713685, 0.035822317, -0.0238614827, -0.0051272572, 0.0242600385, 0.0027544834, -0.0289360713, -0.0301052853, -0.1251944304, -0.0110877492, -0.0684926286, 0.0980833396, -0.0006627684, 0.0376045592, 0.0399936996, 0.018071657, 0.0076282988, 0.011138849, -0.019030951, 0.0315917879, -0.0078557227, 0.0305204317, -0.0113922199, 0.0271963011, 0.010768543, 0.0648810193, -0.021385204, -0.0287875608, -0.0435417369, -0.0156445652, 0.0316375941, 0.0136683481, 0.0570890419, -0.0133510539, 0.0152646471, -0.0680920109, -0.0022023045, 0.0094582932, 0.0194349606, 0.028352676, -0.0017784167, -0.0143052414, -0.0773367435, 0.0899097472, 0.0396234728, 0.0210704673, -0.0418829769, -0.0218161959, -0.0414394513, -0.018263191, 0.0482946299, -0.0211968198, -0.0707646981, -0.0179787204, -0.0617100038, -0.0257832315, 0.0307007264, -0.0228679236, 0.0076618679, 0.0060350467, 0.0022275492, 0.0344769657, -0.0164428987, -0.0138578713, -0.0118733598, 0.0118301278, 0.0066035269, -0.0645272285, -0.027899256, -5.820621174e-33, 0.0115501713, 0.0046300502, 0.0075724102, 0.0002020684, 0.0010944019, 0.0118921511, -0.0705951378, 0.0181125216, 0.0196319912, 0.0265095495, -0.0125047518, 0.0013854422, -0.0040022573, -0.0028043049, 0.0154595012, -0.0380485095, 0.0037852828, -0.0164682698, -0.0153617878, 0.0047536017, -0.0681377277, 0.0169006567, 0.0490210839, -0.0585143417, 0.0531467162, 0.0374943949, -0.0050116456, -0.0219606087, -0.0304576457, 0.0720743462, -0.0030662012, 0.0411762521, 0.0460860841, -0.0267426763, 0.0033619057, -0.0155361444, -0.02168715, -0.032202106, 0.0178320967, -0.0229581315, -0.0005361686, -0.0260038152, 0.0421955995, 0.0039652786, -0.0464433208, 0.0124372886, 0.0225448422, -0.0490762144, 0.0063230665, 0.0048472136, -0.021338569, 0.0274428148, -0.0438211262, -0.023503812, -0.0026953544, 0.0582641549, -0.0181511007, 0.0356329083, -0.0133104827, 0.0239277612, 0.0548642986, -0.0259485506, 0.0250193365, -0.0012587289, 0.0035307196, 0.0324470624, 0.0132546704, 0.0109170359, -0.0130642941, -0.002840274, 0.0086344127, 0.0498254523, 0.0836992487, 0.0226289686, 0.0235914364, 0.0013618227, -0.0898219571, 0.0507240109, -0.0288616251, 0.0028742375, 0.0290763266, -0.0163115729, -0.0204759724, -0.0300284158, -0.0538939647, -0.0248083491, -0.0490957275, -0.0263212305, 0.0452760793, -0.0207088962, -0.0062280386, -0.0313948095, 0.0726858899, -0.00646429, 0.0061049866, -0.0170880016, 0.0144208809, 0.0268941652, 0.0165449549, 0.0317944475, -0.0372809507, 0.0004072503, 0.0486673042, 0.0109065715, 0.0823583901, -0.0112431981, 0.0206915699, -0.0301338844, -0.0445904396, 0.0169636458, -0.0193724148, 0.0175920576, 0.0230549704, 0.0098173264, 0.0230437834, 0.0288995728, 0.008118636, 0.0702157766, 0.0068356972, -0.032304395, -0.046568878, -0.0149562489, -0.0084647452, 0.0158194918, -0.0119367307, 0.0160653275, -0.0110353017, 0.0403106064, 0.0499127284, -0.0284722131, -0.0265627708, -0.0185811911, 2.905e-7, 0.0890007615, 0.0616091602, 0.0047758264, -0.0196721926, 0.0349562578, 0.0008644848, -0.0223867074, 0.0131485285, 0.0439569205, 0.0200796481, 0.0714659914, -0.0005165625, 0.0571873859, 0.0094717881, -0.0191356894, -0.0362533592, -0.0678237006, -0.0112518147, -0.0610965267, -0.0053683841, 0.0736560076, 0.0303545948, -0.0498163179, -0.034876518, -0.0053779203, 0.0171405189, -0.014794698, -0.0751244277, -0.0154549265, 0.0580585524, 0.0060740192, -0.0210947096, 0.0364276804, 0.084401153, -0.0158520173, -0.01744074, -0.0131283691, 0.075608246, -0.0364637822, 0.0071738833, 0.0433897488, -0.0026269443, 0.0240555517, -0.0594065711, 0.0395386405, -0.0175944455, -0.0403839834, -0.0094487378, -0.0236129742, -0.0213909447, 0.1411234885, -0.0651970655, 0.0181204192, 0.0233953446, 0.0302589927, 0.0385865346, -0.0156196728, 0.0250628814, -0.0103597194, 0.056076169, -0.0531705432, -0.0224675201, -0.0136587182, -0.0198018998, -0.0296953954, -0.0481719561, 0.040717613, 2.906223586e-34, 0.0561637357, -0.021447897, 0.056331981, 0.0171795003, 0.0594046079, -0.0118416296, 0.0143289035, 0.0260344762, -0.0143203633, -0.0827704817, 0.0127318334 ]
http://arxiv.org/abs/2406.17876v1
http://arxiv.org/pdf/2406.17876v1
ET tu, CLIP? Addressing Common Object Errors for Unseen Environments
2024-06-25T18:35:13
2024-06-25T18:35:13
Ye Won Byun, Cathy Jiao, Shahriar Noroozizadeh, Jimin Sun, Rosa Vitiello
cs.CV
cs.CV, cs.AI, cs.CL, cs.LG, cs.RO
We introduce a simple method that employs pre-trained CLIP encoders to enhance model generalization in the ALFRED task. In contrast to previous literature where CLIP replaces the visual encoder, we suggest using CLIP as an additional module through an auxiliary object detection objective. We validate our method on the recently proposed Episodic Transformer architecture and demonstrate that incorporating CLIP improves task performance on the unseen validation set. Additionally, our analysis results support that CLIP especially helps with leveraging object descriptions, detecting small objects, and interpreting rare words.
ET tu, CLIP? Addressing Common Object Errors for Unseen Environments: We introduce a simple method that employs pre-trained CLIP encoders to enhance model generalization in the ALFRED task. In contrast to previous literature where CLIP replaces the visual encoder, we suggest using CLIP as an additional module through an auxiliary object detection objective. We validate our method on the recently proposed Episodic Transformer architecture and demonstrate that incorporating CLIP improves task performance on the unseen validation set. Additionally, our analysis results support that CLIP especially helps with leveraging object descriptions, detecting small objects, and interpreting rare words.
[ -0.0058590905, 0.0476428233, 0.0302292053, 0.0426887199, 0.0322624147, 0.0193085782, 0.0040638493, 0.0016593103, -0.0820650086, -0.0119090481, -0.0096782949, -0.0144342547, -0.0140032284, 0.0565626211, 0.0352806859, -0.0633809417, 0.0085049588, -0.0210157838, 0.0001233451, 0.0106850239, 0.0009321324, -0.0275526438, -0.0056614713, 0.0338738561, -0.0252960417, 0.0080806799, -0.0569069646, 0.0081202853, 0.0450532623, -0.0182321817, 0.0072900681, 0.0079391692, 0.0105254045, 0.0236811861, 0.000001856, -0.0130678965, 0.0088431491, 0.007912566, -0.0175491832, -0.0663062185, -0.0119431373, -0.0617702454, 0.0075719818, 0.0061128396, 0.0030647714, -0.0523499437, 0.0553189665, 0.0779811814, 0.0334511809, 0.0650432706, -0.030992901, -0.0010677454, 0.0060898247, -0.0288360976, 0.0465875342, -0.0220884923, 0.0251954962, -0.0503251776, -0.0425805934, -0.0023273213, -0.0242542084, 0.0046989848, 0.0599753968, 0.0135478536, 0.0490401275, 0.0358971171, -0.0653231293, -0.0491094925, -0.0229511037, 0.0265581217, -0.0528272204, 0.0234084204, -0.0352840796, -0.0146045042, 0.0265280567, -0.0251830462, -0.0237568021, 0.042203296, -0.0068549733, -0.0015867921, -0.0678268373, 0.0253258664, 0.006578614, -0.0223791339, -0.0499863625, -0.0152908796, 0.0196616407, -0.0135685038, 0.0294908602, 0.0634635687, -0.0476477593, -0.0472241007, 0.0496833511, 0.0164855085, -0.0112444069, -0.0305166058, -0.0596553683, -0.0035807127, -0.0210541524, -0.0131341079, 0.0617155321, -0.0079298913, -0.0709771439, 0.0457268842, -0.0526038222, 0.0636739731, -0.0222746879, 0.0294911396, -0.0052840882, 0.0170777924, 0.0101403352, 0.0110764215, -0.0497486405, 0.0210584421, -0.0303634498, -0.0294465199, 0.0191057585, -0.0423060767, -0.0352142826, 0.0263810214, 0.0259345397, 0.0330663137, -0.0568426214, -0.0226433948, 0.0077684685, -0.0172749851, -0.0274503436, 0.0416720994, 0.0022503219, -0.051590234, -0.0286099371, 0.0138182985, 0.0303374436, -0.0127787786, -0.0030728402, 0.0839444771, -0.0021195333, -0.0866383314, 0.0015845906, 0.0433670022, -0.0251635201, -0.0045129503, 0.0578058474, -0.0191474482, -0.0097709736, -0.0309429839, 0.0191348381, -0.0252351891, 0.0124573028, -0.002859534, 0.0042927745, 0.0612508431, -0.0333042145, -0.0260289777, -0.0110600442, 0.0261777546, -0.0207421817, 0.0522334017, -0.0197465215, 0.0402527861, -0.0069753276, 0.062190555, 0.0709079579, 0.034289103, 0.0008031628, 0.0083519313, -0.0390454903, -0.0241714288, -0.0431800261, 0.0081605101, 0.0621960759, 0.0368131511, -0.0308959931, -0.038177494, 0.0261618774, -0.0021082887, -0.0223617237, 0.0078949677, 0.0529969931, 0.0048593609, -0.0116330748, 0.0167888068, -0.0246312376, 0.0199972074, -0.0362256616, -0.0215242784, -0.0315488316, 0.002290946, -0.0581254214, -0.0133334501, -0.0031781935, -0.0252492055, -0.0218906719, -0.0098445425, 0.0755767599, 0.0364586636, -0.0411690734, 0.0172891356, -0.0378673673, -0.0254703667, -0.0199723281, -0.0510127209, 0.0208895896, 0.0031187946, 0.0070196833, 0.0196302421, -0.0278711189, 0.0021681304, -0.0113205109, 0.0580071062, -0.0058054668, -0.0230763294, 0.0733067319, -0.020107612, -0.0576668121, -0.0345538929, 0.0163595323, 0.0096925683, 0.052583281, 0.0319007002, 0.0261359569, -0.0435989052, -0.0004420237, 0.0446521305, 0.0013387832, -0.0274307709, -0.0332363509, 0.0298142526, -0.0031031407, -0.0064470223, -0.0259586424, -0.0188950412, 0.0181481466, -0.032627061, -0.0048465915, 0.00660322, -0.0538403019, -0.0489209741, 0.0034216112, -0.0950774178, -0.0113066612, 0.0617239326, -0.0192071106, -0.0362348109, -0.0215407945, -0.0325594731, 0.066657044, -0.0121923974, 0.0142734433, -0.0128723327, 0.0322881304, 0.0277617648, 0.0285999868, -0.0183655955, -0.0153964153, 0.010746046, -0.0116838217, -0.0081713134, -0.0314738452, -0.0515999757, -0.0226829778, 0.0026091922, -0.0133998552, 0.036886476, 0.0408568308, -0.0226585288, 0.0182930175, -0.0532135293, -0.0099452296, -0.0272051841, -0.0292480253, -0.0036541356, -0.0151001392, -0.0122386953, -0.03893473, 0.0191583075, 0.0098241344, -0.0157290623, -0.0002315576, -0.018660102, 0.0099901948, 0.0097293211, -0.004730464, -0.0185485967, 0.0549677275, 0.0450803638, -0.0011790167, -0.0502041131, -0.0322898999, 0.0617644414, 0.0073161004, 0.0144267306, -0.0018966555, 0.0368482023, 0.0240362696, -0.034630619, 0.0289361477, 0.0646704212, -0.0174001716, 0.0437927879, -0.036571987, -0.003281673, -0.0269860271, -0.0128203966, 0.0231296737, -0.0208692718, 0.0037252065, 0.0207303967, -0.0029017152, 0.0256907605, -0.059019018, 0.0017888409, -0.0490911156, 0.0159839503, -0.0181860477, -0.0829573199, -0.0888499618, -0.0274018105, 0.0019773184, 0.0006411647, -0.0564638339, -0.0048189363, -0.0145262135, -0.0140391914, -0.0073359748, 0.0170049947, 0.0255414471, 0.0166350082, 0.0325481892, 0.0460107997, 0.0044049369, 0.0163091179, -0.0538687631, -0.0267264713, 0.0145213483, -0.0436406881, -0.0034280287, 0.0236875731, -0.0278039966, -0.0090632271, -0.0184907783, -0.0151433609, -0.0355259553, 0.0048961155, -0.023444701, 0.0085447989, -0.0512207486, -0.0118332719, -0.0622676834, 0.0187476203, -0.0159549955, 0.0819792971, -0.028820049, -0.0399805717, -0.014622476, 0.0419653468, 0.0318071693, 0.0021442131, 0.0477197394, -0.1030852273, 0.0186024047, 0.0495556369, -0.0148153901, -0.0064073978, -0.0834607258, 0.0357939079, 0.0186941978, 0.0212732069, -0.0035931899, 0.0119337756, 0.0650519729, -0.0496351235, -0.0280677136, 0.0356311016, -0.0013052524, -0.0877301395, -0.0333459079, -0.0180768222, 0.003226537, 0.0442554727, 0.0183306281, -0.0245207585, -0.0335249305, 0.0434935763, -0.0257158522, 0.0312873945, 0.0057078972, 0.0115036331, -0.0776394531, 0.0506483614, -0.0110655138, -0.0655328929, -0.0543326922, 0.0368545465, 0.051006183, -0.0443152376, 0.0011979573, -0.0004144878, -0.0163603183, -0.0196988713, 0.0020161683, 0.0389189906, -0.0622370355, -0.0433346145, -0.0375589989, -0.0268090107, 0.0315473191, -0.0018599859, -0.0177902989, -0.0302644782, 0.0073871845, 0.0270627718, 0.0129329618, 0.0343588851, 0.0713569447, -0.032856483, -0.0094005894, 0.0277958531, 0.0096129188, -0.0175246932, -0.0512281694, 0.0047309822, 0.0170686059, -0.0040037408, -0.0293038506, -0.0510948263, -0.0245994665, 0.0369719788, 0.0415150672, 0.0301277991, 0.012784563, -0.0160140097, 0.023425065, 0.0135611184, -0.0445655845, -0.0156376753, 0.0031170247, 0.0217738152, -0.044517044, -0.0213429686, -0.0316404589, 0.0734134614, 0.0756184533, 0.0055776048, -0.111306712, -0.0213394929, 0.0101593388, -0.0488208607, 0.0394898355, -0.0314742178, 0.0003378661, -0.0180479549, 0.0420456901, 0.0158523265, 0.003426742, -0.0165355783, 0.0192644931, -0.0529552288, 0.0082837148, 0.0175167955, 0.0778193399, 0.0219456162, 0.0463961884, -0.0235192236, -0.0109635144, 0.1054010689, -0.0066004209, 0.0017995065, 0.0377751142, 0.0579068437, -0.0978184566, -0.0245548747, -0.0035367014, -0.0388954878, 0.0412898138, -0.0065182238, -0.0020230804, -0.0112093175, 0.0149078239, 0.0159840174, -0.0801366419, 0.0457225367, -0.0189695936, 0.0568469577, -0.0424033925, 0.0300642867, 0.041974403, 0.0397636294, -0.0528090149, 0.0595798232, -0.0043418128, 0.0307917893, -0.0007000138, 0.0619124547, 0.0045667738, -0.0176754128, -0.0670015588, -0.0368333273, 0.0199399982, -0.013646639, -0.0069349562, 0.0318190977, 0.0040634349, 0.0139202038, -0.0039083026, 0.0492489077, 0.0044139223, 0.01705206, 0.0212405045, 0.010721378, -0.0070471321, 0.0153060649, 0.0081552994, -0.0154556893, -0.09803316, 0.0710370019, 0.026124863, -0.0481539406, 0.0345465727, -0.0145326406, -0.0274495129, 0.0034205229, -0.0228923205, 0.0086614061, -0.0568254404, 0.0125854556, 0.0053818491, 0.0192832816, -0.0339182913, 0.0292629525, -0.007492872, 0.0266603995, -0.0267985668, 0.0169364363, -0.0386014134, -0.0078611886, 0.07202553, -0.0274994876, 0.0149928015, 0.004182233, 0.0441332571, -0.0448608212, -0.0282822307, 0.0247805938, -0.0052052764, 0.0310321394, 0.0452881195, 0.0280292667, -0.0028674752, -0.0218057632, 0.029950792, -0.0370138437, 0.0147931594, 0.0407404676, 0.0227607097, -5.687025506e-33, -0.0575507805, -0.0004539683, 0.0240582265, -0.0089885741, 0.0006734058, 0.0360801965, -0.023138864, 0.0143350344, 0.0305931345, 0.043151509, -0.011708119, 0.0627011657, 0.0280142892, 0.0161307622, 0.0296103004, 0.0026872046, 0.0627245381, 0.0062583247, -0.0206860881, -0.0192591213, -0.0052289865, 0.0520241223, 0.0905458778, -0.018169608, -0.0084009189, 0.0426570103, -0.0330250859, 0.0068645314, 0.0197343808, 0.0027790042, -0.0169251841, -0.0318986662, -0.0121774729, -0.059323132, -0.0551581308, 0.0211687256, -0.0680046901, -0.0163856689, -0.0360217653, 0.0510112718, -0.0742491856, -0.0333923101, 0.0461363681, 0.0129175531, 0.0291972887, -0.0059876512, 0.0342345051, -0.0354707018, 0.0173581764, -0.018409254, -0.0471647047, 0.0220119338, -0.0025996324, -0.0148859778, 0.058975473, 0.0615088753, 0.010535419, -0.0487548485, -0.0309742521, 0.007977766, 0.0394008085, 0.002654732, 0.0071486994, 0.0497720055, -0.0325419828, 0.0945792198, 0.0064183576, -0.0613531843, 0.0011150659, -0.0286405478, 0.0131338062, 0.039046362, 0.0068949512, 0.0084696747, 0.0704962835, 0.0388560183, -0.0032450375, 0.0438793786, -0.0944678858, 0.0837186798, 0.0422511101, -0.0026906389, 0.0197724663, 0.0067578265, 0.0137579292, -0.0105224978, -0.0253769495, -0.0131859677, 0.0379199907, 0.0704567805, -0.0436133929, 0.0582429655, 0.0430394113, -0.0128251286, 0.0573973656, -0.0066732508, -0.0044151889, 0.0345669612, 0.0003675766, 0.0008059255, -0.0314287618, -0.0319296233, 0.0603199154, 0.0516064987, 0.0151501112, 0.0095739961, 0.0724225119, 0.0688321441, -0.0230987519, -0.0226523578, -0.0384717956, 0.0291059911, 0.0110370573, 0.055866491, 0.0448998511, -0.0032474166, -0.0321233012, 0.0414417274, -0.055531811, -0.0032734659, -0.0522191823, 0.0208328534, 0.029938845, 0.0523220673, -0.0281480812, -0.024568554, -0.0431197174, -0.023776399, 0.0296762437, -0.0793264732, -0.0229089931, 0.011770485, 2.595e-7, -0.0139926132, -0.020184841, 0.0694055706, -0.0121351723, -0.0048574181, -0.0037279588, -0.015820846, 0.0446804799, -0.0208269171, -0.0042553302, 0.0493605137, 0.0603350215, 0.0192538444, 0.0205217749, -0.0469459221, -0.0162255336, -0.0453689173, 0.0042482698, -0.0385795645, -0.0411135852, 0.020973647, 0.0997318253, -0.0134474989, -0.0248124264, 0.0346751139, 0.0110245468, 0.0674107969, -0.0006443894, 0.0481313169, -0.021368457, 0.0056747217, 0.1082251072, 0.0218989849, 0.0296167787, 0.0050981785, 0.019910641, 0.068611823, 0.0255072284, -0.0258098617, 0.0025935741, -0.0001835596, 0.086776562, 0.0053484426, -0.0099604791, 0.0615141802, -0.0954507291, -0.0001007949, -0.0359956324, -0.0569212399, 0.0048223282, -0.0183535665, -0.0211498328, 0.0258202497, -0.0041183378, 0.032482069, 0.0043481328, -0.0502831452, -0.0639595166, 0.0236049667, 0.0303100813, -0.0077085863, -0.0119372457, -0.0480992012, -0.0587834753, -0.0191954896, -0.036603, -0.0739639774, 1.535977372e-34, 0.0070965299, 0.0293727536, -0.0134004354, 0.0004145871, -0.0021843766, -0.032391984, 0.0027019351, -0.0362819508, -0.0321221761, 0.0145933889, -0.0454046577 ]
http://arxiv.org/abs/2406.17875v1
http://arxiv.org/pdf/2406.17875v1
Cloaked Classifiers: Pseudonymization Strategies on Sensitive Classification Tasks
2024-06-25T18:30:25
2024-06-25T18:30:25
Arij Riabi, Menel Mahamdi, Virginie Mouilleron, Djamé Seddah
cs.CL
cs.CL
Protecting privacy is essential when sharing data, particularly in the case of an online radicalization dataset that may contain personal information. In this paper, we explore the balance between preserving data usefulness and ensuring robust privacy safeguards, since regulations like the European GDPR shape how personal information must be handled. We share our method for manually pseudonymizing a multilingual radicalization dataset, ensuring performance comparable to the original data. Furthermore, we highlight the importance of establishing comprehensive guidelines for processing sensitive NLP data by sharing our complete pseudonymization process, our guidelines, the challenges we encountered as well as the resulting dataset.
Cloaked Classifiers: Pseudonymization Strategies on Sensitive Classification Tasks: Protecting privacy is essential when sharing data, particularly in the case of an online radicalization dataset that may contain personal information. In this paper, we explore the balance between preserving data usefulness and ensuring robust privacy safeguards, since regulations like the European GDPR shape how personal information must be handled. We share our method for manually pseudonymizing a multilingual radicalization dataset, ensuring performance comparable to the original data. Furthermore, we highlight the importance of establishing comprehensive guidelines for processing sensitive NLP data by sharing our complete pseudonymization process, our guidelines, the challenges we encountered as well as the resulting dataset.
[ 0.02866368, 0.0843037292, -0.0416712053, 0.0251110867, -0.0303871073, 0.0110882334, 0.1178310886, 0.007570155, -0.0022109796, 0.0027941703, 0.0077896025, -0.002199332, -0.0090569546, -0.0005184836, 0.0124301165, -0.0414646529, 0.0256094504, -0.0175567381, -0.0316865481, -0.0526924953, -0.0174158514, 0.0143422717, 0.007861767, 0.0444941707, 0.0150796799, -0.015080546, 0.0113267228, -0.0488065779, 0.0377020352, 0.0279430728, 0.0077438094, 0.0369535498, 0.0093374643, 0.0782423466, 0.00000149, -0.0294723064, -0.0291775819, 0.0043807942, -0.021119291, -0.0483114533, 0.0406262055, -0.0383507349, 0.025858311, 0.0079258587, -0.0677996352, 0.0081373155, 0.0189645365, 0.0531776957, 0.0558721013, 0.051129289, -0.0037703742, -0.0540777408, 0.0036945322, 0.0196278468, 0.0131264376, 0.0373435207, 0.0172985867, -0.0748507753, -0.082316719, 0.0068248203, 0.0279814526, 0.0528869331, -0.008635981, 0.019729726, -0.0303486474, 0.005417672, -0.0223093592, -0.0091576437, 0.0402845964, 0.0244445112, -0.0142904054, -0.0276264735, 0.0328151844, 0.0440795161, 0.0232459754, 0.0065611312, 0.0043446324, 0.0219720956, -0.0077710124, 0.0038344031, 0.0118336463, 0.0242368057, 0.0409043729, 0.0244282223, 0.0221147891, 0.0072505083, -0.0342852622, 0.0210528746, -0.0322686285, -0.0534581728, -0.0289234426, -0.0219773855, 0.04225583, -0.0063612866, 0.0151515324, -0.0025512748, -0.0914601237, -0.0802786797, 0.0158971269, -0.0821173191, 0.0168129541, 0.0128929531, 0.0429932363, 0.0046928348, -0.0378267728, 0.003438924, 0.0104714399, 0.0515507013, -0.006725607, -0.0025775905, -0.0570002422, -0.0043923794, -0.010679529, 0.0586642586, -0.0327067599, 0.0262618456, -0.0711531192, 0.0044082161, -0.0708789378, 0.0081352452, -0.0345160253, 0.0079235323, 0.0026460765, 0.0104924804, 0.0169007964, -0.0425488278, 0.0030111209, 0.0315952338, -0.0152374972, -0.0366511941, 0.0261703786, 0.003133951, -0.0106842481, -0.0011893918, 0.0088110194, 0.079012461, 0.0321188048, 0.0061667864, -0.0289559495, -0.0425272398, 0.0381068215, 0.0072139916, -0.0059902845, -0.022840064, 0.0551721379, -0.0426112972, 0.0048238346, -0.0596017949, -0.0006157864, 0.0113485307, 0.0146801416, 0.1195304766, 0.0085085034, -0.0152079612, 0.0180084482, 0.0126059204, -0.0042503346, 0.0133057935, 0.0210203677, 0.018569693, 0.0463668518, -0.0280226655, 0.0527287833, -0.0158993304, 0.0243805367, -0.0162458327, 0.0037245313, 0.0041758278, -0.009256267, -0.0320812277, 0.0054589035, 0.0121294064, -0.0117726428, -0.0799925923, 0.0276446063, 0.0422087312, -0.0080422871, 0.0200882964, -0.0304380041, 0.0384652354, -0.0468209721, -0.0322402194, -0.0072364882, -0.0171400364, -0.0848953798, -0.0473308936, -0.053307876, 0.0482108295, -0.0221269373, -0.036004629, -0.0019174726, -0.0276557431, 0.0667206943, 0.0251846276, -0.0193491895, 0.0353356749, 0.0351335667, 0.014673668, 0.0138228694, -0.0294749737, 0.019107867, -0.0544701554, -0.042128779, 0.0464717336, -0.010880312, -0.0150540015, 0.0220663194, -0.0836320966, 0.0283352919, 0.0033037337, -0.0503402986, -0.0155246435, 0.0514366552, -0.0187758859, -0.0353564918, 0.0172279812, -0.0053077168, 0.0255994387, 0.0480984077, -0.0203644708, 0.0036998992, -0.0526934005, -0.0269787591, 0.0205982756, -0.0167414695, -0.0613086559, 0.0467581563, 0.0330003463, 0.0500947423, 0.032616172, 0.0025147884, -0.008304094, 0.0080857417, -0.0086806128, -0.0420622267, -0.0227196552, 0.0805999637, 0.0163148046, -0.0208224654, -0.0440624319, 0.0141025614, 0.0606331863, -0.0178765971, 0.0094338823, 0.0295233522, 0.0022946449, -0.0237835981, -0.00190163, 0.0071486235, -0.0281858891, 0.0826635361, -0.0121785831, 0.0198153928, -0.0351847634, 0.039507024, -0.0453861915, -0.0526359417, 0.0329819731, 0.0001658781, -0.0038245586, -0.0216688327, -0.0286654737, 0.0352897607, -0.0023749636, -0.0275542419, 0.0153922122, -0.0357207693, 0.0014827489, -0.0110070342, -0.0094786175, 0.0074626049, -0.0136041753, 0.0016465652, -0.0061124698, -0.0150822848, -0.0189040154, -0.0382615142, -0.0417176001, 0.042671524, 0.0179805402, -0.0152983069, -0.0325961262, 0.0571581535, -0.0104653584, 0.0071135806, 0.0536271371, -0.0187747944, -0.0707048625, 0.0122015746, -0.0261210445, -0.0387447625, 0.0296755433, -0.0419147015, -0.0144066606, 0.0177575331, 0.0299310759, -0.0286737215, 0.023058299, 0.0310674608, 0.0386614464, 0.0485453568, 0.0630347356, -0.0345118269, -0.0312618464, -0.0235757623, -0.0048196823, -0.0326586999, -0.0464059971, -0.0015358925, -0.0445298553, 0.0349981487, -0.0457608849, -0.0260873009, 0.0226078499, -0.0368602127, -0.0552544892, -0.0510871671, -0.0035037962, 0.0058268798, 0.0047368808, -0.0005536085, -0.0256014187, 0.0622024685, 0.0102600399, 0.039701324, 0.0151173612, -0.0480328016, -0.0014546844, 0.0298904777, 0.0054738089, 0.0373868309, 0.0490513444, -0.0412779897, 0.0370820127, -0.001490807, 0.0211679954, -0.0397659242, -0.0454191081, 0.0303948764, -0.0035631557, 0.012618538, -0.0186859015, -0.0705191344, -0.0673143268, 0.0243049692, -0.0048599872, -0.0048987721, -0.0469163135, -0.0386296734, -0.0067768749, 0.0633766204, 0.0150818797, 0.0237761531, -0.0066995132, -0.009770331, -0.0342244692, -0.0456799008, -0.0545615889, 0.0326414816, -0.0015589303, 0.003150278, 0.0605944917, 0.046144221, -0.0151691567, 0.0147857331, -0.0115905851, 0.029743975, -0.042527914, -0.0049516768, -0.0251864567, -0.0025804392, -0.001396552, 0.0494553149, 0.0515254065, 0.0057809013, -0.0187510867, 0.0123331985, 0.0624212995, -0.0134315221, 0.0190500375, -0.0173614342, -0.0940780044, 0.0388188809, 0.0192901548, 0.0768906176, -0.109800227, -0.063125439, -0.009111248, -0.0168103166, -0.007165262, -0.0114434827, -0.0533402301, -0.0113127604, -0.0685254782, -0.0212339554, 0.0092056319, 0.0232964605, -0.0742084458, 0.0122565916, -0.0075782686, 0.0126949726, -0.0479869209, -0.028921999, -0.0304794945, -0.0137309404, 0.0405111164, -0.0291281696, 0.0344149917, -0.0109197386, 0.0033906414, 0.0029945036, 0.001580436, -0.0209912546, -0.0300175156, 0.1108132526, 0.0040682731, -0.0285936501, -0.0142414514, 0.0519395061, -0.0215679072, -0.0165654775, 0.015705701, -0.005513628, -0.0190999676, -0.0111120949, 0.0170443859, -0.0054988996, 0.0422020741, 0.0380749181, -0.02254772, -0.0162556712, 0.0376773141, -0.0460581109, 0.0177049972, -0.0430509523, -0.0472364202, 0.0130936839, -0.0104338266, -0.0652060285, -0.0012302793, 0.0019899467, 0.0016600916, 0.0326735526, 0.0333106928, 0.0285102818, -0.0058930567, -0.0050210487, -0.0643273667, 0.0039722384, -0.0054431045, 0.0358877964, -0.0045546736, -0.026242638, 0.003435123, 0.0077825435, -0.0089057237, 0.0166779868, -0.0414439738, -0.0634801686, 0.0605707131, -0.0102689657, 0.0185198635, 0.039372392, -0.0308333915, -0.0677413121, 0.1039874032, 0.0556978323, 0.0200239886, -0.0105245952, 0.0555392168, 0.0084991604, -0.0151715754, 0.023759244, 0.0347548947, 0.0007835938, 0.0561480559, -0.076089859, -0.0981066972, 0.0307384022, -0.0115706502, 0.0102186622, 0.0299340989, -0.0110600162, -0.0183964334, -0.0082067885, -0.0026430462, -0.0289333779, -0.0241702981, 0.0398320146, -0.0266213007, 0.0604222044, 0.0247151162, 0.0255642384, 0.0289542358, -0.0134573961, -0.0164180323, -0.0530795306, 0.0115966173, 0.0118695386, 0.0368993171, -0.006796353, 0.0298406016, -0.008313504, 0.067901969, 0.0329015292, 0.0188835468, -0.0076532029, -0.0499305949, 0.0165531803, 0.0255842097, -0.0201235749, 0.0316445604, -0.0477119721, -0.058005292, -0.0262460336, 0.0267955549, -0.0364169292, 0.032136552, -0.0141011877, -0.0055076769, -0.0149556203, 0.0050891223, 0.0089726197, -0.0334716588, -0.0475268289, -0.0082975375, -0.0078744916, -0.0301968101, 0.0596187972, 0.0488613769, 0.0109203476, 0.0209410209, 0.0248473641, -0.0388918407, -0.0254345816, -0.010668071, -0.0065317042, 0.0524678119, -0.0174681284, 0.0137487464, -0.0062407125, -0.0485519916, 0.0182738546, -0.0317317247, 0.0193061251, 0.0084753642, -0.0103991749, 0.0135910921, -0.0046340609, -0.0039952914, -0.0959377512, 0.0288439635, 0.013861862, -0.0447683409, 0.023562856, -4.645823671e-33, 0.035001114, -0.0169872157, -0.0282255206, 0.0387013219, 0.0448394679, 0.027963588, 0.0512541644, 0.0140947588, -0.018174028, -0.0063680806, -0.0223217197, -0.0007024615, 0.001699271, 0.0317466296, -0.0422676429, -0.0393149666, 0.0363112241, 0.0026883753, -0.0089392075, -0.0007321718, 0.0946629792, 0.0162640978, 0.0508310013, -0.1160299554, 0.0648704022, 0.0507067405, -0.0501870364, 0.0114909597, -0.0229751375, 0.0460812449, -0.0265388489, 0.0683044121, 0.005799579, 0.0284537021, -0.0036356153, 0.0420344323, -0.0525441431, 0.0051882025, -0.02818406, 0.0452137478, 0.0555101931, -0.0676902905, 0.0281564556, -0.0455848798, -0.036919076, 0.0049515544, 0.0251107216, -0.0213809106, -0.0188626777, -0.0382095203, 0.0092747388, -0.0201335922, -0.0468103923, 0.0286644157, 0.0594156533, 0.027142765, 0.0384325199, -0.0257023852, -0.1110614911, 0.0283253025, -0.0350169279, -0.0338645391, 0.0485190786, 0.0172001291, 0.0279838983, 0.035995651, 0.045136001, 0.0387456305, -0.0036375397, 0.0380832143, 0.0071894769, -0.0007949431, 0.0578450337, 0.0748460889, 0.0024431904, -0.0570871085, -0.0219485797, 0.0115938047, 0.0163836777, -0.0492546149, 0.0500821583, -0.0009278005, -0.0679831579, -0.022718586, -0.0529620461, -0.0649236441, -0.0320264697, 0.0098118205, -0.0296399854, -0.0164284203, -0.068340227, 0.0046474524, 0.0038377666, -0.0189768039, -0.0382765569, 0.0002876416, 0.0670480132, 0.0547350422, 0.0151789812, 0.0039303727, 0.038925495, -0.0216603726, 0.0369272679, -0.0339387432, 0.0571753904, -0.034848623, 0.0788780898, 0.0084314588, -0.0637873635, -0.0179293249, 0.0058785821, 0.0692078322, 0.0377343148, 0.0476307459, 0.01358815, 0.0267655589, -0.0113147022, -0.0306286253, 0.0476528592, 0.0303625613, 0.0101770619, 0.0191221964, -0.0384435579, -0.0094584469, 0.0289355591, -0.0106727071, -0.0050063552, 0.0275307465, 0.0795483962, -0.0313974209, -0.0138306748, -0.0482962951, 2.383e-7, 0.006586967, -0.0090426691, -0.0038544089, -0.0258637294, -0.0198953152, 0.0405578241, 0.0789900795, 0.0603710786, -0.0189356823, -0.0248707756, -0.042175822, 0.0160697252, 0.0279393718, 0.0120083727, -0.0611237586, -0.0860519931, 0.060639333, 0.0197433047, -0.0704536363, 0.0402725562, 0.0021386868, 0.0571089238, 0.0822603405, -0.0150136817, -0.0312350541, -0.0192326214, -0.0123091666, -0.0633616522, -0.0470507517, -0.0048829545, 0.0509572364, 0.0149726607, 0.0382556319, 0.0244344901, -0.0319169462, -0.0067501822, -0.0079666264, 0.0279389042, 0.0303093418, 0.0006747468, -0.0303174257, -0.0677560121, -0.0117059676, -0.0144625222, 0.0403375141, 0.0196858663, -0.0529067032, -0.0024014318, -0.020268986, -0.0109123914, 0.0679592416, -0.0122871269, -0.0086983507, 0.0249867644, 0.0498036817, 0.0074880156, 0.0170173869, -0.0545215681, -0.0058326954, -0.0428278893, -0.0292660054, 0.0090988316, 0.0564833544, 0.0316014253, 0.0244572144, 0.0280039608, 0.0418997332, 2.247809226e-34, -0.0170217529, 0.0193876047, -0.0075546172, -0.0092212744, 0.0386792384, 0.0128901973, 0.0302339774, -0.0019721545, 0.0080683166, -0.0392566174, -0.0029937082 ]
http://arxiv.org/abs/2406.17873v1
http://arxiv.org/pdf/2406.17873v1
Improving Arithmetic Reasoning Ability of Large Language Models through Relation Tuples, Verification and Dynamic Feedback
2024-06-25T18:21:00
2024-06-25T18:21:00
Zhongtao Miao, Kaiyan Zhao, Yoshimasa Tsuruoka
cs.CL
cs.CL, cs.AI
Current representations used in reasoning steps of large language models can mostly be categorized into two main types: (1) natural language, which is difficult to verify; and (2) non-natural language, usually programming code, which is difficult for people who are unfamiliar with coding to read. In this paper, we propose to use a semi-structured form to represent reasoning steps of large language models. Specifically, we use relation tuples, which are not only human-readable but also machine-friendly and easier to verify than natural language. We implement a framework that includes three main components: (1) introducing relation tuples into the reasoning steps of large language models; (2) implementing an automatic verification process of reasoning steps with a local code interpreter based on relation tuples; and (3) integrating a simple and effective dynamic feedback mechanism, which we found helpful for self-improvement of large language models. The experimental results on various arithmetic datasets demonstrate the effectiveness of our method in improving the arithmetic reasoning ability of large language models. The source code is available at https://github.com/gpgg/art.
Improving Arithmetic Reasoning Ability of Large Language Models through Relation Tuples, Verification and Dynamic Feedback: Current representations used in reasoning steps of large language models can mostly be categorized into two main types: (1) natural language, which is difficult to verify; and (2) non-natural language, usually programming code, which is difficult for people who are unfamiliar with coding to read. In this paper, we propose to use a semi-structured form to represent reasoning steps of large language models. Specifically, we use relation tuples, which are not only human-readable but also machine-friendly and easier to verify than natural language. We implement a framework that includes three main components: (1) introducing relation tuples into the reasoning steps of large language models; (2) implementing an automatic verification process of reasoning steps with a local code interpreter based on relation tuples; and (3) integrating a simple and effective dynamic feedback mechanism, which we found helpful for self-improvement of large language models. The experimental results on various arithmetic datasets demonstrate the effectiveness of our method in improving the arithmetic reasoning ability of large language models. The source code is available at https://github.com/gpgg/art.
[ -0.0099420957, 0.0940594673, -0.018350549, 0.0657866523, -0.029823998, 0.0135783283, 0.0059478045, -0.0123023866, 0.0020910779, -0.0503075831, -0.0013968344, 0.0207605921, 0.027004255, 0.0233551227, 0.0229209531, -0.0526956059, 0.0293316767, -0.0704473034, 0.0272787232, -0.0042703566, -0.0098028583, 0.0133560803, -0.020589726, 0.0024449404, 0.00098344, -0.0713911057, 0.0068692574, -0.0533824898, 0.0093613304, -0.0050564744, -0.0396549143, 0.0566132292, 0.0053731552, 0.0308041051, 0.0000021459, -0.0413536094, -0.0330906138, 0.0301221088, -0.0680427924, 0.0401594862, 0.0096551115, -0.0074395002, 0.0004056338, 0.0319125913, -0.0564103983, -0.0115171121, 0.0284209643, -0.0150741767, 0.1012681499, 0.0664177909, -0.0003364614, -0.0674352646, 0.0230751634, 0.0374478027, 0.0142688612, -0.0203711074, 0.0182046201, 0.1094552577, 0.0361986049, 0.0131493574, 0.0050013722, -0.0077721002, 0.0506078675, -0.0394605659, -0.0299693849, -0.0020291195, -0.0171240047, -0.0041389447, 0.0003969081, 0.0275228433, 0.0196878184, -0.0326567218, -0.0163735282, 0.0148084294, -0.0023698746, 0.0200581048, -0.0561947785, 0.034275841, -0.0014789896, 0.0190902688, 0.0129613876, -0.0107972044, 0.0204981137, 0.0016016879, -0.0773796961, 0.0851325765, 0.0042317775, -0.029448539, -0.0410152376, -0.0436658785, -0.0459771864, -0.0518444702, 0.0280546006, 0.0943562165, -0.0278082937, 0.0290452596, -0.0131818131, -0.0159437079, 0.0065158173, -0.0886808261, 0.021032529, 0.0633715466, 0.0182781871, 0.0549148433, -0.0163652357, 0.0764939189, -0.0285291765, 0.0019157411, -0.0079883914, 0.0330255292, -0.0184502415, -0.01806028, -0.0515900142, 0.0270115845, 0.0202766862, 0.0350233503, -0.0396309383, 0.023214601, -0.0340713039, 0.0568479039, 0.0101211108, -0.0594788603, 0.0491641089, -0.0237091593, -0.0190124176, -0.0088167209, -0.0417317636, -0.0315782055, -0.0265607294, -0.0413414426, 0.0064371154, -0.0259276945, -0.0052860747, -0.0270936154, 0.007415581, 0.0708790943, 0.039385058, 0.0103746708, 0.0149759855, -0.0359145328, 0.0427272432, -0.051998049, -0.0018604839, -0.096118547, 0.0072848802, 0.0412420444, 0.0004711763, -0.0474839732, 0.000029303, 0.0217549857, -0.0089774868, 0.0597145632, -0.0206187069, 0.0234994739, -0.0280345362, -0.0105805965, 0.0085163945, 0.0091404207, 0.0770320371, -0.0429639816, 0.0305306036, -0.0208283123, 0.0081850188, 0.0350559987, 0.0127347484, -0.0237265266, -0.0809006095, 0.0261475928, 0.0017594828, -0.0019174856, -0.0339508839, 0.0484144725, -0.0304878242, -0.025914181, -0.024687374, 0.0038613714, 0.0239301082, 0.0450542755, 0.0281580985, 0.0166831091, 0.0366312228, -0.0111220675, -0.0387283564, 0.0206761062, -0.0411820747, 0.0054813619, -0.0010404369, 0.0138194002, 0.0140294144, -0.0795672983, -0.0318076648, 0.0092581399, 0.035053283, 0.0266764946, 0.0579080842, 0.003026204, -0.000808735, 0.0553900972, 0.030147884, -0.0480116121, -0.0398230702, -0.0602919273, 0.0543500334, 0.0241026841, 0.0482323915, -0.064225167, 0.0028169439, -0.0223552566, -0.0171637051, -0.0149813686, 0.0118373735, -0.0237915963, -0.039645277, -0.0298137721, -0.0139754144, 0.0170499943, 0.0117361154, -0.0146511849, 0.0364935733, -0.0142629463, 0.0089022769, -0.0299617574, 0.0181330554, 0.0043319934, -0.0172720458, -0.0086982204, -0.0246721897, 0.020960601, -0.0129171852, 0.0357132554, 0.0020086416, 0.0593486652, 0.0283705443, -0.0023525658, -0.0041461526, -0.0049033025, -0.0146397315, -0.0264008455, 0.0622513667, 0.0188460909, -0.0496941693, 0.0078304503, -0.0383040942, 0.003862448, 0.0563972928, -0.0483558178, 0.0043026307, 0.0071386723, 0.0275290981, -0.0111438464, -0.038151715, 0.0020175888, 0.0723070428, 0.0025166557, 0.0187785327, -0.0012214328, 0.0132007403, 0.030157309, -0.0003377083, 0.0244033951, 0.0251876563, -0.0065393341, 0.0198069662, 0.0256201327, 0.0069468934, 0.0023358979, 0.0093491171, -0.0154872397, -0.0246918183, -0.0112205595, 0.0488294885, 0.0123750195, -0.0046342034, 0.0020667573, -0.0141428532, -0.0174565986, -0.0203637257, 0.0076473784, -0.010731305, 0.0354430191, -0.0182343312, -0.0532227829, -0.005673293, -0.0448066331, -0.0618923344, 0.0757725835, 0.0143611766, 0.0128818713, -0.0468712635, 0.0314045064, -0.0129667828, 0.0028697476, 0.0016665836, -0.0120099997, 0.0138853891, -0.0351205468, 0.0231140982, 0.0286631566, -0.02944538, 0.0348904692, 0.036735028, -0.0040725484, -0.0459405966, 0.0028097297, 0.011969028, -0.0059155063, 0.0080672912, -0.0283797551, -0.0515963994, 0.0855588689, -0.0335772, -0.0153733622, -0.0282962788, 0.0364831872, -0.0039643771, 0.0612907745, -0.0439042002, -0.0020374996, 0.0080741867, 0.0504160747, -0.0001616589, -0.0660745427, 0.0545534603, -0.0039217961, -0.0136962608, -0.051277902, 0.0082960064, -0.0668043196, 0.0987275988, -0.0015830477, 0.0169595275, -0.03027739, -0.0192604624, -0.0081911935, -0.0284664109, 0.0576133616, -0.0425174609, 0.062482737, -0.0268692318, 0.0190415736, 0.0034560701, -0.0208702572, -0.0457202531, 0.0325637348, -0.0565067939, 0.0321567766, -0.0268625468, -0.0164352003, 0.0273447931, 0.0090367561, 0.0350040942, 0.0222028084, -0.0048317728, -0.0154360486, -0.0230404735, -0.0079240063, -0.0139260022, -0.0054172976, 0.007957533, -0.0135832485, 0.0306090508, 0.0087363888, 0.0506700426, 0.0096575581, -0.0598605648, 0.0096649565, 0.0152841927, -0.0364987291, -0.0152102709, 0.0516641177, -0.0188962705, -0.0906685442, -0.0021334549, 0.1117873564, 0.0435141772, -0.0033417821, 0.0369960479, -0.0614653863, -0.0090034837, 0.0508022197, 0.0558036715, -0.0921758935, -0.0023102118, -0.0123587605, -0.0383588262, 0.0253682379, -0.0344570652, -0.000172741, -0.1087220758, 0.0288129747, 0.0389039107, -0.0218467638, -0.0489924811, -0.0388964005, -0.0284090638, 0.0042956346, -0.0359250158, 0.0698548853, -0.0278351791, -0.002329164, 0.0037408066, -0.0198189244, -0.0212981589, 0.0537297316, -0.0112967147, 0.0055720233, 0.0527743027, 0.0252342392, -0.004113799, -0.0311313514, 0.0346702822, 0.0676261336, -0.0201268978, -0.0165798794, -0.0138094593, 0.0170742348, -0.0147515796, -0.0141558051, 0.0140250204, -0.0585220307, -0.0381692499, 0.0062345853, 0.121304065, -0.0102321766, -0.0001295897, -0.0162020698, -0.0153386965, -0.0121492632, 0.0260833278, 0.0075177541, 0.0231563766, -0.034709949, -0.0670409501, -0.0428140014, 0.003623924, -0.0377927385, 0.0106028551, 0.0219921451, -0.0155880991, -0.0279612467, 0.0093500139, 0.0515243076, 0.0232236013, -0.0119328471, -0.0191613436, 0.0807410926, 0.0009626138, -0.0663620755, -0.0044322042, -0.0423134863, 0.054844223, 0.0021238278, 0.0093560005, -0.0048478623, 0.0050846799, -0.029107172, -0.023459509, -0.0543748736, -0.0263188649, -0.031334132, 0.0657971352, 0.0262720417, -0.0085756015, -0.0129495636, -0.0448595285, -0.0274150614, 0.0015675675, -0.0566088669, -0.0338968486, 0.0138736721, 0.0105979238, 0.0095384642, 0.0204732958, -0.0151997693, 0.0055350726, -0.0228229482, 0.0304016825, -0.060323149, 0.0133515168, 0.0533376299, 0.0534953736, -0.0159918908, 0.0373919792, 0.0059140217, -0.0013078303, 0.020013025, -0.0761136413, -0.0446331501, -0.0595538355, 0.0702522472, -0.0189372711, -0.0288887788, 0.002597251, -0.0585766137, 0.0051878248, 0.0010812868, 0.0217802264, 0.0687999874, -0.0012934287, 0.0324115902, 0.046736639, -0.0462198108, 0.0181563143, -0.0067338408, -0.0090193553, 0.0111318538, -0.0087298807, 0.0473740399, 0.0257503204, 0.0140024647, -0.0276586544, -0.0403793231, 0.0189509559, -0.0898103416, 0.0071254419, 0.0172908939, 0.0318037495, 0.0401665308, 0.0129466141, -0.0601922162, -0.0048715323, -0.0117192529, 0.0540152378, 0.0351585932, -0.0371803902, 0.0093992883, 0.0526581779, -0.0011541928, -0.0507022738, 0.0575412922, -0.0361633226, 0.0092179691, 0.0119832773, -0.0309367478, -0.0445787571, -0.0474232249, -0.0249845348, -0.0345938914, -0.0292526912, 0.0158779304, -0.0694619492, -0.1090310812, -0.0588663481, 0.0348309018, 0.0176198799, -0.0041200714, 0.0331691355, 0.0485240668, 0.0012215794, -0.0055861082, -0.0278765261, 0.0327534787, -0.0487136133, -0.0289971605, -0.023196388, -5.372772153e-33, -0.032448329, -0.036032021, 0.0461373776, -0.0334482156, 0.0435741022, -0.0362235457, -0.0470618978, 0.0260576345, 0.0015239333, 0.0231909137, 0.0192858838, 0.0336296484, -0.004469383, -0.0313004516, -0.0052749584, -0.0112737715, 0.0315632671, -0.0065453267, -0.030432811, -0.0184220243, 0.0159480646, -0.0232946221, 0.0649286583, -0.0682595968, 0.0353479795, -0.0274862442, -0.0240320005, 0.0548152551, -0.055076208, 0.0802011415, 0.0111248177, 0.0053859795, -0.0045302836, -0.0131467599, -0.032260865, 0.0123417508, -0.0672326908, -0.0249192789, 0.0260691401, 0.0145502109, -0.0314192139, -0.0256737694, 0.0295602828, 0.0036529691, -0.033459194, -0.0133343572, 0.0155235734, -0.0440392159, 0.0040132068, 0.0127251623, -0.0520099774, 0.0426410995, -0.0371782593, -0.0088081555, 0.0138116656, 0.0133654736, -0.0391982682, -0.0248968061, -0.0064796628, 0.0220084954, 0.0523773767, 0.0138830161, 0.0530025996, -0.0348609798, 0.0134764053, -0.0012997456, -0.0031078602, -0.0110948384, -0.0242879, 0.0172708221, 0.0147183128, 0.051130753, 0.048360493, -0.0010906699, 0.0763414651, 0.0125176487, -0.0568266138, 0.0014033075, 0.0213947557, 0.0565804578, 0.0238138884, 0.0046839146, -0.0527581349, -0.0239771362, -0.0356227383, 0.0375786573, -0.0567641966, -0.046231132, -0.0089207394, -0.0246078856, -0.0451055802, 0.0359880291, -0.0021772967, -0.0579576418, -0.0650911927, 0.0421688855, -0.017246386, 0.0013458682, -0.0143791987, 0.0344747119, 0.0223020185, -0.0151321255, -0.04635204, -0.0531301461, 0.0461919904, 0.058728952, 0.0018436147, 0.0325785503, -0.0559556596, -0.014130515, -0.034632951, -0.0318699777, 0.0169231277, 0.0091855396, 0.0292231143, 0.0179352518, 0.0240598023, 0.0683301017, -0.0025531841, -0.0229368564, -0.021173384, 0.0717357844, -0.0054512164, 0.0286924932, 0.0185174178, -0.0771688595, -0.0547884442, 0.0229625534, 0.0525100641, -0.0146514932, -0.0135743804, 0.0085105393, 2.776e-7, 0.0366674997, 0.0731463805, 0.0432369076, 0.0979199633, 0.022631133, 0.0246897824, -0.0047991774, -0.0108597055, 0.0174239576, -0.0360070802, -0.0075083547, 0.0300335269, 0.0281471759, 0.0365226604, 0.0050697811, 0.0022151263, -0.0218872242, -0.0393370464, -0.05384681, 0.0141593358, 0.0490969792, 0.0436811484, 0.0525032952, 0.0314688496, -0.0025374901, -0.0336408094, 0.028316237, -0.0055014095, 0.0083917864, 0.0147487354, -0.0119493166, -0.0312854536, 0.0169923101, -0.0083732745, 0.0163577665, 0.0275874808, 0.0087714521, 0.0874571279, -0.0085342256, 0.0811576992, -0.0166601166, -0.063348271, -0.0077316253, -0.0324155465, 0.0484223403, -0.0552490503, -0.0493526682, -0.0249434114, 0.0298184678, -0.0376664288, 0.0049097962, 0.0128843617, 0.0090047875, 0.0146821672, 0.020077087, -0.0269544832, -0.0012740793, -0.0421417393, -0.0544915535, 0.0178398658, -0.0333554968, -0.027646713, -0.0142296264, -0.0054911114, -0.0001378662, -0.0255249087, 0.0186309014, 2.886586402e-34, 0.060362462, -0.0072483653, 0.0242708568, 0.0482465364, 0.0773725733, -0.0275059752, -0.0187129434, 0.0589584634, 0.0690758377, -0.0343117192, -0.0247288737 ]
http://arxiv.org/abs/2406.17764v1
http://arxiv.org/pdf/2406.17764v1
BMIKE-53: Investigating Cross-Lingual Knowledge Editing with In-Context Learning
2024-06-25T17:48:56
2024-06-25T17:48:56
Ercong Nie, Bo Shao, Zifeng Ding, Mingyang Wang, Helmut Schmid, Hinrich Schütze
cs.CL
cs.CL, cs.AI
Large language models (LLMs) possess extensive parametric knowledge, but this knowledge is difficult to update with new information because retraining is very expensive and infeasible for closed-source models. Knowledge editing (KE) has emerged as a viable solution for updating the knowledge of LLMs without compromising their overall performance. On-the-fly KE methods, inspired by in-context learning (ICL), have shown great promise and allow LLMs to be treated as black boxes. In the past, KE was primarily employed in English contexts, whereas the potential for cross-lingual KE in current English-centric LLMs has not been fully explored. To foster more research in this direction, we introduce the BMIKE-53 benchmark for evaluating cross-lingual KE on 53 diverse languages across three KE task types. We also propose a gradient-free KE method called Multilingual In-context Knowledge Editing (MIKE) and evaluate it on BMIKE-53. Our evaluation focuses on cross-lingual knowledge transfer in terms of reliability, generality, locality, and portability, offering valuable insights and a framework for future research in cross-lingual KE. Our code and data are publicly accessible via the anonymous repository at https://anonymous.4open.science/r/MIKE.
BMIKE-53: Investigating Cross-Lingual Knowledge Editing with In-Context Learning: Large language models (LLMs) possess extensive parametric knowledge, but this knowledge is difficult to update with new information because retraining is very expensive and infeasible for closed-source models. Knowledge editing (KE) has emerged as a viable solution for updating the knowledge of LLMs without compromising their overall performance. On-the-fly KE methods, inspired by in-context learning (ICL), have shown great promise and allow LLMs to be treated as black boxes. In the past, KE was primarily employed in English contexts, whereas the potential for cross-lingual KE in current English-centric LLMs has not been fully explored. To foster more research in this direction, we introduce the BMIKE-53 benchmark for evaluating cross-lingual KE on 53 diverse languages across three KE task types. We also propose a gradient-free KE method called Multilingual In-context Knowledge Editing (MIKE) and evaluate it on BMIKE-53. Our evaluation focuses on cross-lingual knowledge transfer in terms of reliability, generality, locality, and portability, offering valuable insights and a framework for future research in cross-lingual KE. Our code and data are publicly accessible via the anonymous repository at https://anonymous.4open.science/r/MIKE.
[ 0.0129268169, 0.0136744902, -0.0068948241, 0.073510319, -0.0556188338, -0.0002736509, 0.0008471209, -0.0184002221, -0.0261767264, -0.0463827327, -0.1156246886, -0.0150086572, 0.0343141891, -0.0085986517, 0.0282921921, -0.0518577397, 0.0536368266, -0.0349563435, -0.0435356908, 0.0162140001, -0.0345839448, -0.0001707257, 0.0014453093, 0.0151356263, 0.0122707775, -0.0188727695, -0.0033199743, 0.0270482916, 0.0475719236, 0.0034822747, 0.0349639282, 0.0410274155, -0.006999854, 0.0379902236, 0.0000022439, -0.0545279905, -0.0270663109, 0.0015824198, 0.0134093622, 0.037368156, 0.0127144214, -0.0536686517, -0.0490065515, -0.0233448148, -0.0249580797, 0.020285828, 0.0505827554, 0.0115881702, 0.0559503809, 0.0882433578, -0.0064757089, 0.0202060845, 0.0370704345, -0.0515838712, 0.0123514384, -0.0316405781, 0.05657392, 0.0103545506, -0.0678579137, 0.0021741688, -0.041599229, 0.0680032745, 0.0112993661, -0.0144895064, 0.0125936409, -0.0074425605, -0.028311763, 0.0161853954, -0.0227177404, 0.0730098784, -0.033476159, -0.0049538198, -0.0269216467, -0.0010221655, 0.0117747039, -0.0200352911, -0.0406565182, -0.0098296497, 0.0280042049, 0.0076924548, -0.0081442771, -0.0560397394, 0.0614829846, -0.0026384692, 0.0013316077, 0.0499646403, 0.0360388085, -0.0277769342, 0.0009248179, 0.0097407065, 0.0066240369, -0.0826528817, -0.0219070222, 0.0156049384, -0.0086326487, 0.01095792, -0.0671219453, -0.006830554, -0.0045545129, -0.0600361302, 0.0538394488, 0.0465140715, -0.0431286097, 0.0696598962, -0.0164947398, 0.026390139, -0.0003429692, 0.0187209845, -0.0271582939, -0.0111049786, -0.0490654111, 0.0180919562, -0.0179015994, -0.0089339875, -0.0546402223, 0.0202341676, -0.0605769791, 0.01565044, -0.0280338246, 0.0068358392, 0.0125411116, -0.0423079468, -0.0357550159, -0.013548445, -0.0094823865, 0.0414064825, 0.0179671105, 0.0203105062, -0.0015238214, -0.0295169167, 0.0025675485, -0.0147337643, 0.0010501754, -0.0677988604, 0.0351673923, 0.0668000132, -0.0557731614, -0.0271677841, -0.0629528016, -0.0172167607, -0.0096442224, -0.0773203298, -0.0145549886, -0.0711513236, 0.0408545658, -0.0194574129, -0.0130583281, -0.0578064919, 0.0029530008, -0.007735752, -0.0290448461, 0.0842651948, -0.0013147498, 0.0250400044, 0.0312951915, 0.0224619024, -0.0474391654, 0.0879601315, 0.017082775, -0.0284430571, 0.0790090114, 0.0030285744, -0.0208742227, 0.0582698174, 0.0389529206, 0.0492477901, -0.0201544873, -0.001886268, 0.0469475091, 0.0226853639, 0.0365022272, -0.0031113091, -0.021883402, 0.0078960825, 0.0988916457, 0.0117720859, 0.0177253634, 0.0217169784, 0.0514738299, -0.034456715, 0.0069894306, -0.0245215502, -0.0568379648, 0.0285214558, -0.0192756169, -0.0011321941, 0.013824637, 0.0434775986, -0.0191678237, -0.0292245019, 0.0196791869, -0.0088604456, 0.0127321295, 0.0068678036, 0.0596751496, 0.0381852165, -0.0178640038, 0.0330378115, 0.0480085723, -0.0195699148, 0.0217871647, -0.0088454289, -0.0067242724, 0.0597305857, 0.027256025, -0.0224200226, -0.0416843705, -0.0258301627, 0.0188665725, 0.005719108, 0.0320316292, -0.0301813651, 0.0403114967, -0.0345151238, -0.0205995459, -0.0446359143, 0.0136097381, 0.0080581643, 0.0559160113, 0.0503301173, 0.021328764, -0.0415970646, 0.0172680747, 0.007466984, -0.0717876554, -0.0080043338, 0.0039637201, 0.0137529885, 0.0105167972, -0.0143790077, -0.0104857991, -0.023784779, -0.0120054884, -0.0283588525, 0.0028402703, 0.0286445636, -0.0353062265, -0.0564282499, 0.0524023883, -0.0194200426, -0.0373306125, -0.001458735, -0.0430793017, -0.0217162129, 0.0005267932, 0.0249705426, 0.0538528152, -0.0379202664, 0.0351800993, -0.0042712241, -0.0054909182, 0.0175275449, 0.0303806029, -0.0010453978, -0.0202045497, 0.0346631519, -0.0146720055, -0.0147780972, -0.0099787032, -0.0068072523, 0.0340951942, -0.0194726717, -0.0134948306, -0.0143757053, 0.033308316, -0.0407699682, 0.0156408641, -0.1171920896, 0.0105481744, -0.0191725083, 0.0077509009, 0.0420049727, 0.0487001166, -0.0145526296, -0.0134005845, -0.0162523687, -0.0034687696, -0.0203779228, 0.0458700508, 0.0125589045, 0.0187902171, 0.0251070671, -0.0486699417, 0.0039332551, 0.0310511794, 0.0397251509, -0.0310616475, -0.0205784924, -0.0520436652, 0.0316991918, -0.0705444291, -0.0011586931, -0.0119143724, -0.024211213, -0.0468300097, 0.0018380001, 0.0901577249, 0.0507329293, -0.0402479097, -0.0179536603, 0.0667926818, -0.0612336434, -0.0274333991, 0.0221252944, 0.0289077293, -0.0190799888, -0.0538577288, -0.0187543035, 0.0061414065, 0.0627427697, -0.0181930214, -0.0393716954, -0.0510151237, 0.0008311583, -0.0115822395, -0.0237840805, -0.0891912133, 0.0248494837, -0.005238337, 0.0409953967, -0.0511772968, -0.0364497006, 0.0583652444, -0.0191154871, 0.0121573852, -0.0282093547, -0.0230309516, -0.0163465142, 0.0536242798, -0.0516748577, 0.0361446217, -0.03563191, 0.0222774334, -0.0788481385, -0.0199557506, 0.0373590998, -0.0039277291, 0.0613143221, -0.089791663, 0.0044070883, -0.0076917671, -0.0365119278, -0.0124798249, 0.0165748466, -0.0113482559, 0.0073623955, -0.0031664958, -0.0320812836, -0.0152114099, -0.0259195194, 0.0677707344, 0.0693733096, -0.0003442902, -0.013609115, -0.0021299573, 0.0205838121, 0.0146830017, -0.0008714375, 0.0031157655, -0.0205274373, 0.0113683427, 0.0464839935, 0.012975703, -0.0547165535, 0.0094113424, 0.0068198899, 0.0049830335, -0.0447921343, -0.0161464736, -0.0149221225, 0.0787094906, -0.0616702139, -0.0315915681, 0.0618826672, 0.0070643225, -0.0493284874, 0.0149233583, -0.0455505028, -0.0138494372, 0.0343143493, 0.0373160131, -0.0820666328, -0.0338804312, 0.0249862596, 0.004913378, 0.0182987191, -0.0375966355, -0.01232062, -0.109757483, 0.0186768044, 0.0146713722, -0.0971476287, -0.0525492094, -0.0449287407, 0.0167866256, -0.0081106992, 0.014259018, 0.0187952016, 0.0092662396, -0.0139330532, -0.034076754, -0.0008683249, -0.0933344141, 0.0480528139, -0.0347038992, 0.0313446671, 0.0216919109, 0.0282220393, 0.0122157559, -0.0322187096, 0.0538033321, -0.0042233388, -0.0058145244, -0.0240657069, 0.0275207516, -0.0065350146, -0.0118702063, 0.0435182825, 0.0316716023, -0.0265032053, -0.0401419438, 0.0400554799, 0.1006269231, 0.0163244307, 0.010992582, 0.0370574296, -0.0495768562, 0.032790903, 0.0585995242, 0.0180761609, -0.0319716744, 0.015069169, -0.0360043645, -0.0335135683, -0.0208259448, -0.0022467924, -0.0209816974, -0.067036517, 0.0018961694, -0.0912339389, -0.0453682579, 0.0474555455, 0.0021280635, 0.0351379812, 0.0123925311, -0.0091296965, -0.0075356509, -0.0438754633, 0.0087318439, -0.0551848672, 0.0406106412, -0.0359108001, -0.0169389471, 0.0241935123, -0.0376652628, -0.0053073592, 0.0036949096, -0.051102519, -0.0726489127, -0.0251081493, 0.0838304386, 0.0076966672, 0.0858085528, 0.0015223169, -0.0069580381, 0.059306547, -0.0024663447, -0.0039378814, 0.0585648976, 0.0384121425, 0.0294037536, -0.002312978, 0.0597515479, -0.0667411834, -0.0161086805, 0.052299913, 0.0413268842, -0.0182505716, -0.0126197087, -0.0190213546, -0.0144382063, 0.0039944043, 0.0148147522, -0.0336834192, -0.0390313976, 0.0534804501, -0.0209455118, -0.029604163, -0.0329252258, 0.0001271103, 0.0337412804, -0.0068254243, -0.012345654, 0.0409600325, 0.0148931639, -0.0074057169, -0.0114856819, -0.0093413908, 0.0168270636, 0.0420214385, 0.0418183319, 0.0259223506, -0.0007960905, 0.0200527236, 0.0078560868, 0.0168464072, -0.0398268029, 0.0188268647, 0.0171445645, -0.0239125807, -0.0254565962, 0.0260070525, -0.0433795825, -0.0057489579, -0.0245597549, 0.0194691345, 0.0917146578, 0.0085083591, 0.0625836998, -0.0329486839, -0.0521864444, -0.0322569609, 0.054341998, -0.0054728724, -0.002845177, 0.0180707593, -0.0204491913, -0.0043728538, 0.0516195744, 0.0895608589, -0.0328164846, 0.0967167392, -0.0109207127, -0.0067698844, -0.0406799503, -0.0212753993, 0.031926062, 0.0142983673, -0.0301070996, -0.0144613422, 0.0027002813, -0.0727327988, 0.0149159459, -0.0060252193, -0.0572096929, 0.0296728425, 0.0050602802, 0.0666476637, 0.0121159051, 0.0009812944, -0.0253079887, 0.0048207045, -0.0032725981, -0.0157715883, 0.0191411786, -6.488697225e-33, -0.0078528682, 0.0215491392, 0.0341055989, 0.0131819369, -0.0281029511, -0.0004869425, -0.048104357, -0.0013986892, 0.0131035326, 0.0115895616, -0.0078795552, 0.0014520057, 0.0003327837, 0.0224602129, 0.0132703008, 0.0011554252, 0.0565353669, 0.0327698067, -0.0118931774, 0.0340260155, 0.0335505009, 0.0606387332, 0.1133025885, -0.0340800434, 0.0459763557, 0.0341550037, 0.0025149409, 0.0290071815, -0.0202371497, 0.0308950767, -0.008041936, 0.01902036, 0.0117107155, -0.0143733686, -0.026516743, -0.0327448696, -0.0582902804, -0.0127947088, 0.0069286535, 0.0369148813, -0.0213370938, -0.0713730901, 0.0227054544, -0.0102894185, -0.0018848927, -0.025009403, 0.0161065441, -0.0543414354, -0.0070625064, -0.0240550004, -0.0803412423, 0.0062355492, -0.0062564695, -0.0240454953, 0.0087205106, 0.0283378474, 0.0179832559, -0.027683381, -0.0919314697, 0.0137781082, -0.0136924898, 0.0410737284, 0.0888793543, -0.0091587203, 0.0079722973, 0.0120241409, -0.0007797921, 0.0051070508, 0.0302161407, -0.0112587688, 0.0340163745, 0.0468575284, 0.0173839778, 0.0852610469, 0.010254466, 0.0038743434, -0.0847146809, 0.028537225, -0.0189750139, 0.0246551093, -0.0204791762, 0.0442798026, 0.0367589593, -0.0002872293, -0.0436708741, -0.0196832214, -0.0315844417, -0.0342321433, 0.0282626078, 0.006090309, -0.0122667691, 0.016436331, 0.043145135, -0.0274838619, -0.0143137798, 0.0328019559, -0.0077724257, 0.036825899, 0.0079082279, -0.0155027648, -0.0604646653, 0.0076900525, 0.0072857672, 0.0120673403, 0.0157095138, -0.0146916658, 0.0784153864, 0.0371325798, -0.04885583, 0.0077986545, -0.0300206579, 0.0400294028, 0.0005022796, 0.0309066996, -0.0026576668, 0.0078526838, -0.0248443205, 0.0187478643, 0.0551220775, -0.019511478, 0.0088837501, -0.0061250022, -0.0052042282, 0.018913148, -0.0105438214, 0.0141802225, -0.0303800143, -0.0072092973, 0.041463647, -0.0371499136, -0.0142055964, 0.0215905551, 2.903e-7, 0.0418476984, 0.0355919898, -0.0072656684, -0.0243998747, 0.0243860353, 0.0261847544, -0.0596393794, 0.01715515, 0.0055122995, -0.0002716265, 0.0109399417, 0.0517614782, 0.0632026196, 0.0357839577, -0.0583711714, -0.0061268397, -0.0330330767, 0.0118420888, -0.0446741432, -0.0220911317, -0.0337312669, 0.0385683849, 0.0234077647, -0.0156037295, 0.0015986665, -0.031122921, 0.0883857086, -0.0825952813, -0.0091644712, 0.0057229805, -0.0052543911, 0.0505463518, 0.0069588409, 0.0263513066, -0.0514926426, -0.0097243339, 0.008038193, 0.0798617527, 0.0121105583, 0.0005240453, -0.0096695842, 0.0435068011, -0.0161014833, -0.0720800906, 0.0630304292, -0.0432199165, 0.0204653144, -0.0364681967, -0.0014213343, -0.005485503, 0.047281675, -0.0450178944, -0.0173305608, 0.0265006796, 0.0487569571, 0.0488133505, -0.0665087253, -0.0235324576, -0.0274328124, 0.0210771598, 0.0036278467, -0.0585648529, -0.0489405282, -0.0251591522, 0.0009006417, -0.057055708, 0.0110329445, 2.334821751e-34, -0.0092857024, -0.0023007828, 0.0172630176, 0.0272371043, 0.0182883348, -0.0180827975, -0.0273450688, 0.0008335115, -0.0275747571, -0.0464672148, -0.007240037 ]
http://arxiv.org/abs/2406.17761v2
http://arxiv.org/pdf/2406.17761v2
CaLMQA: Exploring culturally specific long-form question answering across 23 languages
2024-06-25T17:45:26
2024-07-03T16:33:55
Shane Arora, Marzena Karpinska, Hung-Ting Chen, Ipsita Bhattacharjee, Mohit Iyyer, Eunsol Choi
cs.CL
cs.CL, cs.AI, cs.LG
Large language models (LLMs) are used for long-form question answering (LFQA), which requires them to generate paragraph-length answers to complex questions. While LFQA has been well-studied in English, this research has not been extended to other languages. To bridge this gap, we introduce CaLMQA, a collection of 1.5K complex culturally specific questions spanning 23 languages and 51 culturally agnostic questions translated from English into 22 other languages. We define culturally specific questions as those uniquely or more likely to be asked by people from cultures associated with the question's language. We collect naturally-occurring questions from community web forums and hire native speakers to write questions to cover under-resourced, rarely-studied languages such as Fijian and Kirundi. Our dataset contains diverse, complex questions that reflect cultural topics (e.g. traditions, laws, news) and the language usage of native speakers. We automatically evaluate a suite of open- and closed-source models on CaLMQA by detecting incorrect language and token repetitions in answers, and observe that the quality of LLM-generated answers degrades significantly for some low-resource languages. Lastly, we perform human evaluation on a subset of models and languages. Manual evaluation reveals that model performance is significantly worse for culturally specific questions than for culturally agnostic questions. Our findings highlight the need for further research in non-English LFQA and provide an evaluation framework.
CaLMQA: Exploring culturally specific long-form question answering across 23 languages: Large language models (LLMs) are used for long-form question answering (LFQA), which requires them to generate paragraph-length answers to complex questions. While LFQA has been well-studied in English, this research has not been extended to other languages. To bridge this gap, we introduce CaLMQA, a collection of 1.5K complex culturally specific questions spanning 23 languages and 51 culturally agnostic questions translated from English into 22 other languages. We define culturally specific questions as those uniquely or more likely to be asked by people from cultures associated with the question's language. We collect naturally-occurring questions from community web forums and hire native speakers to write questions to cover under-resourced, rarely-studied languages such as Fijian and Kirundi. Our dataset contains diverse, complex questions that reflect cultural topics (e.g. traditions, laws, news) and the language usage of native speakers. We automatically evaluate a suite of open- and closed-source models on CaLMQA by detecting incorrect language and token repetitions in answers, and observe that the quality of LLM-generated answers degrades significantly for some low-resource languages. Lastly, we perform human evaluation on a subset of models and languages. Manual evaluation reveals that model performance is significantly worse for culturally specific questions than for culturally agnostic questions. Our findings highlight the need for further research in non-English LFQA and provide an evaluation framework.
[ 0.0680054724, 0.0485052615, -0.0244154129, 0.0313690491, -0.0218670219, 0.0153665207, 0.0361575261, -0.0090736141, 0.0032182871, -0.0394330658, -0.015077672, -0.0798303261, 0.0132622859, -0.003461343, 0.0227079876, -0.0490896776, 0.0014458514, -0.0313457102, 0.030184757, -0.0137475701, 0.0173514392, 0.021993285, 0.0246106349, 0.0206966549, -0.0046968306, -0.025760863, 0.0084739225, -0.0217556506, 0.0055976328, -0.0157133378, 0.0223739762, 0.0527569875, -0.0382161364, 0.0042216592, 0.000001951, -0.0540453866, 0.0549769737, -0.0217018351, 0.007570467, 0.0829428509, 0.0212359913, 0.0102170752, 0.0245481469, 0.014363097, -0.0280493759, -0.0003962919, 0.025990203, 0.0202390887, 0.0458642729, 0.0321945883, 0.017833909, 0.0062599196, 0.0112963608, -0.0303004831, -0.001233195, -0.0149901239, 0.0245846044, 0.0281165801, 0.0114745516, -0.0111085996, -0.0257357135, 0.0240757838, 0.0485972725, -0.0120803686, -0.0127198528, -0.0201032627, -0.0493804514, -0.0326399505, 0.0434928909, 0.0594396554, 0.0159369223, -0.027047161, 0.0064634965, 0.0244134963, -0.0291075073, 0.016010318, -0.0527282208, 0.0437106639, 0.0016949805, -0.0173923131, 0.0181638505, 0.0115074264, 0.0015196908, 0.0325369798, 0.0295942891, 0.0574787818, 0.0254988354, -0.0040438836, -0.0425218828, -0.0195281599, 0.0014724944, -0.0952441394, 0.0143490313, 0.0514711812, -0.0254463069, 0.0114194006, 0.0026176986, 0.0002610231, 0.0729003549, -0.0541200452, -0.0136098228, 0.0240978766, -0.0326260589, 0.0412068553, -0.0379402675, 0.0590881445, -0.0271980148, -0.0186691657, -0.0428234413, -0.0433537439, -0.0022412837, -0.0600700155, 0.0084020486, -0.0278214794, -0.0078859525, 0.0099165905, -0.0265632309, 0.0027178072, -0.0186531544, -0.0267717838, -0.0193900168, -0.0019624904, -0.0048511983, -0.012794774, -0.0283856224, 0.0605844893, -0.0775983408, 0.0104289763, 0.0012567916, -0.0803330839, 0.0220587794, 0.0147835417, -0.0460508093, -0.025353536, 0.0141335372, 0.0052041993, 0.0012916891, 0.0223082677, -0.0452031121, -0.0204331893, -0.0039041785, -0.0433495231, -0.0214081313, -0.1198404506, 0.0457248464, 0.0190437343, -0.0055763726, -0.0339207426, 0.0344440453, 0.0445746444, -0.0247653816, -0.0007118118, 0.0260636359, 0.0551549941, 0.1107416749, 0.0287636388, 0.0434967205, 0.036991138, 0.0343728252, -0.0028628027, 0.0838601887, -0.020600047, 0.0217065848, 0.0477470979, 0.005032273, 0.0208202358, 0.0086613586, 0.0924832225, 0.0230480302, -0.0157922693, 0.0493924543, 0.0558524281, 0.014578254, -0.0034030816, 0.0130978283, 0.0622491352, -0.059462212, 0.0362561494, 0.0520628653, -0.0122952461, 0.0342772938, 0.0031055866, -0.0229554065, 0.0459400825, 0.0134565039, -0.012348081, 0.0155671667, 0.070835337, -0.0529419146, -0.0436097085, -0.0180133469, -0.0013167199, 0.0163343623, 0.0396959297, 0.0019140891, 0.0098937368, 0.0241999701, 0.0541538261, 0.0265439786, -0.0519655272, 0.0250191577, -0.0263199452, 0.0718368217, 0.0349130034, 0.0041071745, -0.0066807671, 0.0182130542, -0.0275784805, 0.0810784549, 0.0358708017, 0.0479277074, -0.013281038, -0.0086309761, -0.0089750616, 0.0040636831, 0.0216693431, -0.0119972415, -0.0043819542, 0.056490235, 0.048083242, 0.0501096323, -0.0476236679, -0.0338168852, 0.0064473096, -0.0517774373, -0.025138, 0.0117020402, -0.0069989017, 0.0194836352, 0.0294966586, -0.0014387349, 0.0230911989, -0.0027721431, -0.0388243757, -0.0350305699, -0.0143523971, -0.0527940094, -0.002944832, 0.0398724601, -0.0420413353, -0.0151004065, 0.0274892878, -0.0201736297, -0.0351572819, -0.0041704341, -0.0384702235, -0.0171401091, -0.0571447797, -0.0397467427, -0.0008521129, -0.0055527003, -0.0413827263, 0.0600956753, -0.0049902927, 0.0373234227, 0.0338096544, -0.0276357606, 0.0347605124, -0.0380175263, -0.0166399162, 0.0225686077, 0.0002891542, -0.0213840771, 0.0600279607, 0.0210454687, -0.03584411, -0.0168900806, -0.1017870158, -0.0018489714, -0.0173385274, 0.0245835781, 0.0186753478, 0.0402431339, -0.0216132179, -0.0185631737, -0.0992295295, -0.0481578857, 0.0482004285, 0.0220486503, 0.025462117, 0.0026350929, 0.0019169371, -0.0325649828, -0.0376896299, 0.0112518826, 0.0353497714, -0.0009567931, -0.0032823309, -0.041875422, -0.0169396009, -0.0660460964, -0.0415596962, 0.00630495, -0.0713022202, -0.0488849916, 0.0204474442, 0.0448152013, 0.0524202622, 0.0404967144, -0.083080925, 0.0136883911, 0.0172069836, -0.0227277614, -0.011289835, -0.0037891495, -0.0027640115, 0.0093755545, -0.0822702199, 0.0094967848, 0.0405713357, 0.0104326764, -0.0055527142, -0.0110350456, -0.0063718492, -0.0548602082, -0.0242291931, -0.0163561758, -0.0237224028, -0.0055035544, 0.0741937235, -0.0176762473, -0.0190355647, 0.0631868169, -0.0051219375, -0.0269728694, -0.0217647292, 0.0130628161, -0.0730437487, 0.0486608595, -0.0314406678, 0.0362269841, -0.0211587287, -0.0009315489, -0.0395297222, -0.0231286492, 0.0282457992, -0.0434430987, 0.0346592292, -0.0276939757, -0.0110125216, -0.0321458578, -0.0460616611, -0.0372643284, -0.0454756804, -0.0223870557, 0.0071452409, -0.0052863806, 0.0144580239, 0.0400114581, 0.0052959551, 0.0434655137, -0.0611276999, -0.0417473167, 0.0194927864, 0.0295942891, 0.0866020992, 0.0632470623, 0.0151215056, 0.0421982557, -0.0146746896, 0.0334350877, 0.0180241801, 0.0431787521, 0.0103433076, 0.0358096622, -0.0217788704, 0.0176944509, -0.0059835128, -0.0434002019, 0.0053471373, 0.0725440755, -0.0615821332, 0.0040972908, 0.0957089961, 0.0402588993, 0.0086760679, 0.0012366844, -0.0837906599, 0.0568697006, 0.0313712545, 0.0626029074, -0.0790596604, -0.0041374597, -0.0145518966, 0.0244265068, -0.0071681011, -0.001949092, 0.0270606223, -0.1280364096, -0.0029125442, 0.0099356417, -0.0310816858, -0.0087523106, -0.035802938, -0.040747188, 0.0005743556, -0.0118070384, 0.0301214196, -0.0458963141, 0.0195051916, 0.0210036542, -0.0686697662, -0.0107152332, 0.0513555184, -0.0244652312, 0.0230116062, 0.0396568701, 0.0082305316, -0.0411034673, -0.017725518, 0.0211280957, -0.0629550591, -0.024674071, 0.0129433516, -0.0109639317, 0.048157908, -0.0284703225, -0.0532480329, 0.0550816469, -0.0200370178, 0.009134545, 0.066858761, 0.0845395699, 0.0233757719, -0.0208860803, -0.0134933749, -0.0089087766, 0.0123387948, 0.0808904693, 0.0359816812, -0.0630904064, -0.0431808047, -0.0059224018, -0.0358501188, -0.0186720043, -0.0615755431, 0.0278423894, 0.0040035676, 0.0065925689, -0.072726354, -0.0778523535, 0.0277001765, -0.0165416636, 0.0096077714, -0.0017041572, -0.0199948847, 0.0070917411, -0.032741908, 0.0350722671, -0.0397143215, 0.0461736694, -0.0511601195, 0.0107690711, -0.0119592864, -0.0263654068, 0.0090658534, 0.0009834677, -0.0980503261, -0.0130108865, 0.000805183, 0.0674286038, 0.0065768603, -0.0285821818, 0.008839584, -0.0650174692, -0.0071627353, -0.0178929158, -0.031376861, 0.0120154731, 0.0091660684, 0.023583211, 0.0042744726, 0.0564682446, -0.0470031053, 0.0220101755, 0.018817205, 0.038354896, -0.01618612, -0.0356807262, 0.0000080839, 0.0283265207, -0.0676925406, -0.0262521803, -0.0037417095, -0.0156760626, -0.0147370379, -0.0093769114, -0.0148280161, -0.0099308463, 0.0191280022, 0.0140150189, -0.0186284706, 0.0176470131, 0.0066759931, -0.0063726897, 0.0150328856, -0.0502872616, 0.0407952331, 0.0500676781, -0.007151105, -0.0164069775, -0.0007097791, -0.0373410508, 0.0158774499, -0.0188545808, 0.0499510951, -0.028130671, 0.0630206242, -0.0089817671, -0.0080170427, 0.025193464, -0.0237192959, -0.0153008178, -0.0572547875, 0.0447912402, 0.0265656523, 0.0575807355, -0.0219379906, 0.0479355864, 0.0260740761, -0.0230345074, -0.0359169245, -0.0069671003, 0.0254717488, -0.0290121231, 0.0439489633, 0.0252149049, -0.0482743233, 0.0115303127, 0.0847780183, -0.0069088885, 0.0189054366, 0.0234274343, -0.0103949709, -0.00589944, -0.010719521, -0.0019501161, -0.0172644183, -0.0129038682, 0.0073673157, -0.0609457642, -0.0380357057, -0.0039919498, -0.0155297974, -0.0097126933, -0.0052469657, 0.0069111325, 0.0517340899, 0.0108641814, -0.0123563437, -0.076766558, 0.0423984639, -0.0445562601, -0.003430692, -0.0063508307, -6.135199625e-33, 0.0042775446, -0.0361681543, -0.008355801, -0.0054670214, -0.0011983915, -0.0185673777, -0.0233704522, 0.0026842144, -0.0325543322, 0.0392220803, -0.0337153859, -0.0258716941, 0.0030818419, 0.0502881408, 0.0410905853, -0.0197581295, 0.0035642916, 0.0178605989, -0.0316020474, 0.0346759483, 0.0275731049, -0.0015769748, 0.0066320933, -0.0140734818, 0.0542030446, -0.0745315328, 0.0183745753, 0.0354406461, 0.0151696093, 0.1071387976, -0.0251920614, 0.031314306, 0.0402747504, -0.0148921032, -0.0054969206, 0.0515173748, -0.1009291485, -0.020722501, 0.0276429709, -0.0029260027, -0.0363617577, -0.0440464132, 0.0086235609, -0.0161286108, 0.0086376155, -0.0044948449, 0.0340597145, -0.0459828451, 0.0311132316, -0.0416712202, -0.0377136357, 0.0323842429, 0.0117860408, -0.0101489071, -0.0540688224, 0.1229932532, -0.0425558761, -0.0105726216, -0.0280907601, -0.0149365086, -0.0168662407, 0.04538792, 0.0592430457, 0.0131676551, -0.0108693717, 0.0294481013, 0.0083222426, 0.0023388283, -0.0063361982, -0.0009823011, -0.0341150686, -0.0276684687, 0.0597113408, 0.0790740103, 0.0140602654, -0.015969675, -0.0738138556, 0.0013331564, 0.0177935995, 0.0140294489, 0.0121789053, -0.0482370257, -0.0123843271, -0.0257082917, -0.045686923, 0.0352569185, -0.0237685218, -0.022297442, 0.0576695874, -0.0031912497, -0.0135996388, -0.0211155321, 0.055256784, -0.0283119027, 0.0107815573, 0.0343848951, -0.0036461428, 0.0009503079, 0.0452609621, 0.0405884609, -0.0400633812, 0.0577564016, -0.0227500331, -0.0216913577, -0.0014693908, 0.0102694752, 0.0311182346, 0.0038973799, -0.0772892162, 0.0133895902, -0.0088965474, 0.0284485277, -0.0239965748, 0.0311521906, 0.0217388831, 0.012417092, -0.0088149281, 0.0069528283, 0.0093537532, -0.0201934669, -0.0344368964, -0.0473675244, -0.001592485, 0.0202170629, -0.0465695001, 0.0194900557, -0.0433856659, 0.0582792647, -0.0395665541, -0.0129171126, 0.0015482814, 0.023414949, 2.437e-7, 0.0415504575, 0.0165869817, 0.0019050643, 0.0155602871, 0.0288725086, -0.0129147163, 0.0089911809, -0.0155099314, 0.0057045971, -0.0017200357, -0.0308272112, 0.0170138106, 0.0737769827, 0.0464595519, -0.0895252973, -0.0088266637, -0.0875120386, 0.0087885801, -0.0487364903, -0.0482420661, -0.0003641035, 0.0780459121, -0.0055831927, -0.0171708316, -0.0673211887, -0.058558736, 0.0616970994, -0.0037101449, -0.0225391649, 0.0191170722, 0.0146636683, -0.007872411, 0.0538387187, -0.009236712, -0.0272183418, -0.0146646816, 0.0531303287, 0.0819675177, 0.0172267258, -0.0041130907, -0.0135380067, -0.023475593, -0.0052392948, -0.1104890928, 0.0834636763, 0.011556155, -0.0177173354, -0.0308226459, 0.0089001833, -0.0232530497, 0.0428504944, -0.0059146057, -0.0061380174, 0.0331206508, 0.034040764, -0.0117852772, 0.0057104556, -0.0062029185, -0.0370973013, 0.0020357443, -0.0127818696, -0.0313028619, -0.0141151492, -0.0033030843, -0.0469135419, -0.0073697339, 0.0201422274, 2.016199875e-34, 0.0291310828, 0.007650116, 0.0042156675, 0.0336701013, 0.0272347033, -0.0244321749, -0.0059001488, 0.000329566, 0.028426405, -0.0616571419, -0.002466761 ]
http://arxiv.org/abs/2406.17755v1
http://arxiv.org/pdf/2406.17755v1
Accelerating Clinical Evidence Synthesis with Large Language Models
2024-06-25T17:41:52
2024-06-25T17:41:52
Zifeng Wang, Lang Cao, Benjamin Danek, Yichi Zhang, Qiao Jin, Zhiyong Lu, Jimeng Sun
cs.CL
cs.CL
Automatic medical discovery by AI is a dream of many. One step toward that goal is to create an AI model to understand clinical studies and synthesize clinical evidence from the literature. Clinical evidence synthesis currently relies on systematic reviews of clinical trials and retrospective analyses from medical literature. However, the rapid expansion of publications presents challenges in efficiently identifying, summarizing, and updating evidence. We introduce TrialMind, a generative AI-based pipeline for conducting medical systematic reviews, encompassing study search, screening, and data extraction phases. We utilize large language models (LLMs) to drive each pipeline component while incorporating human expert oversight to minimize errors. To facilitate evaluation, we also create a benchmark dataset TrialReviewBench, a custom dataset with 870 annotated clinical studies from 25 meta-analysis papers across various medical treatments. Our results demonstrate that TrialMind significantly improves the literature review process, achieving high recall rates (0.897-1.000) in study searching from over 20 million PubMed studies and outperforming traditional language model embeddings-based methods in screening (Recall@20 of 0.227-0.246 vs. 0.000-0.102). Furthermore, our approach surpasses direct GPT-4 performance in result extraction, with accuracy ranging from 0.65 to 0.84. We also support clinical evidence synthesis in forest plots, as validated by eight human annotators who preferred TrialMind over the GPT-4 baseline with a winning rate of 62.5%-100% across the involved reviews. Our findings suggest that an LLM-based clinical evidence synthesis approach, such as TrialMind, can enable reliable and high-quality clinical evidence synthesis to improve clinical research efficiency.
Accelerating Clinical Evidence Synthesis with Large Language Models: Automatic medical discovery by AI is a dream of many. One step toward that goal is to create an AI model to understand clinical studies and synthesize clinical evidence from the literature. Clinical evidence synthesis currently relies on systematic reviews of clinical trials and retrospective analyses from medical literature. However, the rapid expansion of publications presents challenges in efficiently identifying, summarizing, and updating evidence. We introduce TrialMind, a generative AI-based pipeline for conducting medical systematic reviews, encompassing study search, screening, and data extraction phases. We utilize large language models (LLMs) to drive each pipeline component while incorporating human expert oversight to minimize errors. To facilitate evaluation, we also create a benchmark dataset TrialReviewBench, a custom dataset with 870 annotated clinical studies from 25 meta-analysis papers across various medical treatments. Our results demonstrate that TrialMind significantly improves the literature review process, achieving high recall rates (0.897-1.000) in study searching from over 20 million PubMed studies and outperforming traditional language model embeddings-based methods in screening (Recall@20 of 0.227-0.246 vs. 0.000-0.102). Furthermore, our approach surpasses direct GPT-4 performance in result extraction, with accuracy ranging from 0.65 to 0.84. We also support clinical evidence synthesis in forest plots, as validated by eight human annotators who preferred TrialMind over the GPT-4 baseline with a winning rate of 62.5%-100% across the involved reviews. Our findings suggest that an LLM-based clinical evidence synthesis approach, such as TrialMind, can enable reliable and high-quality clinical evidence synthesis to improve clinical research efficiency.
[ 0.0550547503, 0.0644917637, -0.0041427845, -0.0219923593, 0.0029928559, 0.0059361663, 0.0023057987, 0.0114285164, -0.0313055627, -0.0095003452, -0.0119622378, -0.0142991459, 0.0104143238, 0.0804185644, 0.0102917738, -0.026838487, 0.0067767762, -0.0020998551, 0.1029759347, 0.0202546325, -0.0355502591, 0.0010604342, 0.000574324, 0.0633113459, 0.0166085903, -0.0167300515, 0.026507562, -0.0382069722, 0.0156212868, -0.0750550777, -0.0272807945, 0.0393085741, 0.0069867033, 0.0401471071, 0.0000020001, -0.0505708642, 0.0000921359, 0.0606065355, 0.0006987966, -0.0169239081, 0.0670269132, -0.0194532257, 0.0019734921, 0.0039141523, -0.0338174701, -0.0549294986, 0.0575953424, 0.0414005853, 0.0128267696, 0.0465287417, -0.0194334034, -0.0155058019, 0.0484548211, -0.0412949733, 0.036128372, -0.0116995024, 0.0434018783, 0.0197142921, 0.0543342866, 0.0039940812, 0.042380739, 0.0224934872, -0.0030082748, -0.0363842994, 0.0089933751, 0.0710001364, -0.0415872261, -0.0356467254, 0.022118384, 0.06351877, -0.0020227025, -0.0329512581, 0.0209819898, -0.0001483139, 0.0123889372, 0.0457081422, -0.0662181899, 0.0552720577, -0.0283759255, -0.0211970173, 0.0423249602, -0.014418073, 0.0438114293, -0.0090854736, -0.0032924374, 0.0203910917, 0.0053047449, -0.000458291, 0.0105297547, 0.0194621794, -0.0053654225, -0.0638783053, 0.0012859871, 0.0121016521, -0.0415537953, -0.0067828097, 0.0222986247, -0.0340596549, 0.0911434963, -0.0343614481, 0.0340676345, -0.0218067579, -0.0195443146, 0.0494537987, -0.0103370138, -0.0323536284, -0.0325268544, 0.0351160355, -0.027948115, 0.0154039785, -0.0535065718, -0.0224463399, -0.0206488315, 0.0129119894, -0.0120693222, 0.0010222322, -0.0522515923, -0.0019623071, -0.0267624836, 0.0107581308, 0.0355080515, 0.0229804646, -0.0494975857, 0.0002207331, -0.0155830011, 0.0230108909, -0.0130519969, 0.0372199342, -0.0079037994, -0.014494583, 0.0076483125, 0.0041680438, 0.0058453879, 0.0039317929, 0.0224683955, 0.0358627588, 0.0226135999, -0.0243682768, 0.0114394287, -0.0275208931, -0.0281524472, -0.065498583, 0.0013241533, -0.0965854004, 0.0456601828, -0.0046943729, -0.0031482852, -0.028972419, -0.0092358217, 0.0219175927, -0.0283760391, 0.0334952958, -0.08913102, -0.0063038282, 0.0302335378, -0.0041265478, -0.0337828137, 0.0285881236, 0.0344321318, 0.0123580555, -0.0023706965, -0.0133550679, 0.025826849, -0.0216613319, 0.0157396104, -0.0196701288, 0.0023302417, 0.0118930032, -0.0003150414, 0.071034275, -0.0264929105, 0.0595916845, -0.0403593406, 0.0010628225, 0.0316106826, 0.0720629767, 0.0099648517, 0.0494478308, 0.032258112, -0.0531143025, 0.0306953005, -0.0553855971, -0.0106203444, 0.0253571775, -0.0120501574, 0.0058960598, -0.0041695195, 0.0199404135, -0.0059164315, -0.0401992127, -0.0154219633, -0.0047159698, -0.0392223038, -0.0176894497, 0.0363492183, -0.0084529696, 0.0050658141, -0.0345766023, 0.0493211634, 0.002328269, -0.0250664018, -0.0399212725, -0.0016785483, 0.0393090993, 0.0689347759, -0.0492256023, -0.0014625184, -0.0455069467, -0.0055086114, 0.0202570669, 0.0215229727, -0.0316543989, -0.0051145968, -0.0114138341, -0.0290750396, -0.0213375185, 0.0158173181, 0.0100923916, -0.0040265867, 0.0207137596, -0.0389163867, -0.0861637741, -0.0639463291, 0.0352457017, -0.062610887, -0.0189014189, 0.0413144007, -0.0633189902, 0.0466164127, 0.0194165949, -0.0414737351, 0.0449287407, 0.0403185748, -0.0484329648, -0.0496835262, 0.0224642418, 0.0288414564, -0.0240907725, 0.0357783213, -0.0671430901, -0.0206414461, -0.0427000746, -0.0589163005, -0.025144903, 0.0008661638, -0.0252442248, 0.0267368052, 0.0149692614, -0.0112945559, 0.0171347987, -0.0036507174, -0.0428536572, 0.0534241237, 0.0081539406, -0.0392213725, 0.0428227633, -0.083875902, 0.0126858735, 0.0025778783, -0.0212739818, 0.0495151989, 0.0062873024, -0.0049416833, 0.0209364202, -0.0134426998, -0.0052312217, -0.016199341, -0.1014909819, 0.0282659214, 0.0317766666, 0.0048871557, 0.0060609756, -0.0133573683, -0.0050994945, -0.0068672514, -0.0326247104, -0.0275395177, 0.0081956247, 0.0207072441, 0.0389551669, 0.072928153, -0.0130866319, -0.0949261636, -0.0239677858, -0.0351079665, 0.0640744492, 0.0528371371, -0.0057523539, -0.0042939, 0.0097458242, -0.0121411989, 0.0245248526, -0.0051499559, 0.0183414444, 0.0331362709, -0.0256355703, 0.011107943, 0.0450922027, -0.017676549, 0.0113803372, 0.0906406641, 0.0073605548, -0.0303600486, 0.0112653887, -0.0308201686, -0.0118542286, 0.0234984551, -0.0374577306, 0.0432268828, 0.0457149334, 0.0142202759, 0.0088104596, -0.0201053005, 0.0077762003, 0.0002182104, -0.0057753911, -0.0295982286, -0.018404562, -0.0043126219, 0.0305334404, -0.0186742526, -0.0571767762, 0.1122331992, -0.0098101189, 0.0371509306, -0.0688710809, 0.0251759142, -0.0326051787, 0.0023283372, 0.0133711528, 0.0044148592, 0.0065291212, 0.0289827939, -0.0010834443, -0.0089977803, 0.0359138958, -0.0213954914, 0.0469339788, -0.0542841814, 0.0468729958, 0.008245104, -0.02518766, -0.0239046738, -0.0074480255, -0.0419662111, 0.0042395252, 0.030532008, -0.0070624207, -0.0680645555, 0.0058161607, 0.031988427, 0.0641167387, -0.0214835051, -0.0204703342, -0.0024414377, -0.0316539891, 0.0652839765, -0.0667579249, 0.0134036206, 0.000688453, 0.0616201386, 0.0567226075, -0.0093389889, 0.0334663168, -0.0455705822, -0.0107430182, -0.0349224769, -0.0610146001, -0.0083393473, 0.0207396448, -0.0155800106, -0.0259435046, 0.0037004412, 0.1036733612, 0.0236148853, -0.1070361808, -0.0150884772, -0.0345005356, -0.0160347614, -0.0163538717, 0.0262683071, -0.0731342286, -0.0195284691, -0.0071975705, 0.0163016915, -0.0065117534, -0.0406273864, 0.0074968426, -0.0486292094, 0.0467557274, 0.0084036961, -0.0299595315, -0.0253142118, -0.069808729, -0.0010757484, -0.0349819064, -0.0426284149, -0.0114180669, -0.0006220508, 0.0180004202, -0.0100889318, -0.0422496945, -0.0995017737, 0.0327536017, -0.0533735864, 0.0685955063, 0.0466049165, 0.0421773084, -0.0169882756, -0.0016128535, 0.0031136177, 0.0063601378, -0.0052638366, -0.0093372269, 0.0970463902, 0.0308023971, 0.0346443541, -0.025678575, 0.0363642313, -0.0469553024, -0.0356841199, 0.0357986726, 0.1093226746, -0.0427466482, -0.0461233631, -0.0630160943, 0.0108513385, 0.0033099004, 0.0519247018, 0.0115978466, 0.0086000888, -0.0025118685, -0.0118511301, 0.0420819893, -0.0045676245, -0.0165810436, 0.0256618802, -0.0164188836, 0.0102775171, -0.0646491051, -0.0245130602, 0.0781356543, -0.0166260079, 0.0272563193, 0.0382298231, 0.023790434, -0.0227167681, -0.04175511, 0.0660204142, 0.0267981347, 0.0193458442, -0.0373747014, -0.0382321179, 0.0672149658, -0.0055557154, 0.0126970122, 0.0454642698, -0.0115180165, -0.0125205824, -0.0456663109, 0.0664247721, 0.018801488, 0.003703871, -0.0051527554, -0.021395335, 0.0307485349, 0.0153786847, 0.0157868285, -0.0679203942, -0.018628493, -0.0304898135, 0.0036989965, -0.0364819169, -0.0464419276, 0.0393252522, -0.0110239601, 0.0093374997, -0.0083056791, 0.0067960061, 0.07018736, 0.0417666882, 0.0527351126, -0.0519346148, 0.0048134211, -0.031769149, -0.0254073255, -0.0555386432, 0.0141912159, -0.0359728821, 0.03088893, 0.0265498236, -0.0029110142, -0.0053907982, 0.0086082574, -0.078156516, -0.0172408652, -0.0470248759, 0.0503653921, 0.0036039036, 0.0433860794, 0.0639300421, -0.0257479474, -0.0095240995, -0.0356723182, -0.0428453535, 0.0186656546, -0.0243400279, 0.0119466437, 0.0058397613, 0.0031788372, -0.0272411145, 0.0479375087, -0.0545858219, -0.0357718468, -0.0256977566, -0.0079066288, 0.040358834, 0.01699402, 0.0454714186, -0.0418099985, 0.0019314107, -0.0457261577, 0.0189199075, 0.0024418579, 0.0146983778, 0.0170968454, 0.0158959236, -0.0384905152, -0.0620862171, 0.097750105, 0.0154285552, 0.0378483608, -0.0196319204, -0.0618828535, -0.0265497826, -0.0340341441, -0.0136242239, 0.0331226699, -0.0459138192, -0.0169456974, -0.0598367751, -0.0221532732, -0.0152652031, 0.0197501946, -0.0182718318, 0.0036394882, -0.0105440076, 0.0399226435, 0.0087934593, -0.0284096766, -0.032344386, 0.040828038, -0.0155777624, -0.055516798, -0.0328604467, -5.44202677e-33, 0.0338255242, -0.0265162736, 0.0144347185, -0.0219768006, 0.0253557749, 0.0225048345, -0.075468123, -0.0095652817, 0.0202283058, 0.0862068236, -0.0309866443, 0.0089875041, -0.0109587861, 0.0122154178, 0.0441203192, 0.0217652302, 0.0414545238, -0.0096063688, -0.0060549397, 0.0479482226, -0.0243332908, -0.0091471188, 0.0222917702, -0.039626345, -0.0079636481, 0.0548369847, 0.0073827296, -0.0205229763, -0.0258625634, 0.0720872208, -0.0266087111, 0.0373300388, 0.0248081312, -0.0103882225, -0.0286492482, -0.0305870157, -0.0531233065, -0.0267118737, 0.009181357, -0.0603526905, -0.0171098374, -0.0132471174, 0.0365527011, -0.010207233, -0.0077622226, 0.014811689, 0.0389842503, -0.0599599816, -0.044740703, -0.0533732511, -0.0615028776, 0.0490253419, -0.0167928226, 0.0145596927, -0.0213791449, 0.018498905, -0.007130614, 0.0303157046, -0.0279153269, 0.0361935087, 0.0238951892, -0.0054356861, 0.0052600806, -0.039508231, -0.002173553, 0.0734420568, -0.0033880777, 0.0253275, -0.0018287784, 0.0159823652, -0.0282544978, 0.0430880077, 0.0430368893, 0.022754062, 0.0304067805, 0.0132007506, -0.0809401348, 0.07728111, -0.0592279211, 0.0371057354, 0.0520222746, -0.0055682096, -0.0068844105, -0.0222487878, -0.014822497, -0.0341784135, -0.0221430324, 0.0047449898, 0.024716096, 0.0009493873, 0.0113721322, -0.0325846933, 0.0431674495, -0.0158125143, 0.01333375, 0.0369069763, 0.0015336544, 0.0267658904, -0.0039267768, -0.0153786289, -0.0461179614, -0.0091949189, 0.0365304388, 0.0156450104, 0.0446388349, 0.0207407437, 0.0188301317, -0.0222072117, -0.0505512431, 0.0017922686, -0.0153449364, 0.0161985289, 0.0002489898, -0.0149226543, 0.0004863843, 0.0194252562, 0.0080850767, 0.0451753624, 0.0187425129, 0.0003838785, -0.023329908, 0.0112179797, -0.0057684113, -0.0117637692, -0.0134519786, -0.0009417936, -0.0293151997, 0.0173536446, 0.0624234602, -0.0629820675, -0.008786899, -0.0134372208, 2.746e-7, 0.097168155, 0.0602273606, 0.0005854247, -0.0175344255, 0.0497567169, 0.0198042654, 0.0040450073, -0.0286217742, 0.0301738475, 0.0225128494, 0.0619614013, 0.0177954044, -0.002823778, 0.0492421165, -0.0468548909, -0.0284849871, -0.0687858388, -0.042112425, -0.0543608852, -0.0473671444, 0.021921128, 0.0525635183, -0.0276081618, -0.0128617007, 0.0066108168, -0.0319198631, 0.0226722695, -0.0399576053, 0.0176413879, 0.0563249588, 0.043264579, -0.0138183013, 0.0306230187, 0.0371638238, -0.0357293487, 0.0058781044, -0.0194194205, 0.0497085154, -0.0273004659, 0.0621694401, 0.0436175764, -0.0177793279, 0.0624557026, -0.0402474068, 0.0355095454, -0.0847982988, -0.0549285561, -0.0500103459, -0.0183344148, -0.0083648991, 0.070773907, -0.0360788889, -0.0020886674, 0.0377697758, -0.0064732186, 0.04327875, -0.0283974968, -0.0381741934, -0.0046509905, 0.0635048524, -0.0482436791, -0.0111071309, -0.0046620485, -0.0694679692, -0.0089819515, -0.0480484404, 0.052691713, 2.517123706e-34, 0.0663231388, -0.0109794885, 0.0169102289, -0.005180405, 0.0709240362, 0.0065249014, -0.0188908968, 0.0171854869, 0.0145796137, -0.0852677152, -0.0078071984 ]
http://arxiv.org/abs/2406.17753v1
http://arxiv.org/pdf/2406.17753v1
Measuring and Benchmarking Large Language Models' Capabilities to Generate Persuasive Language
2024-06-25T17:40:47
2024-06-25T17:40:47
Amalie Brogaard Pauli, Isabelle Augenstein, Ira Assent
cs.CL
cs.CL, cs.AI
We are exposed to much information trying to influence us, such as teaser messages, debates, politically framed news, and propaganda - all of which use persuasive language. With the recent interest in Large Language Models (LLMs), we study the ability of LLMs to produce persuasive text. As opposed to prior work which focuses on particular domains or types of persuasion, we conduct a general study across various domains to measure and benchmark to what degree LLMs produce persuasive text - both when explicitly instructed to rewrite text to be more or less persuasive and when only instructed to paraphrase. To this end, we construct a new dataset, Persuasive-Pairs, of pairs each consisting of a short text and of a text rewritten by an LLM to amplify or diminish persuasive language. We multi-annotate the pairs on a relative scale for persuasive language. This data is not only a valuable resource in itself, but we also show that it can be used to train a regression model to predict a score of persuasive language between text pairs. This model can score and benchmark new LLMs across domains, thereby facilitating the comparison of different LLMs. Finally, we discuss effects observed for different system prompts. Notably, we find that different 'personas' in the system prompt of LLaMA3 change the persuasive language in the text substantially, even when only instructed to paraphrase. These findings underscore the importance of investigating persuasive language in LLM generated text.
Measuring and Benchmarking Large Language Models' Capabilities to Generate Persuasive Language: We are exposed to much information trying to influence us, such as teaser messages, debates, politically framed news, and propaganda - all of which use persuasive language. With the recent interest in Large Language Models (LLMs), we study the ability of LLMs to produce persuasive text. As opposed to prior work which focuses on particular domains or types of persuasion, we conduct a general study across various domains to measure and benchmark to what degree LLMs produce persuasive text - both when explicitly instructed to rewrite text to be more or less persuasive and when only instructed to paraphrase. To this end, we construct a new dataset, Persuasive-Pairs, of pairs each consisting of a short text and of a text rewritten by an LLM to amplify or diminish persuasive language. We multi-annotate the pairs on a relative scale for persuasive language. This data is not only a valuable resource in itself, but we also show that it can be used to train a regression model to predict a score of persuasive language between text pairs. This model can score and benchmark new LLMs across domains, thereby facilitating the comparison of different LLMs. Finally, we discuss effects observed for different system prompts. Notably, we find that different 'personas' in the system prompt of LLaMA3 change the persuasive language in the text substantially, even when only instructed to paraphrase. These findings underscore the importance of investigating persuasive language in LLM generated text.
[ 0.0463254713, 0.0483291745, -0.0057436181, 0.0175687317, -0.0338355377, 0.002426987, 0.0135555305, 0.0009108237, 0.0085552298, -0.0441887043, -0.0078761913, -0.0268836133, 0.0130818039, 0.0037754567, 0.0435895175, -0.0739563778, 0.0070382534, -0.002345168, 0.0669151992, -0.015384445, 0.0461261123, 0.0313906856, 0.0206090249, 0.0629093647, -0.0128494808, -0.0395357981, 0.0418510959, 0.0247796904, 0.0143961282, -0.0439762659, 0.0206793696, 0.0413589589, 0.0157612376, 0.0109674549, 0.0000018508, -0.0503958352, 0.0256023314, -0.0084854905, -0.0315633751, 0.0665245205, 0.1282641888, -0.0024529919, -0.0168586019, 0.0216978658, -0.0307355542, 0.0365238301, 0.0609529354, 0.0474782661, 0.0581615493, 0.0629293695, -0.0078612324, -0.0110185323, 0.0120599987, -0.0321217701, -0.0161887947, 0.0045274976, 0.0215580333, -0.031838499, -0.0270065572, 0.0523269102, 0.0253276303, 0.0247087162, -0.0049915994, -0.0064096544, -0.0016764754, -0.0002403195, -0.0085325278, -0.0062135607, -0.0173853096, 0.0381936952, -0.0074076317, -0.0245259628, 0.0349188298, 0.0031521157, 0.0362112857, 0.0330476686, -0.0273727179, 0.034961272, 0.0198238492, 0.0177924335, 0.01105751, 0.014562008, 0.0404247753, 0.0230488069, 0.009107531, 0.0343118273, 0.0066824281, -0.035560105, 0.0488608554, -0.0181212258, 0.0020758319, -0.0724312142, -0.0333016999, -0.0234612543, -0.0182342976, 0.0397104323, -0.0317252204, -0.0184246115, 0.0499057695, -0.1132888943, 0.0444492586, -0.0005265719, 0.0080898497, 0.0128934523, -0.0124579454, 0.0129446974, -0.0256261621, 0.0228489898, 0.0096824402, 0.0387320444, -0.0371368937, -0.022769846, -0.0254671723, 0.0015375949, 0.0155522693, -0.0121285245, -0.0710593164, -0.0027691536, -0.0331476219, 0.0074144537, -0.0487015396, 0.0121345604, -0.0135937147, -0.025608195, -0.007233053, 0.0637979433, -0.0465976335, 0.0225691982, 0.0228142366, -0.0300420355, 0.0310359411, 0.0383793078, -0.0175515004, -0.0046492252, 0.0096620908, 0.0408283286, -0.0068845944, 0.004753632, -0.0451456271, -0.023877304, -0.037697088, -0.051817622, -0.0216725916, -0.062917009, -0.0046274667, 0.0098258629, 0.0276517682, -0.0653775632, 0.0314471014, 0.0174557026, -0.0298534259, 0.0443854108, 0.0193647724, 0.0295343492, 0.0374184996, 0.0309771691, -0.0158241391, 0.0617363341, 0.0480844788, 0.0847966671, 0.0514960587, 0.0043967413, 0.0339171439, 0.0348198153, 0.0015678866, 0.0245800428, 0.0178785902, -0.0178519953, -0.0284161717, 0.0038584436, -0.0165099297, 0.0667004958, -0.0035811553, 0.0218247641, 0.099723883, 0.0438503847, -0.011875594, -0.0012848901, 0.0321294591, 0.0295264032, 0.0109017976, 0.0058048894, -0.0264124032, -0.0274082087, -0.0316458791, -0.0091846287, 0.0373408236, 0.0175942685, -0.017808279, -0.0367857404, -0.0159694571, -0.013744303, 0.0120602865, -0.0114712361, 0.0076680584, 0.0082369139, -0.0521117859, 0.0143622924, 0.0549348891, -0.0432908945, -0.0105356593, -0.0254520662, 0.0125657264, 0.0406667218, 0.0112564275, 0.0091876276, 0.0566828586, -0.0556290187, 0.0475982241, 0.0824928656, 0.0496921316, -0.0229879599, -0.0468677133, -0.0175507031, -0.0144188069, -0.0626869127, 0.0295980964, 0.0491316989, 0.0055808853, 0.0198074821, 0.022040749, -0.0071253804, -0.0475952961, 0.015580086, 0.0051087961, -0.0187817756, -0.0093057826, 0.0229364932, 0.0416599102, 0.0173247308, -0.0317329131, 0.0027520019, 0.0114811948, -0.0413719825, -0.0273940265, -0.0043428624, -0.0075649866, 0.0356716886, 0.0080097774, -0.0219705496, -0.0110161211, 0.0221884903, 0.0086445296, 0.020711828, 0.0250686556, -0.0620040707, 0.0736528113, -0.0075468216, 0.0179199427, -0.0300841257, -0.0054339161, -0.0332824998, 0.0322011933, 0.0335587971, -0.0257336888, 0.0263678487, -0.0271263532, 0.0143593773, 0.0192680545, -0.0043945001, -0.0281905904, 0.0087097809, 0.014052161, 0.0447369553, -0.0480613187, -0.0251508988, -0.0555697978, -0.0585024133, -0.0378789268, -0.014923838, 0.0064546117, 0.0142357918, -0.0138344234, 0.0015862242, 0.0542315356, -0.0691582039, -0.0503935553, -0.0093472824, 0.0429900624, 0.02619607, -0.0002786979, -0.0202928372, -0.0187664684, 0.0285812542, -0.0217014309, 0.0438078344, 0.0552522205, -0.0372922644, -0.0346576348, 0.0414691903, -0.0326927863, -0.0495841876, -0.0210167486, -0.0505375825, 0.0151742026, 0.0408898108, 0.0837626234, 0.0521663837, 0.0403047316, 0.0257389341, 0.0187864136, 0.0453533903, -0.0646405444, -0.0370865837, -0.0430675745, -0.0010285109, -0.0626973137, -0.0818268135, 0.0520766601, 0.0132547831, 0.041181393, 0.0219011959, -0.0395994037, 0.0678253844, -0.0050248979, -0.0458820276, -0.0134773934, -0.0348490812, -0.0139864665, 0.047289744, -0.0217719525, -0.0800902694, 0.0431800447, -0.0341919288, 0.0164195392, 0.012780779, -0.0188811515, -0.0110787274, 0.0219180863, -0.0161333922, 0.0427050106, 0.0038798051, -0.0100221494, -0.0083405366, -0.0711459294, 0.0243739672, -0.0458490998, 0.0373894162, -0.0161488205, 0.0011645369, 0.0539219975, -0.0446291678, -0.0272715539, -0.0877831802, -0.0219204407, -0.0273366868, 0.0518246517, -0.0187727772, -0.034658514, -0.0225458667, 0.0524766371, -0.0039775898, -0.0177332833, 0.0002430378, 0.0027237684, -0.0218321569, 0.0285462346, -0.0232950002, 0.0020273787, -0.0709234402, 0.0138410432, 0.0024183285, -0.0245460402, -0.022456944, -0.0445036292, -0.047947038, -0.0156703163, -0.0143073006, 0.0084773675, 0.010768339, 0.0382652469, -0.0359592587, 0.0469031036, 0.0714117214, 0.0058472119, -0.0267017987, 0.0163605995, -0.0092725176, 0.0330643766, 0.0066368608, 0.0116720041, -0.0596021749, -0.0090028886, 0.0004761645, 0.0192135461, -0.0376048945, 0.0291015115, -0.0297883786, -0.0869439021, 0.0477898903, 0.027012594, 0.0414227881, 0.0005436904, -0.0479148626, -0.0236408133, -0.0003869941, -0.0122346599, -0.0299136061, -0.0152034648, -0.0182003081, -0.0056588026, -0.0350614451, -0.0297087114, -0.0295833033, -0.0152503345, 0.0384518877, 0.0944782645, 0.0474640056, -0.0111003676, -0.0530426241, 0.0393705517, 0.0279955193, -0.012668306, 0.0006861974, 0.0608565696, -0.0092351278, -0.0053559309, -0.0261712354, 0.0307303686, 0.0169235617, -0.0184932016, 0.0218446329, 0.0318196565, 0.0248927604, -0.0452831909, 0.0119302031, 0.0280079879, 0.0410582945, 0.082781218, 0.037047971, -0.0535754338, -0.0006651263, -0.0545264035, 0.0007323793, 0.0049609011, -0.0729220808, -0.0521771684, -0.0307825189, 0.0121297939, -0.0584550463, -0.0242099706, -0.0157758389, 0.0059221033, 0.0574282221, 0.032532461, -0.0090077696, 0.0093018487, 0.0254229829, 0.0140524507, -0.0554548763, 0.0477956459, -0.059902899, -0.0373837501, 0.001717633, -0.0205411278, -0.0173588339, -0.0141479727, -0.0200266, -0.0770380199, -0.0135446871, 0.0363132581, 0.0224675257, 0.0222720671, 0.0389278904, -0.0491112582, 0.074893415, 0.0526525341, -0.0175442845, 0.0308692083, 0.0369897671, 0.0428484753, 0.0189882871, 0.0153097166, -0.0036383539, 0.0435369946, 0.0467519052, -0.0177194308, -0.0317023247, -0.0165379047, 0.0320627876, 0.0539821759, 0.0187879559, -0.0284900349, -0.0108279791, -0.0242008511, -0.0129778013, -0.0295571964, -0.0305560566, -0.0208268203, 0.0108728316, 0.038299799, 0.0060057072, 0.0057889395, 0.0008183248, -0.0535315573, 0.0274145342, -0.0233338121, 0.0140499333, -0.0640883967, 0.0418928638, 0.023172386, 0.0017823048, -0.0130988462, 0.0503063016, -0.0144637739, 0.028058216, 0.0212310068, 0.0289583039, 0.0133746536, -0.0179576464, 0.0086099105, 0.0295560844, -0.042443309, -0.0427886173, 0.0058534769, -0.0081171282, 0.0072290893, 0.0104882875, 0.0801902115, -0.0316774212, -0.0034608108, -0.0235150028, 0.0264186282, 0.0035510939, -0.0310987681, 0.022641588, -0.021709105, -0.0540933646, 0.0075323791, 0.047993511, -0.0561311096, 0.0369182453, 0.0264356509, -0.0251190439, -0.0434102193, -0.0480256937, 0.036595419, 0.0125526134, -0.0013721235, -0.0121910376, -0.0691613257, -0.0112328473, -0.0039769858, 0.0003454699, 0.0005131696, -0.0341205597, 0.0148449931, 0.0244790595, -0.0587368757, 0.0201972704, -0.0555304289, 0.0088141952, -0.0172088593, -0.0050880918, -0.0017453542, -5.87991017e-33, 0.0082403896, -0.0110102585, -0.0331444554, 0.0283314232, -0.0184541214, 0.00497049, -0.0134106996, 0.0378752649, -0.0016119331, 0.0713287368, -0.0008390378, -0.0101140356, 0.0073884632, -0.0060136165, -0.0341363885, -0.0113012763, 0.0621061027, -0.0039410801, -0.055752486, 0.0450157411, 0.03886142, -0.0199362133, -0.0135305729, -0.0998440608, 0.057693176, -0.014973294, -0.0544951074, -0.04205795, -0.0050752298, 0.0217743348, -0.0036689271, 0.0491489358, 0.0130059123, -0.0370002091, -0.0062200068, 0.1169186682, 0.0218537785, -0.0086143734, 0.0773452595, -0.0551035367, -0.0186659284, -0.0651715174, 0.0197074842, -0.0258633476, -0.043732699, 0.0295187794, 0.0319905728, -0.0362290405, 0.0157610644, -0.0105811441, -0.0730112046, 0.0046722088, -0.0268587768, -0.0171116944, -0.0417598002, 0.0701295659, -0.0168073494, 0.0758594051, -0.0093760137, 0.0523976795, 0.0196677595, 0.0237076972, 0.0330987386, -0.0168239232, 0.0036275405, 0.0871365964, 0.0007049259, 0.0214482117, -0.0345811397, -0.0065156464, -0.0077945115, 0.0274003036, 0.0727309659, 0.0903863311, 0.0120912967, -0.0314172357, -0.0554405935, -0.0069437991, 0.0511129722, -0.0525934435, 0.050249245, -0.036316026, -0.0247394983, -0.0372186936, -0.0631626099, -0.0406754352, -0.0202829614, -0.0683585852, 0.0238208864, 0.0147867585, -0.0321779475, 0.0298045631, -0.0094840219, -0.0395905189, -0.0332130268, 0.0165149178, -0.0298496112, 0.0189573076, -0.0097062783, -0.012581028, -0.0644396842, 0.0242633931, -0.0126858009, 0.0069290069, 0.0329146683, -0.0021396566, 0.044980213, -0.0260793176, -0.0877990127, 0.0148652066, -0.0208087768, 0.0219434984, 0.0045248885, -0.0325048044, 0.0171947926, 0.0029491882, 0.0090063214, -0.0006949167, -0.0127838207, -0.0469032004, 0.0037995181, -0.0140646109, -0.0204340946, 0.0465853699, -0.0557472147, -0.0213670135, -0.0369017608, 0.0163504314, 0.0450768173, -0.0602328032, -0.0174147356, -0.0650661737, 2.57e-7, 0.011979308, 0.089820236, -0.0228937659, 0.0267452691, 0.0031667647, 0.0536253117, -0.0149793662, 0.0028252094, 0.0150469067, -0.0647202283, -0.0340207145, -0.003735336, 0.0503877252, -0.0051109781, -0.1048541144, 0.0010537531, -0.0550784171, 0.0026907981, -0.0759747326, 0.0072378945, 0.0664748698, 0.0595696531, 0.0456294455, -0.006442239, -0.0155119766, -0.0221178383, 0.0825791061, -0.0870044902, -0.0288469773, 0.0468592755, 0.002346175, -0.0306049418, -0.0047843428, 0.0161592942, -0.0108942194, -0.0099147446, -0.030782165, 0.0933076292, 0.0351752229, -0.0103882048, -0.0071280389, 0.0014254596, -0.0305243377, -0.0239810254, 0.025798155, -0.0389676318, -0.0363917649, -0.0518488958, -0.022650864, -0.0087760212, 0.0171917677, -0.0382094495, -0.0329057276, -0.017368421, 0.0133182723, 0.0096937036, -0.0018588343, -0.0331669748, -0.0015708121, -0.0198730379, -0.0137138152, -0.0808691233, -0.0372140668, 0.0327103361, 0.0123602552, -0.0633338615, 0.0304676276, 2.21358581e-34, 0.0657182857, -0.0388681404, 0.020866625, -0.0168346465, 0.0369583108, 0.0326648355, 0.0368766263, -0.0084683606, 0.0090680355, -0.0754501969, 0.0280447956 ]
http://arxiv.org/abs/2406.17746v1
http://arxiv.org/pdf/2406.17746v1
Recite, Reconstruct, Recollect: Memorization in LMs as a Multifaceted Phenomenon
2024-06-25T17:32:16
2024-06-25T17:32:16
USVSN Sai Prashanth, Alvin Deng, Kyle O'Brien, Jyothir S V, Mohammad Aflah Khan, Jaydeep Borkar, Christopher A. Choquette-Choo, Jacob Ray Fuehne, Stella Biderman, Tracy Ke, Katherine Lee, Naomi Saphra
cs.CL
cs.CL, cs.AI
Memorization in language models is typically treated as a homogenous phenomenon, neglecting the specifics of the memorized data. We instead model memorization as the effect of a set of complex factors that describe each sample and relate it to the model and corpus. To build intuition around these factors, we break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither. We demonstrate the usefulness of our taxonomy by using it to construct a predictive model for memorization. By analyzing dependencies and inspecting the weights of the predictive model, we find that different factors influence the likelihood of memorization differently depending on the taxonomic category.
Recite, Reconstruct, Recollect: Memorization in LMs as a Multifaceted Phenomenon: Memorization in language models is typically treated as a homogenous phenomenon, neglecting the specifics of the memorized data. We instead model memorization as the effect of a set of complex factors that describe each sample and relate it to the model and corpus. To build intuition around these factors, we break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither. We demonstrate the usefulness of our taxonomy by using it to construct a predictive model for memorization. By analyzing dependencies and inspecting the weights of the predictive model, we find that different factors influence the likelihood of memorization differently depending on the taxonomic category.
[ -0.0145433759, 0.0112044932, -0.0447326452, 0.0228190403, -0.0550800674, 0.0135329803, -0.0640933514, 0.006180427, -0.0227598324, -0.0260270331, -0.0129064787, -0.0530650802, 0.0602273829, -0.0132455537, 0.0318234712, -0.0440001152, 0.0341428109, -0.0599135011, -0.0180647932, -0.0135714058, -0.0230452083, -0.0385360457, -0.027386494, -0.0050677354, -0.005071613, 0.0120856054, -0.0145892026, -0.0298844073, -0.0304767825, -0.029706724, -0.0083533777, 0.0358020365, 0.0241961312, -0.00849149, 0.0000016356, -0.0615269877, -0.0082351556, -0.0565352961, -0.0256765317, 0.0171526838, -0.0099303434, 0.0168428142, 0.0055454089, -0.0198186859, -0.008882652, -0.0110560125, 0.023674408, 0.0108402744, -0.0317378864, 0.0348056629, 0.0094937887, -0.0040853773, 0.0158642884, -0.0084232949, 0.0431177504, 0.0289206021, 0.028476093, 0.0199548583, -0.0239461213, -0.0318229869, -0.0388106778, 0.0266063232, 0.0003309899, 0.0849445984, 0.0262862109, 0.0409160927, 0.0809480399, 0.0151673658, 0.0001306682, -0.0030295146, 0.0394907109, -0.038951654, 0.017574355, 0.025296079, 0.0030125189, 0.0395300649, -0.0729059428, 0.0404627696, -0.0143863214, -0.0002889335, 0.0149271823, -0.0013056828, 0.0343973972, 0.0007030996, 0.0019193841, 0.0204461217, 0.0241614375, -0.0286566876, -0.0270957444, 0.010413412, 0.0349694081, -0.0464873239, -0.0033359975, 0.063826099, 0.0082673905, 0.0564925857, 0.0350644328, -0.0220849887, 0.0665486231, -0.0069626626, 0.0674850941, 0.0207923241, 0.0160725471, -0.0141364411, -0.0627927259, 0.0733458698, 0.0169793162, 0.0032461591, -0.0125876842, 0.0311933961, -0.0762258917, -0.0326761305, 0.0063004182, 0.0620809458, 0.0912994072, 0.0094414083, -0.0439214669, 0.0221662559, 0.0273301639, -0.0278846808, -0.0275860038, -0.0101297097, 0.0003850048, 0.0099604223, -0.0363084637, 0.0046033314, -0.0306890402, 0.0479337126, 0.0307580158, -0.0426922888, -0.0270553827, -0.0143391239, -0.0233355667, 0.0039382405, -0.0267457217, 0.0121884625, -0.0058480492, -0.0413234271, -0.058821205, -0.0584229231, -0.02260332, -0.0429569148, 0.03544816, -0.0535642542, -0.0161019769, 0.0151743796, 0.0210370254, -0.0163505655, -0.024705058, -0.001675093, -0.0079923412, 0.0188380349, 0.0886571109, 0.0350961573, 0.0222517587, 0.0345922485, -0.0016237175, 0.0212667603, 0.023545213, -0.0001319859, 0.0095382938, -0.0267217327, 0.0351033174, 0.0363709293, 0.0059006209, -0.0099484129, -0.0297150072, -0.048551593, 0.0178103987, 0.0235046726, 0.0726174489, 0.039485272, -0.0410720222, -0.0231386758, 0.0579612926, 0.085110262, 0.0313198008, 0.0056644175, 0.0256777015, -0.0357472301, 0.0336953998, 0.0689772069, -0.0149305575, 0.0396228768, -0.0043454752, 0.056288369, 0.0021501465, 0.0075540296, -0.0438480526, -0.0025269401, -0.019859327, -0.0374496654, -0.0086517921, -0.0100609986, 0.020971518, 0.0195461605, -0.0421593599, 0.0681140795, 0.0040487074, -0.0541402176, 0.0230105836, -0.0562499799, 0.0567550436, 0.0556509271, 0.0351249166, 0.0123912469, -0.0057871914, 0.034731742, -0.0374897085, -0.0259573981, 0.0053683366, -0.011362046, 0.001842898, -0.041306451, -0.0157282185, 0.0381749272, 0.0290417317, 0.0468487516, 0.0015413879, -0.0033343304, -0.0140239894, -0.0168255027, 0.0115540382, 0.0183038488, -0.0455945991, -0.0437605307, -0.0237734914, -0.0068872222, 0.0398501009, -0.0001009783, 0.0409251787, 0.0394797251, 0.0631930083, 0.0086470209, 0.0055386182, -0.0050610453, -0.1015075222, -0.0273931287, 0.0205459371, 0.0148268864, 0.0301707573, 0.0240545943, 0.0242139213, -0.0675623566, 0.0356088243, -0.0400052033, 0.1018435359, -0.0219935086, 0.0416410603, 0.0048044045, -0.0032762715, -0.0450566523, 0.0517901517, -0.0593685061, 0.0390800014, 0.0202138796, -0.0245501474, -0.0142480824, -0.0766101927, -0.0116659608, 0.0101380283, -0.0150318937, 0.0205269568, 0.0071030492, -0.0697810501, -0.067171149, -0.0291889552, -0.0351609364, -0.0226253681, 0.0026628245, 0.0240162909, 0.0148505969, -0.0248970669, -0.0130203664, 0.0036977322, -0.0305945911, 0.0123134209, -0.0010924541, 0.0726722032, -0.0005763568, -0.0276903547, -0.0131901316, -0.0362313204, -0.0288711879, -0.0150993848, 0.0235801674, -0.0258468725, -0.008898708, -0.0344087817, 0.025149215, -0.0347298346, -0.03371346, -0.0146824075, 0.0225957092, 0.0339179337, -0.0015644751, 0.1079012901, 0.0583263859, 0.0067899004, 0.014604331, 0.0576368943, -0.0514966547, -0.040865317, -0.0637618229, -0.0591282845, -0.0416217893, 0.0035305333, -0.0467646345, -0.0038549695, 0.0893548951, 0.0597361363, 0.0152828926, 0.0116951233, -0.04084884, -0.0158145241, 0.010318392, -0.0243256297, -0.0330555923, -0.0919920802, 0.0464947596, 0.0010003782, -0.0664238855, 0.0246144272, -0.035209246, 0.020751385, 0.0056705442, 0.0213154517, 0.0186044406, 0.0413812585, -0.0016529392, -0.0160641205, 0.0144897886, -0.0540470742, -0.0495895594, -0.0151821338, 0.0357026979, -0.0589110889, 0.0405667238, -0.0391607247, -0.0183787942, 0.0244931243, 0.0249196682, -0.0583576411, 0.084158577, -0.0054794531, 0.0068721687, -0.0390326083, 0.0167068541, -0.0430599451, -0.019401066, 0.0557139181, -0.0132035334, 0.0029402489, -0.0107838735, -0.0071483655, 0.0466953963, 0.0454416387, 0.0031817346, -0.0047094095, -0.0364944339, -0.0129552828, 0.0403885171, 0.0549423806, 0.0275875293, -0.0029457167, 0.0392335206, 0.0424408428, -0.0352505706, -0.0271777827, 0.0755461752, -0.0146175213, -0.0410485528, 0.0090897689, 0.075659588, 0.0076711923, -0.0476876572, 0.0107336743, -0.0327440463, 0.0887401551, 0.0034329789, -0.0442018583, -0.0230517071, 0.0123872217, 0.0380045362, -0.0022624012, -0.0138489855, -0.0094630206, 0.0025489696, -0.0817268938, -0.020662548, 0.032289803, -0.0659126118, -0.0157861672, -0.0603747405, -0.0150022572, 0.0114245722, 0.0066278949, 0.0589837059, -0.0392024815, 0.0253468044, 0.0507110693, -0.0613977127, -0.0214338526, 0.0283146035, -0.0107170437, 0.0421702974, 0.0458932221, -0.0180952251, 0.0180945471, -0.0660913661, 0.0122547457, -0.0479949377, 0.0227606334, -0.0103923092, 0.0278350823, 0.0180381872, 0.0220388621, -0.0453908928, 0.0307717714, -0.0351690762, -0.0180435311, 0.0068187951, 0.0878808424, -0.0114495112, -0.008305693, -0.0008738612, -0.013804893, -0.0287989657, 0.0890121534, 0.0672698021, 0.008222403, 0.0077293958, -0.0068591805, 0.0394849591, -0.0329969339, -0.040139813, 0.0043372228, -0.0291899219, -0.0240601804, -0.0984917879, -0.0435063094, 0.0043421099, 0.0202816781, 0.0297738239, 0.0559199043, 0.0112115592, -0.060048271, -0.0516548567, 0.003588459, -0.0523311719, -0.0081293834, -0.0617239773, -0.0499877818, 0.0286946986, -0.0311286151, 0.000442981, -0.0088866949, -0.0709355921, -0.0469534025, 0.0269322004, 0.0690722466, 0.0317300893, 0.0423779078, 0.0697188154, -0.0223479308, 0.0471179858, -0.0512611754, -0.002752058, 0.037885122, -0.0056730662, -0.0044647944, -0.0169102345, 0.0059284004, -0.0076110023, -0.0077058887, 0.0122015858, 0.0180673786, -0.0239249002, -0.0352323391, 0.0247802064, -0.0080101807, 0.0458447412, -0.0019291475, -0.0086970301, -0.005936611, -0.0168423504, 0.0285173934, 0.0334037915, 0.0176506937, 0.0099815056, 0.0321087353, 0.0332025997, -0.005910296, -0.0865213722, -0.0295067243, 0.0298510771, 0.0110345101, -0.0268683769, -0.0134967435, -0.0058643054, 0.0056652511, 0.0466726385, -0.0247362647, 0.029324621, -0.06180197, 0.0326366015, 0.0261872597, 0.0255253222, -0.0198270734, 0.0277543664, -0.0325901322, -0.0371514894, -0.039747376, -0.0161580946, 0.0057823821, -0.0275194757, 0.0291068405, -0.0711598098, 0.0463249423, -0.0552192777, 0.0497338697, -0.0597003996, -0.0034757094, -0.011604242, 0.006488469, 0.0284103286, 0.0279943701, -0.0036745595, 0.016711574, 0.1001646668, -0.0174233299, 0.0245965607, -0.0131078297, -0.0065506524, -0.0323607326, -0.05400341, 0.0533726253, 0.0221887771, -0.0250037946, -0.0543002337, -0.0532223135, -0.0520386547, 0.0483600013, -0.0279793702, 0.058204446, -0.0442179218, 0.0527799204, 0.0429021791, -0.0532421134, 0.0309324078, -0.0268049054, 0.04047212, -0.0058039767, -0.019490594, -0.0181750022, -5.698405026e-33, -0.00679253, -0.0058941008, 0.0594297536, 0.0242539924, -0.0205825716, 0.0148250945, -0.0463015325, 0.0570462085, -0.0016694841, -0.0453825481, 0.0066631832, -0.0428523943, 0.0201746896, 0.0259041972, 0.0607208908, 0.0014638044, 0.0128235184, -0.0099973567, -0.0320286416, 0.0043721679, 0.0607328825, 0.0349644348, 0.0487132296, 0.0127388416, 0.0363282077, -0.0335340872, 0.0439649001, -0.0201541372, -0.0454652868, 0.0649954528, -0.0499371402, 0.0321153961, 0.0512276664, -0.038076546, -0.0700193718, 0.0150426226, -0.0706627741, -0.0239146538, 0.0622177869, 0.0030143745, -0.0114460774, -0.0373875611, 0.0646689162, -0.033426892, -0.031342607, -0.0042753932, 0.027224917, -0.0356836319, -0.0976982936, -0.000360132, -0.0217240583, -0.0210910998, 0.032618992, -0.0350292325, 0.0088173589, 0.0030190367, -0.0215513036, 0.0371640436, -0.0603735521, 0.0715049058, -0.0192519855, 0.080568105, 0.0314284116, 0.0509696789, 0.0258853119, 0.0573869869, 0.048503194, -0.0365242809, 0.0056498991, 0.0050357538, 0.0350710377, 0.0015690298, 0.0225290246, 0.0038087687, 0.0805866793, -0.0103470739, 0.0194115359, 0.0154457763, 0.0056278063, 0.0309772156, 0.0379895493, 0.0276458897, -0.0252969991, -0.0321873687, -0.023350453, 0.0332625732, -0.028856013, -0.0206585508, 0.0199696198, 0.0030773082, 0.01478111, -0.0312114004, 0.0605429225, 0.0079152854, -0.0034002757, -0.0262541547, -0.0048912615, -0.0022524248, -0.0142807607, -0.007690568, -0.0703209788, -0.0028250183, 0.0014714735, -0.0052237092, -0.0239546653, 0.0537225865, 0.0166382995, 0.0188261978, -0.0441939645, -0.0229655821, -0.0338881463, -0.0403549448, 0.0222418178, -0.0227267388, 0.0428084917, -0.0003768174, -0.0058190287, 0.0359925292, -0.02277321, -0.0712836087, -0.0271902401, -0.0011250427, -0.062387336, 0.0315273069, -0.0198065564, -0.0082494989, -0.0264039952, 0.0149402115, 0.0454606973, 0.0175747797, 0.0340614207, -0.0138577055, 2.371e-7, -0.0019806609, -0.0057415352, 0.0174049735, -0.017143961, 0.0342964008, -0.0038826261, -0.054498788, -0.0079286611, 0.0300936792, -0.0324397758, 0.0196980163, 0.0444137603, 0.0504530966, -0.0291144028, -0.0483121909, -0.0235947296, -0.0470330268, 0.0214264989, -0.0196201392, 0.0120975161, -0.0020869568, 0.0258573629, 0.0161520373, -0.0259676743, -0.0133841792, -0.0146851679, 0.0642909333, -0.0545063913, -0.0086770784, -0.0224457625, 0.0467158034, 0.0284263548, 0.0232353806, 0.0293174051, -0.0036703392, -0.0206485055, -0.0017175542, 0.0146992654, 0.0380424075, -0.0595624149, 0.0054466212, 0.0062772674, -0.02719176, -0.0169072878, -0.0001537177, -0.0084019089, -0.0067219902, 0.0169846676, 0.0016689888, 0.0292834751, 0.0207677949, 0.0307486523, -0.0171523411, 0.0110446448, -0.0241228361, 0.0412067696, -0.0003080792, -0.0481240861, 0.0052502109, 0.0111498898, -0.054840114, -0.0601724461, -0.0249340851, -0.0303138886, 0.0230600461, -0.0733152851, -0.0447654314, 1.664586785e-34, -0.002074759, -0.0188425146, 0.0482756719, 0.0585234836, -0.0079895994, -0.0649355799, -0.0089699263, 0.0714004487, 0.0190412384, -0.0578757375, -0.0298308972 ]
http://arxiv.org/abs/2406.17744v1
http://arxiv.org/pdf/2406.17744v1
Following Length Constraints in Instructions
2024-06-25T17:29:52
2024-06-25T17:29:52
Weizhe Yuan, Ilia Kulikov, Ping Yu, Kyunghyun Cho, Sainbayar Sukhbaatar, Jason Weston, Jing Xu
cs.CL
cs.CL
Aligned instruction following models can better fulfill user requests than their unaligned counterparts. However, it has been shown that there is a length bias in evaluation of such models, and that training algorithms tend to exploit this bias by learning longer responses. In this work we show how to train models that can be controlled at inference time with instructions containing desired length constraints. Such models are superior in length instructed evaluations, outperforming standard instruction following models such as GPT4, Llama 3 and Mixtral.
Following Length Constraints in Instructions: Aligned instruction following models can better fulfill user requests than their unaligned counterparts. However, it has been shown that there is a length bias in evaluation of such models, and that training algorithms tend to exploit this bias by learning longer responses. In this work we show how to train models that can be controlled at inference time with instructions containing desired length constraints. Such models are superior in length instructed evaluations, outperforming standard instruction following models such as GPT4, Llama 3 and Mixtral.
[ 0.0547591336, 0.026830608, -0.0030776972, 0.0122395763, -0.011256773, 0.0137352673, 0.0047924113, -0.003930314, -0.0043463842, -0.0281171557, -0.0217548218, -0.0581691377, -0.0092038121, -0.0192984901, 0.0044334265, -0.0772672743, 0.0421232916, -0.0250414666, -0.003878773, 0.0091280183, 0.0266778655, -0.0020609559, -0.0251885019, 0.019900769, 0.0051579094, -0.0201220661, -0.0629594028, 0.0225042962, -0.0176677871, -0.0394131504, 0.0337836593, 0.0422224663, 0.0162983593, 0.0106620258, 0.0000017722, -0.0395868272, 0.0022149901, -0.015159376, -0.0609531216, 0.0436426774, 0.0056137117, -0.0011182426, 0.0088779982, 0.0343738869, -0.0201310571, 0.0371109657, 0.0563199595, 0.0533879139, 0.0404976159, 0.0689573586, 0.011047177, -0.0239136834, 0.00708126, -0.0679877773, 0.0403513499, -0.0384130068, -0.0089470344, -0.0212658569, 0.0480699465, 0.0522762798, -0.0369908474, -0.0081061451, 0.0466685295, 0.0306865573, -0.0286028497, 0.0158016142, 0.0136066852, -0.0009810246, -0.0459891334, 0.0719851479, 0.0011894058, 0.0223736148, 0.0117878634, -0.0349588245, 0.0319846347, -0.0208327342, -0.0284614004, -0.0172228143, -0.0267132632, -0.0086554214, -0.0204481818, 0.0385137312, -0.0137412157, -0.0514744781, -0.0023816666, 0.0699894503, -0.0137234619, -0.0619359612, -0.0190510582, -0.0229128897, 0.0096648037, -0.0311194621, -0.0262407903, 0.0562995449, -0.033257246, 0.0213231165, -0.0261611715, 0.0250506289, 0.0297110435, -0.0308744125, 0.0470281802, 0.0628958493, -0.0072626104, 0.0060236203, -0.023282513, 0.0869209766, -0.0253057685, 0.0184792485, -0.0651944652, 0.0880911946, -0.0400717817, -0.0023066278, -0.0348854251, 0.0233664121, -0.0591067448, -0.0064630085, -0.0557343662, 0.0449015684, 0.0028942132, 0.032716047, 0.0190315284, 0.0097998977, -0.0491165742, 0.0070654578, -0.0525146164, 0.0253489725, -0.0593960062, 0.0034760777, 0.002959426, -0.0238218587, -0.0129028196, 0.000665235, -0.0201768205, -0.0228320509, 0.0258407127, 0.1005897969, -0.0054091751, -0.0116140069, -0.025858514, -0.0362492092, -0.0089597031, -0.0528324507, 0.0214640256, -0.027357582, 0.0396924987, -0.0161319412, -0.0057728798, 0.0161003117, 0.0080228364, 0.0406833403, -0.002799947, 0.0623783991, 0.057284195, 0.0005098333, -0.0133999996, 0.0225373935, 0.0121477731, 0.0442411266, 0.0238386691, 0.0561820976, 0.0327056199, -0.0318367518, 0.0339926369, 0.039448943, 0.0116970455, -0.0048328936, 0.0286339391, 0.0265575424, 0.0078321248, -0.0094396109, 0.0440551527, 0.0477864891, 0.0038578007, -0.0442875735, 0.0257037189, 0.0468816981, 0.0298995059, -0.0332631208, 0.0813705996, 0.0006638647, -0.0379132368, 0.078667447, -0.072601527, 0.0347481184, 0.003274566, 0.0493115522, 0.0267182067, 0.0032483856, -0.0458174124, -0.0498444587, -0.0120873498, -0.0201330874, 0.0006179293, 0.0079504224, 0.0080780452, -0.0426694453, -0.0889063925, 0.0191085413, 0.0125814378, -0.0209922828, -0.0421628915, -0.066339381, 0.0109479297, 0.0872298926, 0.0095185647, -0.0233619008, -0.0375082158, -0.0100022787, 0.0012262763, 0.0195762366, 0.0333995149, -0.0219691936, 0.0272160526, -0.0365811102, -0.0342939571, -0.0287031289, 0.0063378466, 0.0252293143, 0.024204677, -0.0140872775, 0.0366983078, 0.0029743612, -0.0043301512, 0.0323336571, -0.0005545683, -0.1076067239, -0.0436244421, 0.0368498527, 0.0364539996, -0.0050237472, 0.0149654113, 0.0113127697, 0.0299079325, -0.0102171721, 0.0395411104, -0.0323248431, -0.0787420347, -0.0342358686, 0.021765653, -0.000321091, 0.0476997681, 0.0093040243, -0.0373909585, -0.0277664643, 0.0089185331, -0.049742844, 0.0432868451, 0.0393858626, 0.0191996396, 0.0304277409, -0.0315996408, -0.0497907512, 0.0326822065, -0.0402565561, 0.0192725472, 0.0321685039, -0.0180684309, 0.0212540869, -0.064759396, -0.0018237151, 0.0133450041, 0.0001659818, -0.0221906248, 0.0188233294, 0.0153597519, -0.0174016245, -0.0305258427, -0.0816256255, 0.0188310463, -0.0115099037, 0.0456958562, 0.0034688746, 0.0243281275, -0.0130355759, -0.042480953, -0.0401519313, -0.0669034272, 0.002946411, 0.0059771626, 0.0103260726, 0.0535546951, -0.0771089494, -0.0505495556, -0.0210959651, -0.0280778501, -0.0040313345, -0.0197610334, -0.0067317877, -0.0640308186, 0.0010943174, 0.0097903376, 0.0555444434, -0.031667985, -0.0476627387, 0.0261303429, 0.0149897765, 0.0330469497, 0.1064897999, 0.0009608476, 0.041462291, 0.0066007604, 0.006065595, -0.040780928, 0.0594900325, -0.0397712439, -0.031059185, -0.0094736172, -0.048212219, 0.0232648123, 0.0820238516, -0.0006352683, 0.0387527198, -0.0887127221, -0.0390306413, 0.0206679199, -0.0094422372, 0.0194658674, -0.0169560462, -0.0392881297, 0.0092308093, -0.0316219367, -0.0928780138, 0.0612648465, -0.054409463, 0.0014291082, -0.0200746134, -0.0056965598, -0.0137932496, 0.0571601391, 0.0263809934, 0.0043381695, -0.0138500379, -0.0142870992, -0.0462909266, 0.0171023812, 0.0172188208, -0.0382482558, 0.0318447426, -0.0288037099, -0.0188082978, -0.0325391255, -0.0239476468, -0.0292356946, -0.0460814312, 0.0248635802, -0.0189192519, 0.0437341332, -0.0104816724, -0.0055668484, 0.0224843211, 0.015802715, 0.0421594866, -0.0295695066, 0.0197287444, -0.0311283953, -0.0476171337, 0.0581241325, -0.0148249445, -0.0462993421, -0.0557188019, 0.0201528054, 0.0232900232, 0.0361844152, -0.0166620426, -0.0370780863, 0.0053553726, 0.014656825, -0.0315549038, -0.0155701106, 0.0742675513, 0.0870371088, -0.0316779576, 0.0248600002, 0.082797192, 0.0267625134, -0.0302266944, 0.0197492782, -0.01155023, 0.0409023948, 0.0181277413, 0.0444479622, -0.0779809356, -0.0284879319, -0.0122839483, 0.0609829687, -0.001141596, 0.0105226794, -0.0201325472, -0.1127380952, -0.0033670394, 0.0039057154, -0.0166122541, -0.0532690585, -0.0261940286, -0.0085913604, 0.0125543987, 0.0112278592, 0.0216789246, -0.0478988402, 0.0026869194, 0.0576262474, -0.0164174642, -0.0224494915, 0.0342849456, -0.011585501, 0.0646040514, 0.0774799734, -0.0332638398, 0.0279810093, -0.0433490388, 0.011440455, -0.0171995983, 0.0356572531, 0.0040300069, 0.061556235, -0.0233273935, 0.0146984737, -0.0140911518, 0.0407056548, -0.0043468117, 0.0260559674, 0.0300265588, 0.0499537252, -0.0147160925, 0.0093259513, -0.0378110744, 0.006540745, 0.0348810889, 0.0463701636, 0.015512039, -0.0358672999, -0.048030246, -0.0357514545, -0.0363990776, -0.0705287978, 0.0328095146, 0.0052267853, -0.0564363748, 0.0119944299, -0.033367496, -0.0745992735, -0.0061691618, 0.0515521392, 0.0272128265, -0.0182629246, -0.0032405837, 0.0109002115, -0.0134026008, 0.0032075755, 0.0134875104, 0.0330154151, -0.0982082114, -0.0236415137, 0.0095343208, -0.0111311907, 0.0108257197, 0.0107951155, -0.088333793, -0.0656496584, -0.0396759883, 0.1065542325, 0.0289581437, 0.0652515814, 0.0147537515, 0.0056224153, -0.0425310768, -0.0029979974, -0.0256212559, 0.016026428, 0.0310832672, 0.0063290703, -0.0215724893, -0.0207807999, -0.0446137749, 0.0422836095, 0.0066408631, 0.011614236, -0.0541088954, -0.0348374657, 0.019396523, 0.0021286849, -0.0178605095, -0.0144353621, -0.0092766024, 0.0126336552, 0.0729631484, -0.0165144764, 0.0159831606, -0.0223045927, 0.0056671854, -0.0104058823, -0.0250033978, 0.019698197, -0.0781973675, -0.0446623564, -0.0183670484, -0.0614188984, 0.0523647033, -0.0088474602, 0.0047818865, 0.0155652193, 0.0061615431, -0.0248564277, 0.0244759899, 0.0095971385, 0.055071827, 0.0207281727, 0.0799008012, 0.0273882169, 0.0437122881, -0.0320938006, 0.0403442234, -0.0223627407, 0.004942535, -0.0455135815, -0.0254159719, 0.0138508985, -0.0427446216, -0.016183177, 0.0371395536, 0.0384222306, -0.0193689689, 0.0556934439, -0.0078254575, -0.009275428, 0.016936373, 0.0535639301, 0.0180995315, -0.0404519141, 0.0758239105, 0.0152233345, 0.0499734096, -0.0160827525, -0.018922409, -0.063043341, -0.0374672972, 0.0263969433, 0.0149839884, -0.0032347883, -0.0063487636, -0.0174935888, -0.0322147347, -0.0347607508, 0.0272962842, 0.0466343611, -0.021915134, 0.0660392269, 0.0385114141, -0.0068858415, -0.0689945891, -0.0008863414, -0.0421900749, -0.0087036509, 0.0058569536, -0.0493831933, -5.499216406e-33, -0.0296285134, -0.0020107545, 0.0457201973, -0.0078873783, -0.0432191417, -0.0086276662, -0.0094763506, -0.0007928602, 0.0015483897, -0.042149581, -0.0058616786, -0.0210440997, 0.0234062616, -0.0159866624, 0.0224527922, -0.0149193723, 0.0432872288, -0.0173171274, -0.0605230927, 0.0236596651, 0.0120979575, -0.017329162, 0.0226688553, -0.0221017245, 0.0290189516, -0.00865799, 0.0145926988, -0.0089180646, -0.0427253954, 0.0901856497, -0.0009751751, -0.0043500769, 0.0391562134, -0.024507273, -0.006190144, 0.0516868792, -0.032682322, -0.0371279195, 0.0359581113, 0.0215010773, -0.0364400558, -0.0438754335, 0.0132544935, -0.0050036982, -0.0341528095, -0.0026997658, 0.0281200837, 0.0126779582, 0.0179009326, 0.0271542333, -0.0554227754, -0.0060152188, 0.0019289033, -0.0584004149, 0.0077808658, 0.0605237894, 0.0262075216, -0.0206652377, 0.0045088688, 0.0015374662, 0.0641283095, 0.0704763681, 0.0516081378, 0.0472306907, -0.008039332, -0.0047209035, -0.0288143065, -0.0084146177, -0.0371370316, -0.0098906076, -0.0237176586, 0.0173854679, 0.0165075287, -0.0264626518, 0.0598077849, -0.0027115252, -0.029406026, 0.0355995633, 0.0257095098, 0.0584999807, 0.0160185136, 0.0064448644, -0.0965977907, -0.008454835, -0.002510211, 0.0146990679, -0.0039093876, -0.0627394766, 0.0689534992, 0.0520737469, -0.0059044277, 0.007186017, 0.0656661242, 0.025602337, -0.0105039375, -0.0115723247, -0.0033038876, 0.03057882, -0.0201293975, -0.0154882157, -0.040415924, -0.0364309251, 0.0488714017, 0.0005481715, 0.0222009178, 0.0053335326, 0.0109511325, 0.0322041214, -0.0619177334, 0.0013999338, -0.0307486597, 0.0211017821, 0.0024278401, 0.0207582712, 0.0439736694, 0.0049529322, -0.0254739486, 0.0768847167, -0.0647577122, -0.0685037822, -0.0433811881, -0.0280661099, 0.0179504268, 0.0231819041, -0.0503422618, 0.0028855142, -0.0106888153, 0.0204736013, -0.0520959422, -0.0455548577, -0.0109632369, 0.0251281429, 2.54e-7, -0.0078487117, -0.0096171219, 0.086513862, 0.041414313, -0.0153256375, 0.0089356555, 0.0165675934, 0.0255414285, -0.0156280492, -0.0355128199, 0.0099359676, 0.0133622652, -0.0040469728, -0.0133439526, -0.0428534523, -0.0051576965, -0.0424132273, 0.0475790985, -0.0454269238, 0.0005398868, 0.0677705631, 0.0580672584, -0.0008515508, 0.0321142375, 0.0050605, 0.0093149003, 0.0044561038, -0.0078592002, 0.0219557714, 0.0021106235, 0.0202584192, -0.0282525551, -0.0177231263, 0.0412731878, -0.0111933863, 0.0101622269, -0.0371062867, 0.0898389667, -0.0232647993, -0.0282765757, 0.0175014399, 0.0005216917, -0.0153051969, -0.0203328766, 0.0394347608, 0.0073851175, 0.0153308539, 0.0046955864, -0.0181353688, -0.0148154851, -0.0234463047, 0.007154264, 0.0354758054, -0.0004059112, 0.0188553929, -0.0327749401, 0.0101914741, -0.0585736595, 0.0114918984, 0.0224293321, -0.0289943572, -0.0458670855, -0.056472294, 0.0108737135, -0.0076037217, -0.0818823799, -0.0485118441, 2.350899621e-34, 0.0097311875, -0.0501372293, 0.0498563461, 0.0349404849, 0.0094233947, -0.0152918519, -0.0184490401, 0.0466655232, 0.09389323, -0.0747403577, -0.0513251945 ]
http://arxiv.org/abs/2406.17739v1
http://arxiv.org/pdf/2406.17739v1
Find Parent then Label Children: A Two-stage Taxonomy Completion Method with Pre-trained Language Model
2024-06-25T17:25:02
2024-06-25T17:25:02
Fei Xia, Yixuan Weng, Shizhu He, Kang Liu, Jun Zhao
cs.CL
cs.CL, cs.AI
Taxonomies, which organize domain concepts into hierarchical structures, are crucial for building knowledge systems and downstream applications. As domain knowledge evolves, taxonomies need to be continuously updated to include new concepts. Previous approaches have mainly focused on adding concepts to the leaf nodes of the existing hierarchical tree, which does not fully utilize the taxonomy's knowledge and is unable to update the original taxonomy structure (usually involving non-leaf nodes). In this paper, we propose a two-stage method called ATTEMPT for taxonomy completion. Our method inserts new concepts into the correct position by finding a parent node and labeling child nodes. Specifically, by combining local nodes with prompts to generate natural sentences, we take advantage of pre-trained language models for hypernym/hyponymy recognition. Experimental results on two public datasets (including six domains) show that ATTEMPT performs best on both taxonomy completion and extension tasks, surpassing existing methods.
Find Parent then Label Children: A Two-stage Taxonomy Completion Method with Pre-trained Language Model: Taxonomies, which organize domain concepts into hierarchical structures, are crucial for building knowledge systems and downstream applications. As domain knowledge evolves, taxonomies need to be continuously updated to include new concepts. Previous approaches have mainly focused on adding concepts to the leaf nodes of the existing hierarchical tree, which does not fully utilize the taxonomy's knowledge and is unable to update the original taxonomy structure (usually involving non-leaf nodes). In this paper, we propose a two-stage method called ATTEMPT for taxonomy completion. Our method inserts new concepts into the correct position by finding a parent node and labeling child nodes. Specifically, by combining local nodes with prompts to generate natural sentences, we take advantage of pre-trained language models for hypernym/hyponymy recognition. Experimental results on two public datasets (including six domains) show that ATTEMPT performs best on both taxonomy completion and extension tasks, surpassing existing methods.
[ 0.0416950844, 0.0355267748, -0.0172756966, 0.0041202521, -0.0252685212, -0.0118327141, -0.0441421755, 0.0247969981, -0.0021549729, -0.0903560221, -0.0187542941, -0.0174772274, 0.0044078487, 0.0076411343, -0.0121472916, -0.0710083246, 0.0509655438, -0.0018890146, 0.0137187587, -0.0128507456, -0.0349610746, 0.0282886717, 0.0062614693, 0.015587802, -0.036204014, -0.0263139568, -0.0389848538, 0.0194722041, 0.0017319804, -0.0500302427, 0.014150966, 0.041450005, 0.0201083384, 0.0777638331, 0.0000017823, -0.092423588, -0.037676692, -0.0309792683, -0.011051923, -0.0317522846, 0.0441919565, 0.0008453475, 0.0307447389, -0.0253431927, 0.003567341, -0.030974485, 0.0336745605, 0.0247527696, 0.0224696416, 0.0057899808, 0.0129904216, -0.0529947244, 0.0334860831, -0.0047068391, 0.0028746142, -0.045625031, 0.0171818081, 0.0352959521, 0.0113538364, 0.0280532744, 0.0012554314, 0.0086603891, -0.017359674, -0.0313589387, 0.0189223997, 0.0010596069, 0.0173195004, -0.0240433086, -0.0088426927, 0.0728395283, -0.049407687, -0.0013639948, 0.0164734386, -0.015802918, -0.0043265028, 0.0346213579, -0.0162775479, 0.0212253127, 0.0004136738, -0.0364396051, 0.0277924649, -0.0392591394, 0.0455115661, 0.0289867334, -0.0093186637, 0.1065070778, 0.0483163744, -0.0221142136, 0.0032725723, 0.009137678, -0.0000429068, -0.0855170637, 0.0181249436, 0.0625716597, -0.020717049, 0.0614954904, -0.059825711, 0.0416123718, 0.0267861858, -0.037095204, -0.0265319459, -0.0050107255, 0.0067016589, 0.0519218594, -0.0202281736, -0.0398138575, 0.0161549356, -0.0105687631, -0.048834499, -0.0161423385, -0.0651840493, -0.0263558663, -0.0641713217, 0.0542701669, -0.0366426744, -0.0209308155, -0.0449253023, 0.0150613599, 0.0033013483, 0.0069364212, -0.0050575356, -0.0279275123, -0.0355482846, -0.0305770934, -0.0402533971, -0.010498873, -0.0449412689, 0.0617589019, 0.0220469441, 0.0318658538, 0.0172224473, 0.0240024906, 0.0475510284, 0.0327374041, 0.0015108775, 0.039324265, 0.0592288077, -0.0385026634, -0.0691844225, -0.0510645695, -0.0155461831, -0.0160229411, 0.0095009981, -0.0634508505, 0.0039268374, -0.0392924882, -0.0095765954, -0.0032784047, 0.0265811402, 0.0176650994, -0.0460140631, 0.0767250359, -0.095102258, -0.0081698578, 0.0739182606, -0.0031821893, 0.0112225395, 0.0304633938, 0.0216307845, 0.030359121, 0.0799278468, -0.0046383073, 0.0080910353, 0.0045025381, 0.0002571727, -0.0117380675, -0.0009531518, -0.0342481881, -0.0454619192, 0.0609103516, -0.0150593501, 0.0312949456, 0.0134411594, 0.0351252109, 0.0326781571, 0.0501551777, 0.0296991877, -0.009748186, 0.0725807101, -0.0097469902, -0.0367274806, -0.0128325112, -0.0309683755, -0.0257866401, -0.0616703779, 0.0307889488, -0.0142772049, 0.0112160556, -0.0151671134, -0.0443329029, -0.0223228447, -0.0279730167, -0.0241007563, -0.021264879, 0.0927860364, 0.0594816357, 0.0096174078, 0.0406065993, 0.0636410788, -0.007165438, 0.0604081191, -0.00931066, 0.0413565151, 0.0636132658, 0.013291507, -0.0650144592, -0.0404372849, 0.0096876584, -0.0107741375, 0.0442067087, -0.0126105594, -0.0494299978, -0.0387718081, -0.0170699023, -0.0273470487, 0.030236721, 0.0165758058, 0.0274637621, 0.0117499456, 0.0146489944, 0.0523665845, -0.063805379, -0.0288771149, 0.0462570377, -0.0206143092, 0.0021149775, 0.0112323267, 0.0271902382, -0.0183426272, 0.0380876213, -0.0133924028, -0.0007367375, 0.0216055624, 0.0229838472, -0.0752818808, -0.0193575881, -0.0095691606, -0.0561771803, 0.0483629145, -0.0034406951, 0.0033156872, -0.0149065442, -0.0206237175, -0.0283814501, -0.00095241, -0.0056568412, 0.0481552519, -0.0490790084, 0.0568140559, -0.0039896709, -0.051997114, -0.0304145813, 0.0898912996, 0.019512916, 0.0449750461, -0.0601390079, -0.0012632323, 0.0010624563, 0.028458355, -0.0270516984, 0.0205305051, -0.0317354351, 0.0258516222, 0.0324155986, 0.0154004758, -0.0586382672, 0.0317687057, -0.0604432933, -0.0070258318, 0.0163412821, 0.0265012421, 0.0029126045, -0.0006257475, -0.0074098292, -0.0293438081, 0.0169974193, 0.077196151, -0.0151786525, 0.0157546178, 0.034097977, -0.0013828805, -0.0457711369, -0.0280791484, -0.003508572, -0.0091777677, 0.0505562201, 0.036030855, 0.0113223093, -0.0137414932, -0.0122582763, -0.0665590242, 0.0239050686, 0.0418828428, 0.0087847393, -0.0142830731, -0.0049616699, 0.0025197295, -0.0294775553, -0.011234357, -0.0201767348, 0.0633453652, 0.0188479926, 0.0002997391, -0.0260951221, -0.0407428816, -0.0224383418, 0.0196954142, 0.0228360947, 0.0277287737, 0.0938919485, -0.0262527056, -0.035447944, 0.0031202752, 0.0180886369, -0.03427146, 0.0317110904, -0.0594294034, -0.0350252688, -0.014733444, 0.0711568743, -0.0091090919, -0.0532673411, 0.0667503625, -0.0341917984, 0.0114686824, -0.0318757072, -0.0100645153, -0.0198696889, 0.0803416595, 0.0117678335, 0.0596237332, 0.0868442804, -0.0922102705, -0.0332004428, 0.0451253355, 0.0105763227, 0.0009746773, 0.0334092863, -0.0500633195, 0.013479813, 0.019539332, 0.0273284074, -0.0380179919, -0.022200346, -0.0076133832, 0.0293399971, 0.0001749806, -0.0274163503, 0.0231047105, -0.0167683456, -0.0106162177, 0.0731428415, 0.029856978, -0.0297927503, -0.0376334824, 0.06831377, 0.0445977524, -0.0043790732, -0.0160889011, 0.018576568, 0.0200854708, 0.0116318073, -0.0179573856, 0.0016973224, -0.0036289725, -0.0032925433, 0.0218650643, -0.0944198146, -0.052515313, 0.0126157263, -0.0284190942, -0.0131001277, 0.0056012999, 0.1174455062, 0.0700430945, -0.0172382798, 0.0264023878, -0.0216211006, 0.0210222974, 0.0558582172, -0.0065479935, -0.0816676468, 0.0139065692, 0.0238331296, -0.0088724038, 0.0243034158, -0.0397045538, 0.0300303288, -0.1048875153, -0.0056968746, -0.0085227992, -0.0549402535, 0.013929517, -0.0212740395, -0.0081936568, 0.0614541993, -0.0279534943, -0.0046601156, 0.0061366912, -0.0108103156, -0.0106286481, -0.0161515158, -0.0401866287, 0.0641755387, -0.041619394, 0.0124500738, 0.054431051, 0.0409575477, -0.0406168997, -0.0000311542, 0.019816706, -0.0141883772, -0.031586621, -0.0065320381, -0.0075653978, -0.0094996979, 0.0122873466, 0.0101698246, -0.0129195871, -0.0455755815, -0.0055294726, 0.0480960868, 0.1031811684, 0.0220134631, -0.013972979, -0.0147005841, -0.0039030677, 0.0093316343, 0.0559468977, 0.0121572101, -0.0342465751, 0.0286401045, -0.0175536908, -0.0082362778, 0.0247282181, 0.0170000177, 0.0302142855, -0.0729932487, -0.0108603146, -0.1002536565, -0.0435816459, 0.0598264448, 0.0392035618, 0.0266898256, -0.0001446654, 0.0133311329, 0.0101346327, -0.0504929163, 0.034913145, -0.0215179399, 0.0015958252, -0.0240586121, -0.0211566556, 0.0612449124, -0.0108097997, -0.0487076864, -0.0057916455, 0.0410139859, -0.0389171019, -0.0438460186, 0.0138592301, 0.0140058529, 0.0144848153, -0.0423793346, -0.029352041, 0.0214151423, 0.0089468025, -0.0324871503, -0.0137804402, 0.0094816424, 0.0335246921, -0.0134161441, 0.0484023169, -0.0569392964, 0.0408921726, -0.0032589277, 0.0576933548, 0.0125144804, -0.0583740883, 0.0423119217, -0.0040538502, 0.0765933841, -0.0595418289, -0.0049383771, -0.0190399569, -0.0094650034, -0.0158678032, -0.061346937, -0.0016646605, 0.0013354396, 0.023395123, 0.0219519585, 0.0188014563, -0.0276941527, -0.0329455733, 0.0161556993, -0.0007810433, -0.0519553944, 0.0068542748, 0.0486084633, 0.0139794834, 0.0053617293, -0.0181947872, -0.0481087416, -0.0319268666, 0.0065018386, -0.0691377297, 0.0230578389, -0.0102924081, -0.0187538471, -0.0715842098, 0.0019544293, 0.0058574663, -0.0197963808, -0.0436962284, 0.0340125449, 0.0773245618, 0.0201945044, -0.042426236, -0.0411010198, -0.0291897487, -0.0081423791, 0.0192302205, -0.0002085413, -0.0900573805, 0.0365140699, 0.015866451, -0.0408526435, -0.004666422, 0.0589995645, 0.0397690013, 0.0357645303, -0.0115942201, -0.0076097581, -0.0019495105, 0.0057771443, -0.0187783446, 0.0642958581, -0.0466069989, 0.0278453287, -0.0454056337, -0.0745211765, 0.0031397187, 0.0181656778, -0.0117548471, 0.0472409129, 0.0664808154, 0.065827325, 0.0079341233, -0.0266258772, -0.035848897, 0.0086538391, -0.0102259479, 0.0051218788, 0.0098065911, -5.48316099e-33, -0.0606188625, -0.0722756311, 0.015931543, -0.0121574188, -0.0011987201, -0.0003318224, -0.0300740749, 0.0454742201, 0.0155793903, 0.0401504375, 0.0293309633, -0.0105694793, 0.0033655127, 0.0379551426, 0.0311429277, 0.0214894302, 0.0122338254, 0.009717199, 0.0074967169, 0.0139634181, 0.0225860346, 0.0279138908, 0.075700745, -0.0392305292, 0.014500415, 0.0046979478, 0.0216531213, 0.0198670868, 0.0011204216, 0.0406598151, 0.0009076038, 0.0772267208, 0.0119596301, -0.011570801, 0.0105550792, 0.0102598462, -0.0679410249, -0.029270418, -0.0088769011, -0.0177023262, -0.0461680777, -0.0181755163, 0.0237388834, -0.0015921127, 0.0278697945, 0.0156212514, 0.0424591899, -0.0197625719, -0.0570088066, -0.0469356999, -0.1083008647, 0.0124152768, 0.0239443742, -0.0050705103, -0.0103289243, 0.0240940247, 0.0100630084, -0.0029769642, -0.0226642005, 0.0367540233, 0.0409559719, 0.0160761718, 0.0314428098, 0.0412328839, -0.0264029149, 0.0483785905, 0.0023637768, 0.0075008422, 0.0087198382, 0.0544813573, 0.0181963407, 0.0319157243, -0.004194384, -0.0101405075, -0.0090867486, -0.014809181, -0.0241002124, 0.0117868641, -0.0505558737, -0.0107830092, -0.0364463516, 0.0125488937, -0.0075151566, -0.0138001572, -0.0559978783, -0.0074082189, -0.0409343578, -0.0525302887, 0.0128754647, 0.0033197319, -0.0099844206, 0.0181694925, -0.0413754731, -0.0016324811, -0.0575368665, -0.0284532383, 0.0293838326, 0.0401785187, -0.0380912945, 0.0699756294, -0.0196519755, -0.0768764094, -0.0095968517, -0.0231264234, -0.0188387334, 0.007075638, -0.0126971733, -0.0014614888, -0.0470059402, 0.0243471432, 0.0127537129, -0.0542446934, -0.0015864416, -0.0269195698, 0.0486155041, 0.0136342393, 0.0073826425, 0.0602919906, 0.0498780124, -0.0310181603, -0.0336920805, 0.036542587, -0.0075587188, 0.0135719515, -0.0062256395, -0.0081755584, -0.0080607375, -0.0280669872, 0.020890709, -0.0297254622, 0.0008767958, -0.0325483531, 2.606e-7, 0.0171873476, 0.0490323193, -0.0240897834, 0.0076599168, 0.0152825555, -0.0080395564, 0.0313607901, 0.0228096545, -0.0328974836, -0.0430661589, -0.0106421988, -0.0722049624, -0.0035960607, 0.0035417469, -0.1092090905, -0.0382921584, 0.011985166, 0.0013681545, -0.0844780058, 0.006183987, 0.0598282889, 0.0788267776, 0.0156562645, 0.0021344, -0.0433655865, -0.0488353707, 0.0346694812, 0.0131515162, -0.0029702012, 0.0572510548, 0.004696779, 0.0610029772, -0.0000748618, 0.0345971473, -0.0349886566, 0.0391532853, 0.016214788, 0.0263506025, -0.0240165982, 0.0151451174, -0.004531431, -0.0130490223, 0.0327499285, -0.0682788715, 0.0333611518, -0.0610891506, 0.0091783516, 0.019514136, -0.0206489842, 0.0050584031, -0.0378971547, -0.0469167344, -0.0099832062, 0.0023273742, 0.0076512075, 0.0235503521, -0.0129322428, -0.0949984118, 0.0345759541, 0.0768938661, -0.0431291163, -0.0312617682, 0.0349239744, -0.0210400335, 0.0021838231, -0.024492912, 0.0097478945, 2.152454364e-34, 0.0094201025, 0.0021329124, -0.0096861646, 0.053157106, 0.0250239596, -0.0081912884, 0.0180920977, 0.0153338648, 0.0295983404, -0.0532089621, -0.0012547713 ]
http://arxiv.org/abs/2406.17737v1
http://arxiv.org/pdf/2406.17737v1
LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users
2024-06-25T17:24:07
2024-06-25T17:24:07
Elinor Poole-Dayan, Deb Roy, Jad Kabbara
cs.CL
cs.CL, cs.AI, cs.LG
While state-of-the-art Large Language Models (LLMs) have shown impressive performance on many tasks, there has been extensive research on undesirable model behavior such as hallucinations and bias. In this work, we investigate how the quality of LLM responses changes in terms of information accuracy, truthfulness, and refusals depending on three user traits: English proficiency, education level, and country of origin. We present extensive experimentation on three state-of-the-art LLMs and two different datasets targeting truthfulness and factuality. Our findings suggest that undesirable behaviors in state-of-the-art LLMs occur disproportionately more for users with lower English proficiency, of lower education status, and originating from outside the US, rendering these models unreliable sources of information towards their most vulnerable users.
LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users: While state-of-the-art Large Language Models (LLMs) have shown impressive performance on many tasks, there has been extensive research on undesirable model behavior such as hallucinations and bias. In this work, we investigate how the quality of LLM responses changes in terms of information accuracy, truthfulness, and refusals depending on three user traits: English proficiency, education level, and country of origin. We present extensive experimentation on three state-of-the-art LLMs and two different datasets targeting truthfulness and factuality. Our findings suggest that undesirable behaviors in state-of-the-art LLMs occur disproportionately more for users with lower English proficiency, of lower education status, and originating from outside the US, rendering these models unreliable sources of information towards their most vulnerable users.
[ 0.0385592505, 0.0766503215, -0.0388005786, 0.0200502332, -0.0161413606, 0.011904249, 0.0549075082, -0.008985227, 0.0416241065, -0.0577895716, 0.0055669649, -0.0481768027, 0.0063178297, -0.0192792844, 0.0007331275, -0.0280754566, 0.0347205065, -0.0204158928, 0.0160220247, -0.0462333485, -0.0063822703, 0.0299924575, -0.0075221234, 0.0658017769, -0.0087618819, -0.029398134, 0.0380068459, -0.005426826, -0.0117420228, -0.0253775958, 0.0408060551, 0.0710853711, -0.0165564287, 0.0544663034, 0.0000019642, -0.0494065657, 0.0199376047, -0.0060038799, -0.0416780896, 0.038137719, 0.0424570665, -0.0036438005, -0.0004450391, 0.0469901264, -0.0326826312, 0.0101448763, 0.0212661866, 0.0431817174, 0.0182559825, 0.025360316, -0.0012094539, 0.000645722, 0.0165178217, -0.0029185528, 0.0260118991, -0.0049310718, 0.0232064743, -0.0003242029, -0.0235188585, -0.0469876789, 0.0178936999, 0.0373335034, 0.0017042145, -0.0006152194, -0.0348821878, 0.0225410126, -0.0516662896, -0.0421068482, 0.0082739471, 0.0499824062, 0.0033914116, -0.0154941687, -0.0043009249, -0.0140058985, -0.0019320018, 0.0331656188, -0.0235332642, 0.0372199453, -0.015610422, 0.0039980672, -0.0040369462, -0.0286649968, 0.0432984494, 0.0217237473, -0.0076274271, 0.0582721457, 0.0029602365, -0.0407273397, -0.001523249, 0.0046288087, 0.0315011628, -0.061642196, 0.0280755647, -0.0231216624, -0.0574373752, 0.0154352244, -0.0191536509, 0.0063612433, 0.0569959357, -0.0727650821, 0.0161906797, 0.0433334447, -0.0207864176, 0.0212776754, 0.0025894637, 0.0317027271, 0.0016637015, 0.0393501073, -0.0003154113, 0.0137629937, -0.0029421193, -0.0107962918, 0.0228742212, -0.0150777623, -0.02074017, 0.0231016744, -0.0555073284, 0.0193988383, -0.0110718273, 0.0146452831, -0.0212132372, -0.0022275182, -0.0133653451, -0.0313631594, -0.0755030736, 0.0650473237, -0.0202622525, 0.0293204747, 0.0108184656, -0.0800098553, 0.0334234573, 0.0020332541, -0.0362806097, -0.0323129967, 0.0025008291, -0.001440202, 0.0219909418, -0.0173068382, -0.0705942214, -0.0222001467, -0.023379432, -0.0678900257, -0.0280944835, -0.0622243024, 0.014774411, 0.0405702479, -0.0040146671, -0.075796783, 0.009101213, 0.0432154387, -0.0432841927, 0.0359873176, -0.0099735931, 0.029518187, 0.0414451137, 0.0299754683, 0.0081731118, 0.0679320022, -0.0002253518, 0.0340306647, 0.0319704376, -0.0508118384, 0.0286622047, 0.0417059548, -0.0083728367, 0.0402936935, -0.0135115357, -0.0075587495, -0.0228584725, -0.0001522456, 0.0403304771, 0.0374325365, 0.0058941096, 0.0147438403, 0.1142865941, 0.0498903096, 0.0163970999, 0.0095970612, -0.0249680467, 0.0421136245, 0.0103118252, -0.0079568382, -0.0647414848, -0.0147760557, -0.0291885529, 0.0014963843, 0.0478200205, 0.064314343, -0.0149041573, -0.0386890396, -0.0252313279, -0.0133423842, 0.0207195338, 0.0696633086, 0.0064307861, -0.0312831812, -0.0135445436, 0.0591498315, 0.054749392, -0.028822409, -0.0048241625, -0.0623375289, 0.0153972283, 0.0773385018, 0.022317091, 0.0041002845, 0.034803573, -0.0381500125, 0.0329640284, 0.0834963396, 0.0635315329, 0.0249533132, -0.017079955, -0.0396437831, -0.0150276767, 0.0327718742, 0.0114620086, 0.0118453363, 0.0192637574, 0.019658912, 0.0199150462, -0.0466710292, -0.0539469682, 0.0150594842, -0.0268938299, -0.0101134926, -0.0262549538, 0.0019491839, 0.0190708414, -0.0046402584, 0.0051523233, 0.0069470056, -0.0082673458, 0.0079514822, 0.001399031, -0.0172684416, -0.0522819534, 0.0123702111, 0.0240962934, -0.0131723499, -0.0003774491, 0.0212439634, 0.0162032805, -0.0044478141, 0.0184292998, -0.0224597398, 0.0640704259, -0.0367607139, 0.0321337432, -0.0194922872, 0.0218875632, -0.0291609094, 0.0630492046, -0.0550472364, 0.0343120769, -0.009302929, -0.0465504639, -0.0176357739, -0.0161931179, -0.0518580563, 0.0309407189, -0.0019002067, 0.0314246081, 0.0210341495, -0.0070821089, -0.0436444804, 0.0043178569, -0.0652659833, -0.0133820176, -0.022054391, 0.0015854571, 0.0072169448, -0.0155110518, -0.0336707346, 0.0040922929, -0.0464524738, -0.0239317846, 0.0256718397, 0.0410994329, -0.0178496502, -0.0138349291, -0.0139564229, -0.0307358, -0.0185490493, -0.0101132933, 0.0053222724, -0.0112735257, -0.0405372716, -0.0338271186, 0.0086862324, -0.032617189, 0.0127881933, -0.0145599591, -0.0235077627, 0.0126840156, 0.0369688682, 0.0236601196, 0.0163679738, 0.0207737852, 0.0289411321, 0.0584760644, 0.0074803727, -0.0844141841, -0.0123628555, 0.0133747021, -0.019464219, -0.0395372957, -0.0798400193, 0.0098975962, 0.0649441928, 0.0505980924, 0.0286252312, -0.0363346972, -0.0043167626, -0.0590297431, -0.0581600964, -0.0044538891, 0.0069765942, -0.0449415222, 0.0437223352, 0.0021061581, -0.0619601868, 0.05649212, -0.0176246818, 0.0188739933, -0.0112205446, -0.0168280546, 0.014040322, 0.0206253678, -0.0013430637, 0.0398287624, 0.0152434334, -0.0415677652, -0.0199063607, -0.0259025842, 0.0093715806, -0.0642712712, 0.0601357855, -0.0313876756, 0.0024502096, 0.0227807947, -0.0309597515, -0.054563988, -0.0878408775, -0.0263710804, -0.0444528088, -0.0016621413, -0.0327266231, -0.0517654754, -0.0212455895, 0.0490457378, 0.0263603088, -0.0136249317, -0.0177825801, 0.0096703647, -0.0541317649, 0.0506171398, 0.0223445669, 0.0730973184, -0.0415142961, 0.0374416895, 0.0527502745, 0.0375222415, 0.0206134245, 0.0378085747, -0.0355056748, 0.0456213728, -0.0093826298, -0.0356840789, 0.0540907383, 0.0483602323, -0.0458413623, 0.0138002485, 0.0988990068, 0.0296736974, -0.0463890769, 0.0049239509, -0.0818336159, 0.0589330569, 0.0356856249, 0.0050042709, -0.0918268263, -0.0106954426, 0.0029454227, 0.0558526069, -0.0482058898, -0.002034392, -0.0079619335, -0.1006293893, 0.0210229401, 0.0062340805, 0.0313549824, 0.0038842519, -0.0211383961, -0.0094142193, 0.0409513749, 0.0884429067, -0.0192228425, -0.0044933208, -0.0279076397, 0.0089658741, -0.1034289226, -0.0639335439, 0.050217893, -0.0151242474, 0.0496193282, 0.0453462228, 0.0107267825, -0.0394472443, -0.0589398965, 0.0525574051, 0.0008913784, -0.0146996537, -0.0064812824, 0.0279488005, 0.0096866218, 0.0065102307, -0.0557849184, 0.019161161, -0.0096275648, 0.0014034755, 0.0287156682, 0.0410354957, -0.0243347622, -0.0175969396, 0.0102629103, 0.0092630414, 0.0198432095, 0.0775540471, 0.0286171082, -0.0723688751, -0.0184606835, -0.0532717854, -0.017395772, 0.0092287483, -0.0590152517, 0.0801245943, 0.0293687172, -0.0120603414, -0.0375929363, -0.0419941097, 0.0286764633, -0.0081178136, 0.0070278239, 0.0114725539, 0.0196802486, -0.0411192626, 0.0390443504, 0.0432997234, -0.045271758, 0.0473413505, -0.0698991716, 0.0159076005, -0.0200362317, -0.0195004903, 0.0225270409, -0.0069625634, -0.0183559265, -0.0521369241, 0.0580313653, 0.0170672387, 0.0307614952, 0.0302187186, -0.0045291344, -0.0778350011, 0.0462429971, 0.0135093825, -0.0386787355, 0.0151976962, 0.0289066788, 0.0060414439, -0.0182829592, 0.041311685, -0.0104989652, -0.0063808821, 0.0041190698, 0.0219167043, -0.0422722399, -0.0143399807, 0.0349684097, 0.023262931, 0.0192302484, -0.0234235618, -0.0064004478, -0.0216945969, -0.0071150707, -0.0284131318, -0.0136479, 0.0055027199, -0.0349345058, 0.021942379, -0.0153066004, 0.0203483552, -0.0247406792, -0.0519290678, 0.0023587272, -0.0588824823, 0.0530017242, -0.0304820482, 0.0189979412, 0.0302125197, -0.012603729, -0.0378074087, 0.038568005, 0.0184036605, 0.0537071638, 0.0084238313, 0.0920078158, -0.0260119569, -0.0226280093, 0.0088914428, -0.0033244547, 0.010063855, -0.0787945613, 0.0046051722, 0.0112589253, 0.0139591666, -0.0003635469, 0.0668615997, 0.0010366405, 0.0107896943, -0.0218658559, -0.019876644, 0.0304998774, -0.0279954094, 0.0494882278, -0.0103066945, -0.0422791876, 0.0035636127, 0.0620524697, 0.0097270822, 0.0296686273, -0.0262393635, -0.012848598, -0.0263960492, -0.0592464879, -0.0216309279, 0.0195525996, -0.0649673492, -0.0343907848, -0.0946377441, -0.0449581183, -0.043465551, -0.0016099354, -0.0188392904, -0.0568739884, -0.0113835288, 0.0567377508, -0.0167401731, -0.0267667677, -0.0579354651, -0.0111214602, 0.0025989711, 0.0003552266, -0.0076011135, -6.39230375e-33, -0.0021700067, -0.0284572244, -0.0004444259, -0.0557575002, 0.0114531051, -0.0057829362, -0.0009402742, 0.0457382947, -0.0419811159, -0.0123099145, -0.0147545198, 0.0044591082, 0.0270541385, 0.0223893598, 0.0120386668, -0.0167316925, 0.032455489, 0.0196159277, -0.0309126843, 0.0178814437, 0.0290970001, -0.0121715348, 0.0460593142, -0.0708892271, 0.0844185874, -0.0205465667, -0.0152584333, -0.0076394705, -0.0284608547, 0.0573001467, -0.026063256, 0.087121442, 0.0129226008, -0.0454883091, -0.0097216796, 0.0286769811, -0.0097906021, 0.0006595354, 0.0559545681, 0.0112147061, 0.0250310935, -0.0276538134, 0.0343547985, -0.0068551828, -0.0372508056, 0.0063841846, 0.0213820338, -0.0150742754, 0.0054815281, 0.0204199236, -0.0390576944, 0.0091677671, -0.0086204866, 0.0268834326, -0.0587159321, 0.1052613929, -0.0217615161, -0.0076290094, -0.0614831708, 0.0278696753, 0.0263321437, 0.0081257671, 0.091798991, -0.023154987, 0.0043822606, 0.0850317404, 0.0148483254, 0.0182757229, -0.0197209381, -0.041285526, 0.0366865546, 0.0098603815, 0.0219210554, 0.0579020195, -0.0062595312, -0.0608080588, -0.0393903926, 0.0048942766, 0.0334255882, 0.0337494425, 0.0655050352, -0.025454659, -0.0799442753, -0.0635517016, -0.0392564461, 0.0324941613, -0.0271959081, -0.0417501256, 0.0626204461, -0.0086871972, -0.0022454571, -0.0079709105, 0.0309883859, -0.0280103963, -0.0604677536, -0.0130079733, -0.0075709349, -0.0039621992, -0.0046182228, 0.0517445691, -0.0629328266, 0.0015207713, -0.0491660871, -0.0006841926, 0.016544342, 0.0307216682, 0.00459912, -0.0148476446, -0.0866910964, 0.0188337192, -0.0173535645, 0.0390584432, 0.022926772, -0.0084357103, 0.0602228045, -0.0317794643, -0.011815994, 0.02727223, 0.0086214067, -0.0228054617, -0.0536085851, -0.031048907, -0.0228235219, 0.0621776134, -0.0365749374, -0.0113348262, -0.0527864583, 0.0144887082, 0.0665855482, -0.028977247, -0.0213725474, -0.0405675136, 2.499e-7, 0.0226734243, 0.0611780174, -0.0454702228, 0.0112017719, -0.014309736, 0.0413782187, 0.0107904868, 0.0126789846, 0.0549441278, -0.0010479362, -0.035699252, -0.0076830257, 0.0429352894, 0.0468265787, -0.0612021163, 0.0125879589, -0.0601613596, 0.0414195955, -0.0781517774, 0.0178880356, 0.0232046638, 0.0688962713, 0.0029074196, -0.0270618573, -0.0251136683, -0.0599669963, 0.0344609171, -0.0491343439, -0.0442268327, -0.0035728025, 0.0563437045, 0.011655652, 0.0339898504, 0.0380273648, -0.0157618839, 0.0014482499, -0.0210555159, 0.0489160679, 0.0150259687, -0.0226280931, -0.016481895, -0.0294773467, 0.0149986623, -0.0837155282, 0.0619679503, 0.0022002601, -0.0156574883, -0.0047863284, 0.0131055089, -0.0437763929, 0.039489869, -0.0422095321, -0.0324099623, 0.0027093284, 0.0316109322, 0.0051902407, -0.0062939841, -0.0133195613, -0.0176369734, 0.0058116252, -0.0121346489, -0.0742815211, -0.0160591397, 0.0457255654, -0.0277264211, -0.0339827649, 0.035304226, 2.319889988e-34, 0.0305302162, 0.0475695617, 0.0473625772, -0.0115177324, 0.0440440476, 0.0481585674, 0.0668870881, 0.0130462609, -0.0026551117, -0.0551494658, -0.0031500463 ]
http://arxiv.org/abs/2406.17716v2
http://arxiv.org/pdf/2406.17716v2
ViANLI: Adversarial Natural Language Inference for Vietnamese
2024-06-25T16:58:19
2024-07-01T15:19:51
Tin Van Huynh, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen
cs.CL
cs.CL
The development of Natural Language Processing (NLI) datasets and models has been inspired by innovations in annotation design. With the rapid development of machine learning models today, the performance of existing machine learning models has quickly reached state-of-the-art results on a variety of tasks related to natural language processing, including natural language inference tasks. By using a pre-trained model during the annotation process, it is possible to challenge current NLI models by having humans produce premise-hypothesis combinations that the machine model cannot correctly predict. To remain attractive and challenging in the research of natural language inference for Vietnamese, in this paper, we introduce the adversarial NLI dataset to the NLP research community with the name ViANLI. This data set contains more than 10K premise-hypothesis pairs and is built by a continuously adjusting process to obtain the most out of the patterns generated by the annotators. ViANLI dataset has brought many difficulties to many current SOTA models when the accuracy of the most powerful model on the test set only reached 48.4%. Additionally, the experimental results show that the models trained on our dataset have significantly improved the results on other Vietnamese NLI datasets.
ViANLI: Adversarial Natural Language Inference for Vietnamese: The development of Natural Language Processing (NLI) datasets and models has been inspired by innovations in annotation design. With the rapid development of machine learning models today, the performance of existing machine learning models has quickly reached state-of-the-art results on a variety of tasks related to natural language processing, including natural language inference tasks. By using a pre-trained model during the annotation process, it is possible to challenge current NLI models by having humans produce premise-hypothesis combinations that the machine model cannot correctly predict. To remain attractive and challenging in the research of natural language inference for Vietnamese, in this paper, we introduce the adversarial NLI dataset to the NLP research community with the name ViANLI. This data set contains more than 10K premise-hypothesis pairs and is built by a continuously adjusting process to obtain the most out of the patterns generated by the annotators. ViANLI dataset has brought many difficulties to many current SOTA models when the accuracy of the most powerful model on the test set only reached 48.4%. Additionally, the experimental results show that the models trained on our dataset have significantly improved the results on other Vietnamese NLI datasets.
[ 0.0532222465, 0.0701498464, -0.0249648578, 0.043488428, -0.0438763648, 0.0248614959, 0.0128460312, -0.0255510546, 0.0743760243, -0.040447522, -0.0222140979, -0.013193219, 0.0185943283, 0.0178447403, 0.0355065949, -0.0456946939, 0.032540001, -0.0056112744, 0.0033132709, -0.0118832514, 0.0014593628, 0.0047119386, 0.0049158488, 0.0253731441, 0.0059419526, -0.013983272, -0.0060317996, 0.028428359, 0.0020874976, -0.0544022135, -0.0028405453, 0.0551858321, -0.0097632315, 0.050122153, 0.0000020416, -0.0368092693, 0.0107030561, -0.018056456, -0.0306928977, 0.018484408, 0.0436071493, -0.0590620823, -0.0016393366, 0.0228052866, -0.0521699376, -0.0513469875, 0.0514575168, 0.0573855191, 0.0588481948, 0.0575133525, -0.0236268882, -0.0492214151, 0.0101142395, -0.0624510311, -0.042976208, -0.0369724929, 0.0353823677, -0.010629138, -0.0268402137, -0.033687219, -0.0132722417, 0.0327576362, 0.048484087, 0.0253242478, -0.0456602052, 0.0150650721, -0.0221807472, -0.0238147601, -0.0079592532, 0.0693180785, -0.0151586328, -0.0287250001, -0.0332102999, 0.0036379939, -0.0040757689, 0.0167799499, -0.0044014947, 0.0172865726, -0.0182645041, 0.0290462729, 0.0908903554, 0.0199040174, 0.0260289647, 0.03939588, 0.0106559442, -0.0199245065, 0.0006226411, -0.0357300006, 0.0276009478, 0.0146661531, -0.087952666, -0.0340187475, 0.0079854606, -0.0119322799, 0.0090380218, 0.0109603517, -0.0288131312, -0.0300876275, 0.0452547558, -0.0846699327, 0.0239568017, 0.0149098067, -0.0644500256, 0.0511031039, -0.0122344326, 0.0040317406, -0.0450794511, 0.0129325772, -0.0379269049, -0.0300539043, -0.0027149289, 0.0038639696, -0.0375389792, 0.0359514691, -0.0356832407, -0.0033454422, -0.0407221094, -0.0027714483, -0.0180610716, -0.011313444, -0.0049487096, 0.024995653, -0.0683238879, -0.0108406171, -0.0027132137, 0.002191935, -0.0336160697, 0.0662658364, -0.0153402453, -0.0835216641, -0.0067121312, 0.0509020239, -0.0107174618, -0.0430516638, 0.0183066968, 0.0646056458, 0.0116878254, 0.007636894, -0.0566420145, 0.0154277738, 0.0105418731, -0.0582282394, 0.0036833251, -0.0445577502, 0.0172620453, -0.005444909, 0.0204166509, -0.0580402166, -0.0000664345, -0.0008069406, -0.0178434197, 0.1101502478, -0.0062426073, -0.0194659662, 0.0262390748, 0.0075513846, -0.0165078007, 0.0804732218, -0.0090463161, 0.0463888198, 0.031587746, 0.0054080207, -0.0210336372, 0.0248829033, 0.0152025316, -0.0024882876, -0.0094325943, 0.0361020118, -0.0179190449, 0.0227862205, -0.0020573901, 0.0626069084, -0.0231829491, -0.0341207199, 0.0340849608, 0.0270838924, -0.0058323275, 0.0340522528, 0.0447002277, 0.0083143516, 0.0323030315, 0.0176931396, -0.0524635278, 0.0107915606, -0.0487769246, -0.0202469248, 0.0219621081, 0.0312383249, -0.0328595899, -0.0747493804, -0.0156514943, -0.0386597626, 0.006633359, -0.0032209987, 0.0568728, 0.0675907359, -0.0146119315, 0.014502624, 0.0495520048, -0.0224728193, -0.0359597765, -0.0435851663, 0.0045798095, 0.0031525323, 0.0285403114, -0.0259316992, 0.0270584952, -0.0015568679, 0.0088190157, 0.0790495053, 0.0597154461, -0.0235212725, 0.0173827745, -0.0209496766, 0.0237570275, 0.0042256638, 0.0489620194, 0.0703914165, 0.0219679885, 0.0166138317, 0.0228122026, -0.0139605924, -0.0376323089, 0.0531693548, -0.0144894579, -0.0329100825, 0.0249158684, 0.0347959325, 0.0471992902, 0.0150784114, -0.014489268, -0.0020269686, 0.0375131927, -0.0536923781, 0.0158065595, 0.0033592265, -0.0003241176, -0.0887214541, -0.008705535, -0.0612463392, -0.0279828086, 0.0806176513, -0.0679379255, 0.0000570348, 0.0203703493, -0.0080110254, -0.0136940535, -0.0794048086, -0.0206618104, 0.0174935628, 0.0627068207, -0.0063818116, 0.0837716088, -0.0230182502, -0.0082186619, 0.0414755307, 0.0186507311, 0.0194912851, -0.0276250634, -0.0381206162, 0.0436127111, -0.0023925889, -0.008530369, 0.0010981341, -0.0158873647, 0.0090573058, -0.0172720999, -0.0914943293, 0.0081261015, -0.04239057, 0.00008445, 0.0172419157, -0.0053307922, -0.0296078194, -0.0661575049, 0.007304023, -0.0202094968, 0.0042829025, 0.0287249926, 0.0070420527, -0.015445794, 0.0171256159, -0.0322194956, -0.0257819351, 0.0012565835, 0.0431444757, 0.0069765998, -0.0571139492, -0.0331611149, -0.0105213616, -0.0148089649, 0.0312269703, -0.0206280071, -0.0611149967, 0.0178037584, 0.0048320554, -0.0385157466, 0.0939797238, -0.0062814085, 0.0201198272, 0.0252691209, 0.0137885325, -0.0042278795, 0.0118225468, -0.0143459598, 0.0095256045, -0.0329622924, -0.0254322905, -0.0422460102, 0.0651000589, 0.0298702996, -0.0089603001, -0.0260406937, 0.0215928536, -0.0275179036, -0.0210394561, 0.0155413793, 0.0114863226, -0.0169243719, 0.0102737425, -0.0229542963, -0.0316586755, 0.1175079867, 0.0096536539, 0.0054116054, -0.0541393012, -0.0002353899, -0.0372162908, 0.0961872041, 0.0305671692, -0.018286122, 0.0225226264, -0.0122160185, -0.0257200599, 0.0135847554, 0.0378352553, -0.0240637455, 0.0234461892, -0.0384950563, -0.0124699622, -0.0133611625, -0.0290306676, -0.0787529871, -0.0285005271, -0.0448326543, 0.0002641311, -0.0088740727, 0.030474497, 0.0020247446, -0.0075015258, -0.0573977679, 0.0174781345, 0.0061585866, 0.0272042118, -0.0353983082, 0.0383114927, 0.0422788411, -0.0259983838, 0.0085939663, 0.0151920384, 0.0619816408, 0.0561135225, -0.0335540399, 0.0192007553, -0.0137502914, 0.0179499779, 0.0170090273, -0.0133764539, -0.0052575907, -0.0020476109, 0.004561469, -0.0437360741, 0.0322058462, 0.0720449537, 0.0168458, 0.0060096704, 0.0292130932, 0.0334439203, 0.0147663457, 0.003634054, 0.004135638, -0.0792259127, -0.0406254195, -0.0056554209, -0.0095804054, -0.0188485645, -0.0199481826, -0.0205658879, -0.1146165058, 0.0118183261, -0.0222347323, -0.035855718, -0.0238273982, -0.0555989705, -0.0017762921, -0.0622338988, 0.0257282052, -0.0160556175, 0.0242196806, -0.0081600314, 0.0125551578, -0.0285377447, -0.0441882387, 0.0435548536, -0.0376506634, 0.0199461654, 0.0458145216, 0.0465440266, -0.0093860067, 0.0004093865, 0.0320482515, -0.0209358688, 0.0058084093, -0.0082959039, 0.0707199648, -0.0034666173, -0.001555839, -0.0231732, 0.0531149358, -0.0344461426, -0.0098450035, 0.0166155323, 0.0663337335, -0.0064996658, -0.0152602959, -0.0019866547, 0.0242973175, 0.0046454123, 0.0824966058, 0.0277880169, -0.0411395654, -0.0359233692, -0.0377694517, -0.0236973371, -0.0344448686, -0.0304348152, 0.0420297831, -0.0132641261, -0.076050587, -0.0754292011, -0.0829653144, -0.0005077263, 0.0390184112, 0.0113496799, 0.0380634852, -0.0099485442, 0.0035249263, -0.0158083364, 0.029201841, -0.0092461761, 0.0036829973, 0.0387157947, -0.0239851456, 0.0186399445, -0.0392639302, -0.0031901014, 0.0293716863, -0.0768173337, -0.0741761923, 0.0284060929, 0.0427105688, 0.0221876092, 0.08534237, -0.0395302847, -0.0307533462, 0.0022933898, 0.0524672046, 0.0053016124, -0.0389283299, 0.0467063934, -0.0324949473, 0.006470798, 0.065897353, 0.0112222964, 0.0132089294, 0.0490097739, -0.0095804669, -0.0369304419, -0.00227872, 0.0268497374, 0.0180386733, 0.087313734, -0.0196490288, -0.0434936136, -0.0487043411, 0.0222246256, -0.0281725284, -0.034914881, -0.0132065937, 0.0119408295, 0.0175464209, 0.0173356142, -0.0383851081, -0.0227535218, -0.0224709641, 0.0092407148, -0.0049853562, 0.0232750177, 0.0487398021, 0.0880601183, -0.0042671645, 0.0019517874, -0.0285146311, -0.0052411682, 0.0038465299, 0.0563819595, 0.0082889432, 0.0349231623, 0.0290323868, -0.0069449688, 0.012601058, 0.0036055471, -0.0430395529, -0.013513891, -0.0447015576, 0.0303534362, 0.0674112737, -0.0300514251, 0.0220830645, -0.0547787324, 0.0096350955, 0.0022812267, 0.0339325219, -0.0030013244, -0.0372125208, 0.0236320719, -0.0141595993, -0.0162226222, 0.0047378978, 0.0526013374, 0.0187667999, 0.0686153322, 0.0171637032, -0.007238525, -0.0461308733, -0.0373573862, 0.0243537035, 0.0127528552, -0.030133279, 0.0305708461, -0.0336306803, -0.0900589004, -0.0043848408, 0.0254910514, 0.0242460668, 0.0427691042, -0.0018056373, 0.0658719614, 0.026390966, -0.0227202307, -0.0505727977, 0.0144060748, -0.0488795266, 0.0008401834, -0.0223522857, -6.562610106e-33, -0.0285681114, -0.0099051809, 0.0009169983, -0.0296406746, 0.0089160167, -0.0047330586, -0.0382688642, -0.0053608995, -0.0531265028, 0.0094457129, -0.0172120035, 0.0123563157, 0.0290586948, 0.0289105326, -0.0060178502, -0.0463648513, 0.0061353934, 0.0344627798, 0.0092768921, -0.0348619111, 0.0027439038, 0.022465501, 0.0843596682, -0.0713451058, 0.0436625779, 0.0050389394, -0.0102042574, 0.0057619419, -0.0497305207, 0.046354726, -0.0244078059, 0.0420095325, 0.0256617051, 0.0025319003, 0.0111299232, -0.0221354943, -0.0810828879, -0.0653789937, 0.0497631058, 0.0632084757, -0.0528411753, -0.0276812557, 0.0109545551, -0.0176830813, -0.0007084918, 0.0140584158, 0.0236288607, -0.0534117781, 0.0293701198, 0.0269759111, -0.0400071777, 0.038163662, 0.0227914788, -0.0066244272, -0.0079216938, 0.059786588, 0.0078717377, -0.0399739444, -0.0977504775, 0.0192494784, 0.0016969651, -0.0129261855, 0.0326165259, -0.0068673864, 0.0448873155, 0.0268013999, 0.0465275273, -0.0161331091, -0.0493009463, 0.0070817866, 0.0015173682, 0.0455191061, 0.0669098496, 0.0833396986, 0.093143627, 0.005592796, -0.094871901, -0.0119802374, 0.0242768861, 0.0271151848, 0.0056691389, -0.0038992674, -0.0857218429, -0.0375026017, -0.017540941, -0.1029223427, -0.0162331946, -0.0605633445, -0.0019333025, 0.0219082888, -0.0563741289, 0.0286954492, 0.0124753909, -0.0598941222, -0.0018593837, -0.0030855099, -0.010232524, -0.0351025723, -0.0065852245, 0.0436370894, -0.0585320331, -0.0074220817, 0.0141201951, -0.0182186086, 0.0185102653, 0.0079835868, 0.0723984763, -0.0049639787, -0.0396874622, 0.0106081897, -0.0298399553, -0.0088898027, 0.0021502441, 0.0140918018, 0.054673966, 0.0382592753, -0.0000492162, 0.013773161, 0.027502019, 0.0199705232, -0.0679306313, 0.0164787024, -0.0011709203, 0.0063353465, -0.0332958698, -0.0064290673, -0.0224201735, -0.0135821793, 0.008230147, -0.0674963817, -0.0140824402, 0.0168717541, 2.668e-7, 0.0463004448, 0.0027634064, 0.0548869111, 0.0100778947, -0.0102758547, 0.0305737015, 0.0003647005, 0.0024476689, 0.0072250427, -0.0187413264, 0.0276928023, 0.042413123, 0.0651179701, 0.0423473045, -0.0866612718, 0.0101986155, -0.0584563091, 0.0153593589, -0.0306063425, -0.037420895, 0.0506559387, 0.1298654079, 0.0071230987, -0.0132118491, -0.0290203933, -0.0215010755, 0.0478152484, 0.0135597838, -0.0208163615, 0.0072326432, 0.0431726947, -0.0136270775, 0.0063664084, 0.0265203677, -0.0195492133, -0.037617825, 0.0339725055, 0.0163792353, -0.0090472363, -0.0277131572, -0.0308867898, -0.0134507669, -0.0118109668, -0.0267723426, 0.0920797139, -0.0251341127, -0.0552655272, -0.0151283126, -0.0043874304, 0.0414394848, 0.0586065501, -0.0062387278, -0.0274964776, 0.0125306258, 0.0421874486, -0.0315079354, -0.0286680888, -0.0149299363, -0.0387464166, -0.0344790965, -0.0319245569, -0.0000983689, -0.0055137272, 0.0231464822, 0.0411032401, -0.0162911676, 0.0053443168, 2.065351151e-34, 0.0235221256, 0.0062209554, 0.0052044899, -0.0205016974, 0.0559040084, -0.0294372272, -0.002876546, -0.0067831995, 0.0056237667, -0.0440899655, -0.0149436546 ]
http://arxiv.org/abs/2406.17706v1
http://arxiv.org/pdf/2406.17706v1
FedBiOT: LLM Local Fine-tuning in Federated Learning without Full Model
2024-06-25T16:45:47
2024-06-25T16:45:47
Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, Jing Gao
cs.LG
cs.LG, cs.CL, cs.DC
Large language models (LLMs) show amazing performance on many domain-specific tasks after fine-tuning with some appropriate data. However, many domain-specific data are privately distributed across multiple owners. Thus, this dilemma raises the interest in how to perform LLM fine-tuning in federated learning (FL). However, confronted with limited computation and communication capacities, FL clients struggle to fine-tune an LLM effectively. To this end, we introduce FedBiOT, a resource-efficient LLM fine-tuning approach to FL. Specifically, our method involves the server generating a compressed LLM and aligning its performance with the full model. Subsequently, the clients fine-tune a lightweight yet important part of the compressed model, referred to as an adapter. Notice that as the server has no access to the private data owned by the clients, the data used for alignment by the server has a different distribution from the one used for fine-tuning by clients. We formulate the problem into a bi-level optimization problem to minimize the negative effect of data discrepancy and derive the updating rules for the server and clients. We conduct extensive experiments on LLaMA-2, empirically showing that the adapter has exceptional performance when reintegrated into the global LLM. The results also indicate that the proposed FedBiOT significantly reduces resource consumption compared to existing benchmarks, all while achieving comparable performance levels.
FedBiOT: LLM Local Fine-tuning in Federated Learning without Full Model: Large language models (LLMs) show amazing performance on many domain-specific tasks after fine-tuning with some appropriate data. However, many domain-specific data are privately distributed across multiple owners. Thus, this dilemma raises the interest in how to perform LLM fine-tuning in federated learning (FL). However, confronted with limited computation and communication capacities, FL clients struggle to fine-tune an LLM effectively. To this end, we introduce FedBiOT, a resource-efficient LLM fine-tuning approach to FL. Specifically, our method involves the server generating a compressed LLM and aligning its performance with the full model. Subsequently, the clients fine-tune a lightweight yet important part of the compressed model, referred to as an adapter. Notice that as the server has no access to the private data owned by the clients, the data used for alignment by the server has a different distribution from the one used for fine-tuning by clients. We formulate the problem into a bi-level optimization problem to minimize the negative effect of data discrepancy and derive the updating rules for the server and clients. We conduct extensive experiments on LLaMA-2, empirically showing that the adapter has exceptional performance when reintegrated into the global LLM. The results also indicate that the proposed FedBiOT significantly reduces resource consumption compared to existing benchmarks, all while achieving comparable performance levels.
[ -0.0048592244, 0.0779442266, -0.0237358473, 0.0560917258, -0.0504147969, -0.0037646382, 0.0498327315, -0.0111038545, 0.0086029246, -0.0084636789, -0.0761822686, -0.0275226831, 0.0204450749, 0.0323803276, -0.0073936344, 0.0083681475, 0.0415510461, -0.0152878631, -0.03088101, -0.0401250012, -0.0359721668, -0.00170876, 0.0250069611, 0.0040306798, 0.0306915175, -0.0414599776, -0.0151846381, 0.0177129135, 0.0048980243, 0.026680192, 0.0398769639, 0.0411745049, 0.0135127967, 0.0768393055, 0.0000021781, -0.0275452342, -0.0444030911, 0.0214960705, -0.0315661132, 0.0884731114, 0.0171573851, -0.0026801541, -0.023031259, 0.0357158706, -0.0339746326, -0.0118508162, 0.0291574895, -0.013328963, 0.0836292803, 0.0783935562, 0.002079969, -0.0354283117, 0.0054081697, -0.0394620784, -0.1231549233, -0.0456108563, 0.0142975412, -0.024460068, -0.041829206, -0.0257883128, -0.0311601143, 0.0070526553, 0.0142368488, 0.0308112595, -0.0118654827, -0.0013838911, -0.0614461266, -0.0292152334, -0.0222826265, 0.04694492, 0.0056478581, -0.0084203184, 0.0130263222, -0.0187958758, -0.0224419031, 0.0046699601, -0.0011252976, 0.0161448009, 0.026967857, 0.0424503349, 0.0122225303, -0.0290237814, 0.0319492109, 0.0071174749, -0.0110217454, 0.0940016508, 0.001840096, 0.0144025199, -0.0529506952, -0.0101598771, -0.0166083779, -0.0614272989, 0.0328903683, 0.0411407389, -0.0254514534, -0.0144900652, -0.0879801959, -0.0629331097, 0.0198128242, 0.0206349306, 0.077003397, 0.0092163254, 0.0016887725, 0.049253691, -0.0204693377, 0.0113523155, -0.0314116552, 0.031520538, -0.0171666183, 0.0071278308, -0.0166101214, 0.0290494971, -0.0032895894, 0.0390895195, -0.0660323799, 0.0027491809, -0.0163723566, 0.0320257731, 0.0079858974, 0.0360128395, 0.0086442381, -0.0273082871, -0.0376617312, -0.0077328105, 0.0123097207, 0.0249084737, -0.0340063646, 0.0110372882, 0.0330291651, -0.0531443916, -0.0121554872, -0.0089877313, -0.0347755887, -0.0120460466, 0.017598452, 0.0558258407, 0.0130969556, -0.0018284456, 0.0069967853, -0.0522624515, 0.0560681857, -0.083925657, 0.0021034875, -0.0376315191, 0.0185860731, -0.0069952291, 0.014403698, -0.0440674573, -0.0205747839, -0.031347353, -0.0283818021, 0.0520656183, -0.0467091352, 0.0020855498, -0.006950004, 0.007328792, 0.0156223318, 0.1055157855, 0.0208958611, 0.0044786823, 0.0431455187, 0.0073680379, -0.0485515557, 0.038286373, 0.0635784566, 0.0591339692, -0.0257362146, 0.0102038458, 0.0004083851, 0.0444412492, 0.000901557, 0.0209667943, -0.0055425428, -0.0196575224, 0.0447392352, 0.0562537909, 0.0580508485, 0.0135324914, -0.052569177, 0.0194785483, 0.0042490968, 0.0319096036, -0.0671754926, 0.0241221748, -0.0199645944, -0.0287271086, 0.023230467, 0.1122396961, -0.0285616256, -0.0313866846, -0.0145038962, -0.0438196845, -0.048316054, 0.0156210037, 0.0208959542, -0.001072711, -0.0122614084, 0.0757944509, 0.044709444, -0.0228977595, -0.0194397457, -0.0339147411, -0.0090482347, 0.0621905699, 0.0779167563, -0.0022941725, -0.0512866415, -0.0066783461, 0.030149851, 0.0212613344, 0.0175523423, 0.0213818904, 0.0186304022, -0.0539132021, -0.0708050728, 0.0230709706, 0.0120267924, -0.004592163, -0.0049790088, -0.0065308618, -0.0317543708, -0.0229081884, -0.0008113403, 0.0306892227, 0.0111068804, -0.0055415961, 0.0104770605, 0.0237823911, 0.0485531501, -0.0292618573, -0.0143924607, -0.0387563594, 0.0127569214, -0.0096647572, -0.0123731932, -0.0386433043, -0.0017101022, 0.000277902, 0.021600645, 0.0136301592, 0.0361969657, 0.0430855453, -0.0885092616, -0.0310972463, -0.0037948077, 0.0083532026, 0.0246819649, -0.0140170464, 0.024366226, -0.0233891625, 0.0448675752, -0.0248881318, 0.0036470401, -0.0412857607, -0.0237967614, -0.013348449, -0.0189353842, 0.024378825, -0.0172537453, 0.0345636047, 0.009196315, 0.0141121438, 0.0167898219, -0.0264852885, 0.0235959925, -0.0066431854, -0.0009206834, 0.017083738, 0.0384967998, -0.0348162577, 0.0566481613, 0.0293682776, 0.0200860426, -0.021164747, -0.0203384925, -0.0129317017, 0.0693890527, -0.0432379395, 0.026446905, 0.0201309174, 0.0092244074, -0.0490336046, -0.0167273693, -0.0307609905, 0.0319129266, 0.0136635881, -0.0329365581, -0.0078113703, -0.0808898807, 0.0256010797, -0.0613693707, 0.04438667, 0.0028519575, -0.0382315181, -0.0192874279, 0.0319681689, -0.0112804715, 0.0824697167, -0.0276308097, -0.0103338929, 0.0650032908, 0.0266649574, -0.0786962211, 0.0097454321, -0.0152013618, -0.0422890857, -0.0609007515, 0.0421498418, 0.0149004953, 0.0659607947, -0.0595748648, -0.0298260599, -0.0223426502, 0.0407281034, -0.0156034073, -0.0216618702, -0.0259589832, 0.0358872749, 0.0047325594, 0.0566601157, 0.0142477052, -0.0942550004, 0.0891349167, -0.0302351546, 0.0142117282, -0.0083092386, -0.0475892052, -0.0642460138, 0.0148713998, -0.0132498033, 0.026521476, -0.0220167059, -0.0445896126, -0.0478586257, 0.0274894126, 0.0341458172, -0.0379616395, 0.006371785, -0.0037751065, -0.0107953716, -0.0505428836, -0.0617613047, 0.0147716878, -0.0257610995, -0.0456287898, 0.016195219, -0.016866684, -0.0327722356, -0.0380271003, 0.0087737255, -0.0095960936, 0.0055052149, -0.0121820299, 0.0199069753, -0.0367223471, -0.0247369856, 0.011483537, 0.0021573086, -0.0066580204, -0.0289660115, -0.0295327138, 0.0387945212, 0.0394981429, -0.0225336831, 0.0336841568, -0.0328572243, 0.0605951026, -0.0177509878, -0.0229444858, -0.0111629767, 0.0599131472, -0.0471421108, -0.0091594998, 0.0727721974, 0.0090728439, -0.0452684946, -0.0000766489, -0.0353846513, 0.0668171123, 0.0148674194, 0.039814204, -0.0799084082, -0.0312825441, 0.0467523783, 0.0550565571, 0.032198038, -0.0207434371, -0.0217898749, -0.1102538481, 0.0181704555, 0.0236415882, -0.0030314759, -0.0363989286, -0.0365218706, 0.0219419096, -0.0031644544, 0.0230117161, -0.0165316146, 0.0290210731, -0.0072053149, 0.0156077305, 0.0076189782, -0.0493630804, -0.0080258362, -0.0153990397, 0.0794083402, 0.0261316076, 0.001214503, 0.0348806344, -0.0742888898, 0.0166076366, 0.0417421795, -0.0169339981, -0.0388242193, -0.0125070568, 0.0042925142, -0.0218081102, 0.0196294412, 0.0127569139, 0.0172999594, -0.0016436682, 0.0068159159, 0.0948340818, -0.009884988, 0.0257968884, 0.0034181727, -0.0416125581, 0.0153706037, 0.0410055742, 0.0276146345, -0.0182290971, -0.0220053494, 0.0117024528, -0.0107076894, -0.0042504529, 0.0038277507, 0.0249028876, -0.0255648363, 0.0208410509, -0.0115132332, 0.0324248374, 0.0050407117, 0.052899722, 0.0146063231, 0.010039011, 0.0276722796, -0.0039447867, -0.0449175611, 0.0010107268, -0.0874509513, 0.0329597034, -0.0721437559, -0.0289233681, 0.0392431468, 0.0162815321, -0.0180755891, 0.0164516177, -0.0643320456, -0.0785219893, 0.0214662012, 0.0251253955, 0.0219608843, 0.0493227392, -0.0185845196, -0.0693139806, 0.0637907311, 0.0589190982, -0.0780178905, 0.0477843545, 0.0372642353, -0.0751656517, 0.0167381205, 0.0599391535, -0.0646961182, -0.0015001385, 0.0570605397, -0.0006134315, 0.0080878921, -0.0982145965, -0.0481443331, 0.0663151368, 0.0581548773, -0.0112431925, -0.027447708, -0.0067507029, 0.0649567395, 0.010362125, 0.0104434872, -0.031957183, -0.0220330302, -0.0061641973, -0.0272179246, 0.0287409853, 0.0083018159, 0.0561129004, 0.0091601368, -0.010808384, 0.0121895811, -0.0102972016, -0.0434983149, 0.0641680062, 0.0173855219, 0.0269009378, 0.0234778617, 0.0564440899, 0.0709505454, -0.0361748263, 0.0090670316, -0.0135993976, 0.0545612164, -0.0602390319, 0.0403406844, -0.0141523341, -0.1076711193, -0.002864049, 0.0346512198, 0.0353397503, 0.0429203324, 0.0102020642, -0.0178970248, -0.0107923718, -0.014152525, 0.0025052754, -0.0439040028, -0.0308106504, 0.0427479371, -0.0168477371, -0.0049831914, 0.0513196588, 0.0826515332, -0.0017098489, 0.0411212966, -0.0548186339, 0.031598065, -0.0474952757, -0.0204734355, 0.0020065235, 0.032523144, -0.033204779, -0.0225548223, -0.0267668236, -0.0611489825, 0.0047201738, 0.004700948, -0.0089828102, -0.0261596981, 0.0151256882, 0.0285031665, 0.0067163208, -0.0018641717, -0.0539518185, 0.0010385173, 0.0130746271, -0.0025105465, 0.0446324907, -6.80466276e-33, -0.0471461713, 0.0095692445, 0.0080269258, 0.031926211, -0.0238636695, 0.0302441157, -0.0028946083, -0.0235801972, 0.0178913046, -0.0246592388, -0.0531205758, 0.0157247223, 0.0334450342, 0.0160576776, -0.0147411656, -0.028428616, 0.0347482301, -0.0521970876, -0.0527257286, 0.0173565596, 0.0329854377, 0.0475543961, 0.0223673545, -0.0337275192, -0.0335301012, -0.0178137198, -0.0060019293, 0.045063816, -0.0376977399, 0.0235969815, -0.0167743694, 0.0188087504, -0.0020199073, -0.0265491959, -0.0250990279, -0.0334345847, -0.0460328162, 0.0043059643, -0.0071305414, 0.0189931914, -0.0391189344, -0.0395536497, 0.053948421, 0.0165435746, -0.0352658965, -0.0247118175, 0.0015344921, -0.0078818509, -0.0365006216, -0.0275616515, -0.0277978089, -0.0100370608, 0.0305565149, -0.0052451449, 0.0465166681, -0.0335891247, 0.0154024046, -0.0062638004, -0.0397628918, 0.0112922769, 0.0202231389, 0.0371573046, 0.0808152631, 0.0221233442, -0.0040774262, 0.0399142504, 0.0472314283, 0.0107017942, -0.0128891878, 0.0328080095, 0.0275288951, 0.0214244835, -0.0144751007, 0.0462500788, -0.0133843794, -0.0380245708, -0.0457194969, -0.0072702565, 0.0410922915, -0.0165007822, -0.0115408599, 0.0350688845, 0.0048564775, -0.0087651657, -0.0352448896, 0.0181172416, -0.0197049454, 0.00666897, -0.0123264361, 0.0113642346, -0.0279489085, -0.0516183749, 0.0350867324, -0.0392209366, -0.0011246947, -0.0034281239, 0.0586569309, 0.0049328841, 0.016947262, 0.0232064128, -0.0731453598, -0.0469097346, -0.0289867353, -0.0360321999, 0.0272656232, 0.00717569, 0.0670188069, 0.0304992497, -0.0654753521, 0.0053106439, -0.0352407545, 0.0118604209, 0.0251016952, 0.023380788, 0.0530019961, -0.0307091512, 0.0171631612, 0.0218624622, 0.0083921589, 0.0050579025, -0.0116726793, -0.0090181101, 0.0061502578, 0.0201351419, -0.0060270559, 0.0034674162, -0.0532445163, 0.0355967321, 0.0389978923, -0.0261360966, -0.005236702, -0.0379410163, 2.951e-7, 0.0230782069, 0.0459817164, -0.0163958278, -0.0325500295, 0.032961335, 0.0350381844, -0.0435267352, 0.0375269465, 0.003140897, 0.0275102668, 0.0053067878, 0.0284658037, 0.055885423, 0.0031728102, -0.0434296057, 0.0308244713, -0.0300398935, 0.0530871302, -0.0502500609, -0.0587280877, -0.0531337745, 0.0817480907, 0.039936237, -0.0442282148, 0.0181043763, -0.0143603226, 0.0293186512, -0.0330685377, -0.052238293, 0.0131101878, 0.0201983694, 0.102790229, -0.0298428182, -0.0360503085, -0.0110679371, 0.0236746259, -0.061740566, 0.08435563, 0.0336946845, 0.024307223, -0.0029612437, -0.0398774333, -0.0484066382, 0.0243485048, 0.0469119735, -0.0270132385, -0.0257130265, -0.018204106, 0.039255619, -0.067603521, 0.0409304388, -0.0040403116, 0.0201132838, 0.0275081899, -0.0004583366, -0.0254764128, -0.0420671813, -0.0688579157, -0.0054213987, -0.0199031755, -0.0243356302, -0.0499996841, -0.0333599262, 0.0229147915, 0.0205625109, -0.0317941383, 0.0116256, 2.836496797e-34, 0.0240344778, -0.0253707413, 0.02832748, 0.0407824367, -0.025119083, -0.0362560786, 0.001029058, 0.0181054156, -0.0129285129, -0.0379336588, 0.0079593761 ]
http://arxiv.org/abs/2406.17692v1
http://arxiv.org/pdf/2406.17692v1
From Distributional to Overton Pluralism: Investigating Large Language Model Alignment
2024-06-25T16:32:33
2024-06-25T16:32:33
Thom Lake, Eunsol Choi, Greg Durrett
cs.CL
cs.CL, cs.LG
The alignment process changes several properties of a large language model's (LLM's) output distribution. We analyze two aspects of post-alignment distributional shift of LLM responses. First, we re-examine previously reported reductions in response diversity post-alignment. Our analysis suggests that an apparent drop in the diversity of responses is largely explained by quality control and information aggregation. Alignment suppresses irrelevant and unhelpful content while shifting the output distribution toward longer responses that cover information spanning several responses from the base LLM, essentially presenting diverse information in a single response. Finding little evidence that alignment suppresses useful information, it is natural to ask the opposite question: do aligned models surface information that cannot be recovered from base models? Our second investigation shows this is not the case and the behavior of aligned models is recoverable from base models without fine-tuning. A combination of in-context examples and lower-resolution semantic hints about response content can elicit responses from base LLMs that are as similar to alignment-tuned LLM responses as alignment-tuned LLM responses are to each other. Taken together, these results indicate that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior, providing further evidence for the Superficial Alignment Hypothesis. They also show that in-context alignment can go surprisingly far as a strategy for imitating aligned LLMs without fine-tuning. Our code and data is available at https://github.com/thomlake/investigating-alignment.
From Distributional to Overton Pluralism: Investigating Large Language Model Alignment: The alignment process changes several properties of a large language model's (LLM's) output distribution. We analyze two aspects of post-alignment distributional shift of LLM responses. First, we re-examine previously reported reductions in response diversity post-alignment. Our analysis suggests that an apparent drop in the diversity of responses is largely explained by quality control and information aggregation. Alignment suppresses irrelevant and unhelpful content while shifting the output distribution toward longer responses that cover information spanning several responses from the base LLM, essentially presenting diverse information in a single response. Finding little evidence that alignment suppresses useful information, it is natural to ask the opposite question: do aligned models surface information that cannot be recovered from base models? Our second investigation shows this is not the case and the behavior of aligned models is recoverable from base models without fine-tuning. A combination of in-context examples and lower-resolution semantic hints about response content can elicit responses from base LLMs that are as similar to alignment-tuned LLM responses as alignment-tuned LLM responses are to each other. Taken together, these results indicate that current alignment techniques capture but do not extend the useful subset of assistant-like base LLM behavior, providing further evidence for the Superficial Alignment Hypothesis. They also show that in-context alignment can go surprisingly far as a strategy for imitating aligned LLMs without fine-tuning. Our code and data is available at https://github.com/thomlake/investigating-alignment.
[ 0.0208503269, 0.0410461277, -0.0365810096, 0.0338490047, -0.055195339, -0.0038789667, -0.0095026102, -0.020132171, -0.0649216101, -0.0211254749, -0.0502667464, -0.0430250913, 0.0366390757, -0.0142103862, 0.0351996571, -0.0137910787, 0.0539601967, -0.0134565914, -0.0012004658, -0.0293155331, 0.0103800325, 0.0274328291, -0.0043067443, 0.0366836973, 0.0085573923, -0.0236731432, -0.0238660891, 0.0122575285, 0.0172105264, 0.0053221397, -0.0181067362, 0.0378015749, 0.0068998951, 0.0395847112, 0.0000016722, -0.066036135, 0.0102968039, 0.0050027776, -0.0382374041, 0.0360615142, 0.0003741868, -0.0172513202, -0.0017390569, -0.0059662238, -0.0343357138, 0.0049414812, 0.076004371, 0.0121422233, -0.0125820301, 0.0515366644, -0.0002210025, -0.0467795953, 0.0380872153, 0.00038673, -0.0155455926, -0.0044935602, 0.0000984098, -0.0022730313, 0.0051726433, 0.0124341734, -0.0348500684, 0.0186481811, -0.0003012073, 0.0094309421, 0.0105055794, 0.0251222048, -0.0177083798, -0.0567025803, 0.0014354453, 0.0542086884, 0.0175981093, -0.0210974962, -0.0324239545, -0.0347287282, 0.0341577381, 0.0425495319, -0.0409240611, 0.0268246159, 0.0106904386, 0.0346472338, 0.0209690332, -0.0327730253, 0.0166461747, 0.0040682023, -0.0196624808, 0.0683159977, 0.0321325175, -0.0495073125, -0.0173618458, 0.0077703148, -0.0242383461, -0.0551988259, -0.0363948345, 0.0201320108, -0.032705389, 0.0323304012, -0.0221975483, -0.0194709506, 0.0421864428, -0.0221219081, 0.0554126129, 0.0699386224, -0.0167757459, 0.0073492215, 0.0166646633, 0.072594665, -0.0634132996, 0.0269606691, -0.049926728, 0.0151434727, -0.0038122907, -0.0364057422, 0.0267358664, 0.0341105945, -0.0393144786, -0.014262381, -0.0389534347, 0.0563029014, 0.0067612207, 0.0132018132, -0.0257427916, 0.0039032879, -0.0237116572, -0.0037489401, -0.0260383952, 0.0959757343, -0.0370902307, 0.0286955424, 0.0156647321, -0.0382838249, -0.0019966853, 0.0241885353, -0.0270410366, -0.0486921035, -0.0256935433, 0.0348318629, -0.0145107415, -0.0049583078, 0.0139053054, -0.0092743831, -0.045647379, -0.0617478676, -0.0071664471, -0.0766205937, 0.0371082202, 0.0636866763, -0.0221193712, -0.0502082147, -0.0059205112, -0.0205274113, -0.0323378704, 0.0299008414, 0.0045683216, 0.0585635193, 0.0129387742, -0.0224921294, 0.0157930572, 0.0747436434, 0.0125176553, -0.0245921873, 0.0329640619, -0.0411867276, 0.0026179401, 0.0544021018, -0.0118411882, 0.0159851536, -0.0290665664, 0.0058329203, 0.0091257179, 0.0338252373, 0.078836374, 0.0300156716, 0.0058484222, -0.0248450954, 0.0865386501, 0.0464152694, 0.0135763995, 0.0072593554, 0.0045694201, 0.0062964852, -0.0035105657, 0.0616622977, -0.0729219317, 0.0127910469, -0.0374829099, 0.0627677664, 0.0190025866, 0.0561929233, -0.0309861377, -0.0217643492, -0.0518724211, 0.0243490543, -0.0092807347, 0.0100910692, 0.0361108556, -0.0278712679, -0.0087049948, 0.0450880602, 0.0429246426, -0.0355517603, -0.0045323055, -0.0421328768, 0.0578463525, 0.0106047578, 0.014862841, -0.0144028841, -0.0457501225, 0.0310500339, -0.0283576343, 0.0972952321, 0.057390634, 0.0251754466, 0.0105369454, -0.0231297668, -0.0233825035, -0.0337590873, 0.0086604245, 0.0004605264, -0.006738591, -0.0266847033, -0.0195211004, -0.041853562, -0.0098629091, 0.0247526746, -0.0106290597, -0.0063329423, -0.0612291396, 0.0090958821, 0.0314220414, 0.0050576292, -0.0002006701, 0.0287330076, 0.0146253137, 0.0003736414, 0.0251825619, -0.002715423, -0.037593253, -0.0091425767, 0.0505595878, 0.0146090621, -0.0185633209, 0.0428926013, -0.0000459363, -0.0330460332, 0.0068746861, -0.0100841094, 0.044724144, -0.0038538401, 0.0540025309, -0.0193479005, 0.0254370254, -0.0599037074, 0.0489343889, 0.0111266263, 0.0232095215, 0.0060284054, -0.0023749336, -0.0072392407, -0.0190556049, -0.018708868, 0.032077577, -0.0271735378, 0.0065167407, 0.040556971, -0.0453318357, -0.0492518246, -0.0272729602, -0.0552696735, 0.0419531912, -0.0116373682, 0.037681289, 0.061457593, 0.0208428819, -0.0172197763, 0.0391605347, -0.0491861477, -0.0027184102, 0.0286800396, 0.0546178445, 0.0382775404, -0.017387718, -0.0395479389, -0.0640740246, -0.0134185636, -0.0262476951, 0.0324501172, -0.0173776168, 0.0014267068, -0.0823405609, 0.0168036632, -0.0657658428, -0.0012659405, -0.0121856453, -0.0447705165, 0.0217402261, 0.0173270162, 0.0631330535, 0.0388722159, 0.0099395886, 0.030041052, 0.0587585196, 0.0200667493, -0.0764452964, -0.0045871139, -0.0436014682, -0.0462123528, -0.0610982031, -0.0496175662, -0.0130082099, 0.0903572962, 0.0279721152, 0.0028508946, -0.0490370467, -0.0080317194, 0.0000285556, -0.0479758605, -0.0043390333, -0.0370754264, -0.0398857556, 0.0434292182, -0.0280360561, -0.0791531652, 0.0427758619, -0.0125269927, -0.0245892666, -0.0150918327, -0.0028983187, -0.0338053182, 0.053820882, -0.0184481665, 0.0159889348, -0.0131337149, -0.031066753, -0.08901117, -0.0229841582, 0.008459473, -0.025267873, 0.1113638356, -0.0422107726, -0.0025077683, -0.0354113504, -0.0276030656, -0.0321173333, -0.0483438186, 0.006878552, 0.0226776693, 0.0002949252, 0.0055235652, -0.0609727055, 0.0095650349, 0.0662076995, -0.0121956058, -0.0330403484, 0.0190635715, -0.026491167, 0.0168014802, 0.0644613653, 0.063975513, 0.0441538729, -0.0429942161, 0.0030359651, 0.0204771385, 0.0133414567, 0.0204207804, -0.0238449257, -0.0384101383, 0.0688051358, -0.0355427526, -0.036561273, 0.0261200219, 0.0942706317, -0.008615844, 0.0100418944, 0.1040039435, 0.0223014783, -0.0267395247, 0.0166364554, -0.0537148155, 0.0619782396, 0.013652185, 0.0028504913, -0.0519204326, 0.0156205855, 0.0065245447, -0.0309118591, 0.0004192878, 0.0148893353, -0.0113604488, -0.1476614028, -0.0002591842, 0.0256497189, -0.0640689358, -0.0269542579, 0.0018552145, 0.0653943345, 0.016408246, 0.0514136478, 0.049232617, -0.020121349, -0.0081003029, 0.0134565504, 0.0154567482, -0.0361021534, -0.0207530744, -0.0537290536, 0.0612235703, 0.0601336509, 0.0082308818, -0.0186215583, -0.0345043838, 0.0433240794, -0.0357429348, 0.0236428138, 0.0224100836, -0.0063384324, 0.0210635047, 0.0169018395, -0.0185308736, 0.0169398561, -0.0101668928, -0.0116439089, -0.0067265746, 0.0495238602, -0.0118979421, -0.0092266584, -0.0423174873, 0.0116017843, -0.0252789985, 0.0777931362, 0.0318572707, -0.030666152, -0.0104379291, -0.0023578936, -0.0151961735, 0.0390702225, -0.0193968602, 0.0324431285, -0.0040860418, -0.0177710652, -0.0429043286, -0.0484387614, -0.0209233314, -0.0086202975, 0.0218774732, -0.0022821152, -0.0348139144, -0.0463088267, 0.0166086163, 0.0461026914, -0.0077485135, 0.0220214315, -0.0481983498, 0.0008452333, -0.0151120834, -0.044065658, 0.0013992133, -0.0144793103, -0.0508498885, -0.1121155694, 0.0200855937, -0.012563386, 0.0314468779, 0.0415874198, 0.0396147035, -0.0405946225, 0.0189115591, 0.0407688729, -0.0201353561, 0.0716517717, 0.0023135392, -0.0003957976, -0.0057446701, 0.0451032631, -0.0419990197, -0.005347068, 0.0528348349, 0.0618474223, -0.0065655569, -0.0514384396, 0.0298018698, 0.0178369675, 0.0479648449, -0.0142804887, 0.0050513139, -0.0208403058, -0.0050053187, -0.0260278918, -0.0127161676, 0.0078477338, -0.0551756918, 0.0311871804, 0.0019513069, 0.0176472608, 0.0356609225, -0.0071319863, 0.0622177236, -0.0603221655, -0.0013843094, -0.0217702519, -0.0336729586, 0.0178555585, 0.0553295724, -0.0303749405, 0.0350539349, -0.0295235571, 0.0330884419, 0.0132580716, 0.0703006759, 0.0137092397, 0.0077567515, -0.0358885303, 0.0123798857, -0.0482160971, -0.0225909688, 0.0152465198, -0.0119322743, 0.0256731194, -0.0191382244, 0.074149847, -0.0044669015, -0.0042239442, -0.0354264714, 0.0318193249, -0.0239496678, 0.0173345841, 0.0668972582, 0.0317682587, -0.0349109285, -0.0304632559, 0.0351455808, -0.033162389, -0.0010977724, -0.0218369048, 0.0174788218, -0.0068660807, -0.0484237857, 0.0313737206, -0.0016816903, -0.0375055447, -0.0121549005, -0.080522947, -0.0459704399, 0.009818757, 0.0039374824, -0.0051324465, -0.0496963486, 0.0090554282, 0.0789122134, -0.0201565493, -0.038241975, -0.0149101382, 0.0060919863, -0.0413215905, 0.0179208983, -0.0462339744, -5.960456512e-33, -0.0347836986, -0.0066757435, 0.017610075, -0.0521509908, -0.060333401, -0.0011600044, -0.0168194324, 0.0465368815, -0.0535397194, 0.0077965921, -0.0119874505, -0.0128134321, 0.0122957611, 0.0073721362, 0.0307678767, -0.0199361611, 0.0177653525, 0.031347163, -0.0578498095, -0.0309012271, 0.0302208476, -0.0334943756, 0.0713022128, -0.0499660671, 0.0365485474, -0.0245192908, 0.0021998009, -0.0020288869, -0.0309371296, 0.0843579397, -0.0100659914, 0.0694457293, 0.0334584378, -0.0143422512, -0.0108668124, 0.0403449424, -0.0058466904, -0.0372415595, 0.1130781621, 0.0031079771, -0.0078074737, -0.0364862718, 0.0529242642, -0.0062052123, -0.0433909968, -0.0422644056, 0.0383023061, -0.0263678003, -0.0258787256, 0.0112815648, -0.1014884487, 0.0033544046, -0.0039219367, 0.0026817033, -0.0355532616, 0.0287378188, 0.019576218, 0.0469313301, -0.0255194418, 0.0177870337, 0.0245078448, 0.0408831313, 0.1011369228, 0.0541198477, 0.0119711841, 0.0331236459, -0.013072269, 0.008718349, -0.0282636397, -0.0475127436, 0.0213988032, 0.0012256166, 0.0289787296, 0.0691116154, 0.0016141381, 0.0074449074, -0.0234694481, -0.0033315138, 0.0130541343, 0.0177020822, 0.0370838605, 0.0506150499, -0.052085638, -0.0113747288, -0.031647142, 0.0089354003, -0.0450217947, -0.0264530126, 0.0247631297, -0.0391138084, 0.0203212462, 0.0129728122, 0.0430274382, -0.0227586385, -0.0130747166, 0.0212263595, -0.0405852646, 0.0392897837, -0.0070841783, 0.0104590068, -0.0443307348, -0.0168487802, -0.0212657619, 0.0054790149, -0.002407725, 0.0498511046, 0.0116939303, 0.0065402361, -0.0686344951, 0.0050022923, -0.0371096209, -0.0290114153, 0.0031913826, -0.0538116507, 0.0565835126, 0.0005552942, -0.0124805756, 0.044365719, 0.0373341553, -0.0142319361, -0.0472383536, 0.0044397968, -0.0299766362, 0.0067036436, -0.0332679302, 0.0090148607, -0.0146175399, 0.0173139237, 0.0420181938, -0.016360756, -0.0399884954, 0.0144193545, 2.381e-7, 0.0713907033, 0.0052164299, 0.0269725304, 0.035301663, 0.0106407478, 0.0096564814, -0.0122171305, 0.0310438983, 0.033545319, -0.0196575392, -0.0170700606, 0.006497975, 0.0407030247, 0.005590193, -0.0237304363, 0.0393055305, -0.0827264786, 0.0397056527, -0.0513903014, 0.0522447005, 0.0186159406, 0.0921752155, 0.0035204908, 0.0088892076, -0.0322229452, -0.0230080653, 0.0576788411, -0.0378372669, -0.0066232341, 0.0435363017, -0.0209785532, -0.0334315635, 0.0237273294, 0.00451884, 0.0218677819, 0.0481475443, -0.0214359686, 0.0557246357, -0.0108366804, -0.0017892036, -0.0105595151, -0.0261545423, -0.0052122399, -0.0895145088, 0.0410461463, -0.0514144674, -0.0502229966, -0.0391989015, -0.0215049498, -0.0070217829, 0.0169722214, -0.0221916661, 0.0224323645, 0.0322514735, 0.0270697232, -0.0060418337, -0.0484828576, -0.038324412, -0.0722208768, 0.0253342483, -0.0536781959, -0.072999157, -0.0007642579, 0.0006960919, 0.0120698083, -0.0686959252, -0.0113438684, 2.189606643e-34, 0.0067587728, -0.0054057427, 0.0400150903, 0.048755046, -0.0018236203, -0.0428137258, -0.0132044666, 0.0463404097, 0.0280351844, -0.0843844563, -0.0190648064 ]
http://arxiv.org/abs/2406.17681v2
http://arxiv.org/pdf/2406.17681v2
VarBench: Robust Language Model Benchmarking Through Dynamic Variable Perturbation
2024-06-25T16:13:53
2024-06-26T15:21:49
Kun Qian, Shunji Wan, Claudia Tang, Youzhi Wang, Xuanming Zhang, Maximillian Chen, Zhou Yu
cs.CL
cs.CL
As large language models achieve impressive scores on traditional benchmarks, an increasing number of researchers are becoming concerned about benchmark data leakage during pre-training, commonly known as the data contamination problem. To ensure fair evaluation, recent benchmarks release only the training and validation sets, keeping the test set labels closed-source. They require anyone wishing to evaluate his language model to submit the model's predictions for centralized processing and then publish the model's result on their leaderboard. However, this submission process is inefficient and prevents effective error analysis. To address this issue, we propose to variabilize benchmarks and evaluate language models dynamically. Specifically, we extract variables from each test case and define a value range for each variable. For each evaluation, we sample new values from these value ranges to create unique test cases, thus ensuring a fresh evaluation each time. We applied this variable perturbation method to four datasets: GSM8K, ARC, CommonsenseQA, and TruthfulQA, which cover mathematical generation and multiple-choice tasks. Our experimental results demonstrate that this approach provides a more accurate assessment of the true capabilities of language models, effectively mitigating the contamination problem.
VarBench: Robust Language Model Benchmarking Through Dynamic Variable Perturbation: As large language models achieve impressive scores on traditional benchmarks, an increasing number of researchers are becoming concerned about benchmark data leakage during pre-training, commonly known as the data contamination problem. To ensure fair evaluation, recent benchmarks release only the training and validation sets, keeping the test set labels closed-source. They require anyone wishing to evaluate his language model to submit the model's predictions for centralized processing and then publish the model's result on their leaderboard. However, this submission process is inefficient and prevents effective error analysis. To address this issue, we propose to variabilize benchmarks and evaluate language models dynamically. Specifically, we extract variables from each test case and define a value range for each variable. For each evaluation, we sample new values from these value ranges to create unique test cases, thus ensuring a fresh evaluation each time. We applied this variable perturbation method to four datasets: GSM8K, ARC, CommonsenseQA, and TruthfulQA, which cover mathematical generation and multiple-choice tasks. Our experimental results demonstrate that this approach provides a more accurate assessment of the true capabilities of language models, effectively mitigating the contamination problem.
[ -0.0150005706, 0.1116496101, -0.0259464495, 0.0310791228, -0.0432360657, -0.0094744908, 0.0038597914, -0.0031800929, -0.0331641398, -0.0497521535, -0.0142905368, -0.0362251848, 0.0062289061, 0.0440837033, 0.0593248419, -0.0316394828, 0.0546100922, -0.031656649, -0.0094251512, 0.0012431161, -0.0156355333, -0.0292780735, 0.0090763541, 0.0307900403, -0.0153251952, -0.0120109087, -0.0029631308, -0.0134279942, 0.0042276834, -0.0157259423, 0.0189361684, 0.086811766, -0.0110086845, 0.0482972302, 0.0000022189, -0.0321116075, -0.0042703515, 0.018118402, -0.0277917907, 0.0383180603, 0.0553713404, 0.0139472447, -0.0041151703, 0.0350958817, -0.0287874248, -0.0367510989, 0.0255647264, -0.0523133241, 0.0598448068, 0.0645267069, 0.0037368487, 0.0410202332, -0.0149122048, -0.0381702483, -0.0146474838, -0.0618139729, 0.0127279181, -0.0004120886, -0.0059971907, -0.0083906958, -0.0222581699, 0.0350564048, 0.0339754, 0.0123624504, -0.0268246382, 0.0291573051, -0.0801339969, -0.0134306354, -0.0034993687, 0.0695848018, -0.0181836989, -0.0374907702, 0.0039143632, -0.0177208986, 0.0041732322, 0.0134037687, -0.0381792113, 0.0042415094, 0.0120575568, -0.0080447076, -0.0103897145, -0.0106412806, 0.0357470624, -0.0037155342, 0.002060306, 0.0562815033, 0.0436856188, -0.0294029936, -0.0081769526, 0.0133023495, -0.0318605751, -0.0433262251, 0.0073524499, -0.0070650219, -0.0054865717, 0.0342294425, -0.0742054358, -0.0759078339, 0.0258021392, -0.0513422675, 0.0580790229, 0.0304986574, -0.0060030813, 0.0090436721, -0.0019927246, 0.0913252309, 0.0209992826, 0.046464853, -0.0423918441, 0.0276933182, -0.0105054751, -0.0127305267, -0.0009873082, 0.0197684858, -0.0116957007, 0.0144218579, -0.0401749313, 0.0053563747, -0.0272722803, 0.0126510253, -0.0033804437, 0.0029988538, -0.0407758653, 0.007235019, 0.0133593539, -0.0039467309, -0.041985631, 0.0213803258, 0.0241100173, -0.0727919564, -0.0171163939, 0.0075604143, -0.019491829, -0.0231603552, 0.0076423353, 0.0796656907, 0.0140208015, -0.0182952024, 0.0130877197, -0.0481301397, -0.0205754172, -0.0652957261, -0.0056566042, -0.0713842213, 0.0170875993, 0.0053638397, -0.0144343218, -0.0217530373, -0.0014927647, 0.0315453, 0.0095184827, 0.0254119001, 0.0198675077, -0.0085956473, -0.0159869213, 0.0056999275, -0.0051875142, 0.0812697038, 0.035773851, -0.0073501309, 0.0464624763, -0.0226261355, 0.0052493126, 0.0692391694, 0.0445082411, 0.0039487686, -0.0591361634, 0.0092503857, 0.0571270138, 0.0227361675, -0.0041710394, 0.0511348844, -0.0397168063, -0.0145758875, 0.0794371143, 0.0466596559, -0.0194981527, -0.0054037306, -0.0152543737, -0.027131252, 0.0318849348, 0.0019630766, -0.0595036075, -0.0337958932, -0.0190618839, 0.0166723114, -0.0013295941, 0.0307437349, -0.0157106053, -0.048756931, -0.0280640759, -0.0112611102, 0.0142906317, 0.0103841117, 0.0528182052, 0.0059428168, -0.0223862436, 0.0632253438, -0.0008274455, -0.0750767365, 0.0162018053, -0.0420055315, 0.0040440587, 0.034697447, 0.0695381612, -0.0071123084, -0.0267381743, -0.0125446795, 0.02553159, 0.031247152, 0.0666941106, -0.0147052528, 0.0348661654, -0.0269140303, -0.0240795203, -0.0003536675, 0.0305932779, 0.0204890072, 0.0300960485, 0.0123301698, 0.0136900954, -0.0749322549, -0.0388797037, 0.0503801517, -0.0443048999, -0.0373188965, 0.0174803603, -0.0235055052, 0.0199482273, -0.0019412852, -0.0239248984, 0.0275620576, 0.0242175106, -0.0101986649, 0.0096382555, 0.0066680596, -0.0392044, -0.0295260344, 0.0524618886, -0.0264669303, -0.006503596, 0.0841541439, -0.0427382961, -0.0229186993, 0.0350478627, -0.0034978869, -0.0251780357, -0.0241257884, 0.0456495546, -0.0427286029, 0.0220367592, -0.0303267837, 0.0493191518, -0.0090149585, 0.0364490561, -0.0278344378, 0.0563285984, 0.0222728569, -0.0095287403, -0.0361311622, 0.002985314, -0.0220506229, 0.0386708975, 0.0079149762, 0.007243346, -0.0369620994, -0.0291347448, -0.0627688095, -0.002315178, -0.0274257101, 0.0292571038, 0.0220629778, -0.0132725155, 0.002858686, 0.008482418, -0.0335273966, -0.0308154561, 0.02419338, 0.0193805546, 0.0153885558, 0.0090159299, -0.0069368058, -0.0581392758, -0.0253978949, -0.0317989551, 0.0836160034, -0.0035789746, 0.0469658449, -0.0263228621, 0.0255531427, -0.0412241518, 0.0174809135, 0.0055799703, -0.0268051028, 0.0249293651, -0.0010886687, 0.0323594697, 0.0793883726, -0.0044901194, -0.0056865574, 0.0586036779, -0.0097452849, -0.0669040978, -0.0116389086, -0.0105077811, -0.0229575764, -0.0049871271, -0.0223275926, 0.0178447887, 0.0442577451, 0.0088172909, -0.0222712681, -0.0438296981, 0.0621283799, -0.0390918776, -0.0305798519, -0.0461938605, 0.0054722363, -0.0585680604, 0.0391399078, -0.0287081301, -0.078489095, 0.0704375952, -0.0293859448, 0.0172852725, -0.0449134894, 0.0200427808, -0.0207151715, 0.0273029227, -0.0038484689, -0.0101985922, 0.0007816466, 0.003252421, -0.0383763537, -0.0073073744, 0.0412640199, -0.066391319, 0.0533013232, -0.0565831028, 0.004465498, 0.0038821106, -0.0571076609, -0.0476159267, -0.0338936038, 0.0211412087, 0.0241322741, 0.0338170603, -0.0107840607, -0.0023634434, -0.0079764919, 0.0032452948, 0.0692456216, -0.0394908935, 0.0068585933, -0.0188614689, 0.0194277726, 0.0429304093, -0.0222208034, 0.0229375567, -0.0656190589, 0.0128953522, 0.0498627983, 0.0735425279, 0.0287000425, -0.0242731478, 0.0154654104, 0.0074511194, -0.0032718738, -0.0014089154, 0.0344440565, 0.0290937852, -0.0641912818, -0.0453309231, 0.0856520087, 0.0280429255, -0.0004436657, 0.0221029818, -0.0321916379, 0.0829609632, 0.0348059349, 0.0314081721, -0.0857224762, 0.00591473, 0.022814149, 0.0603389032, -0.0028602732, -0.0324148089, -0.0151160983, -0.1450094134, 0.0116185071, 0.0201666281, -0.0665871426, 0.0245358758, -0.0340634771, -0.001510903, -0.0236962195, 0.0456111915, 0.0153512442, 0.0148344794, 0.0076166489, 0.0332733393, -0.025789123, -0.1098464206, 0.0232538376, -0.0434703901, 0.0461340994, 0.0525779948, 0.0106808273, -0.0374902524, 0.003453935, 0.0537255742, 0.0032032044, -0.0367178209, 0.0269100573, 0.045141317, 0.0158351343, -0.0231130142, 0.0106365774, 0.0117321592, -0.028282417, 0.0048524756, 0.0111027481, 0.0656312481, 0.0245219562, -0.0241197888, -0.0002721481, -0.0308362022, -0.0015874733, 0.0752604231, 0.0464575551, -0.0471014827, -0.0282278918, -0.045884937, -0.0070635588, -0.0587450974, -0.0154911401, 0.0079395203, 0.019570034, -0.0317193717, -0.0685616359, -0.0403601788, 0.0249529779, 0.063045986, -0.0030967854, 0.0086336965, 0.0060259034, -0.0195679236, -0.0422464386, 0.0445066169, -0.0140153533, 0.0240508467, -0.0608028956, 0.0034713282, 0.0024864089, 0.0351435244, -0.035793025, 0.0577671826, -0.1281450093, -0.1100882366, 0.0172838848, 0.0823279098, 0.022504814, 0.0308031645, -0.0069826627, -0.0595959648, 0.0351918414, 0.0099934237, -0.013271247, 0.0363402218, 0.0181010012, -0.0719899982, 0.0099422205, 0.0174752027, -0.0328649394, 0.0036350072, 0.0163812209, 0.0319962949, -0.0061946074, -0.0110225463, 0.0125455596, 0.0315395147, 0.0158415772, -0.0152010862, -0.0167123936, -0.0161840711, 0.0084763961, -0.0483587608, -0.0173135288, -0.0351204388, 0.0167514551, 0.0117349457, 0.006368787, -0.0061714044, -0.0149577186, -0.012881659, 0.0673577636, -0.0355792157, 0.0131307393, -0.0160299744, 0.0580479912, 0.0303623155, -0.0019974594, -0.0143812448, -0.0009966849, -0.0293052476, 0.073018305, -0.0363870673, -0.002613148, 0.0135958157, 0.0220568515, 0.0205709264, 0.0099404119, -0.0618358403, -0.0522265658, -0.0445397831, 0.0077645769, -0.0015300803, 0.0339977667, 0.0404692367, -0.0266770981, -0.0178928226, -0.0418184623, -0.0121920751, -0.0077181263, -0.0409083404, 0.0268821456, 0.0185573455, -0.0202850346, 0.0148645472, 0.0827505961, 0.0503537506, 0.0343274772, -0.0239906814, 0.0314177908, -0.0459287278, -0.0099594835, 0.0032145714, 0.000411326, -0.0247242991, -0.0241437107, -0.0543032847, -0.1045504361, 0.0388551652, 0.0342499204, 0.0284438152, -0.0035191318, 0.0039269738, 0.0532322191, -0.0227503143, 0.0066461246, -0.0282312538, 0.004271518, 0.0066057811, -0.0296054296, -0.0278171822, -6.91079741e-33, -0.0222538207, -0.0154750003, 0.0054915375, -0.0465563908, 0.0007611893, 0.0180234239, -0.0308355466, 0.0306542944, -0.0116138384, -0.0030870403, -0.0166510846, 0.0158966724, 0.0183380339, 0.0249072053, -0.0017875403, -0.0297643896, 0.0368947499, -0.043481376, -0.0650424883, 0.0154188592, 0.0120654209, 0.0324322283, 0.0761905462, -0.0734236389, 0.007009645, 0.006283382, -0.0356108546, -0.0012548582, -0.0349100977, 0.0728232339, 0.0085065356, 0.035237316, 0.0026712522, -0.0296842884, -0.035773281, -0.0121857459, -0.0310466141, -0.0321369283, 0.0122338459, 0.0258731358, -0.0064754412, -0.0063077891, 0.0523440726, -0.0150974542, -0.0487431549, -0.0108551262, 0.0080656577, -0.0329450704, -0.0599128753, -0.0148727531, -0.0272394717, 0.0278572422, -0.0052410145, 0.022927383, -0.0019272942, 0.0690585151, -0.0257264022, -0.002643307, -0.0414127633, 0.0104251532, 0.0567870475, 0.053061787, 0.0722322166, 0.0316204838, 0.0221053381, 0.0767420158, 0.0231788754, 0.0143848807, 0.0325137228, -0.0094374428, 0.090106301, 0.01391852, 0.075557746, 0.0668181404, 0.0455976427, -0.0196038242, -0.0600700565, 0.0129183475, 0.0146272853, -0.0055950382, 0.0357813425, -0.0328718089, -0.0679443628, -0.03366993, -0.0915003717, -0.0103012221, -0.0581916012, -0.0171520822, 0.0721701533, 0.0035163076, -0.037418332, -0.0106078126, 0.0328541324, -0.0325018689, -0.0368002988, 0.0517481938, -0.0209370274, 0.0290527157, -0.0121189859, 0.042982094, -0.0294722859, -0.0144446902, 0.0038709952, 0.0274844896, 0.040311221, -0.0025197545, 0.0069389311, 0.0322791636, -0.0646598265, 0.0108351922, -0.0345462076, 0.020309424, 0.0334128663, 0.0459102727, 0.0577238239, -0.0283446573, 0.0100845825, 0.0376649462, 0.0263839066, -0.0000474951, -0.0441730022, 0.0094200335, 0.0289062355, 0.0441186354, 0.0055846712, -0.0242365915, -0.0242096111, -0.0091111064, 0.0557927005, -0.0309463721, -0.0053840065, 0.0112893591, 2.868e-7, 0.0447458997, 0.0598571599, 0.005418241, 0.0286996309, -0.0218247734, 0.0181656014, -0.0606483966, 0.0191356, 0.0081975628, 0.0041463673, -0.0282914285, -0.002338683, 0.0758820474, 0.036670547, -0.056013152, -0.0156421103, -0.0640190542, 0.017447792, -0.0650358498, -0.0344630554, 0.019829547, 0.0890927017, 0.0057084793, -0.0154089304, -0.0226286631, -0.0388958938, 0.0551089644, -0.0261649694, -0.0302051678, 0.0303664953, 0.0260091219, 0.0287681632, 0.0177735351, -0.0143785691, -0.0397435203, 0.0201772805, -0.0224491246, 0.0696376711, 0.0176258963, -0.0310600474, 0.011010942, -0.0245178305, -0.0246251021, -0.0357716158, 0.042904187, -0.0298957117, -0.0044120993, -0.0465602577, -0.0277455971, -0.0330225155, 0.0551450588, -0.0442583784, -0.0182515904, 0.0222013202, 0.0243553873, -0.0024606187, -0.0070953686, -0.0405334719, -0.049166292, -0.001793657, -0.0454338565, -0.0376821719, -0.028249152, 0.0277636871, -0.01122416, -0.0277185086, 0.0282794945, 3.183853311e-34, 0.0538518392, -0.0300196446, 0.0319730751, 0.0254127961, 0.0408509858, -0.0093601476, 0.0028013536, 0.0049259663, 0.0074428348, -0.0610949844, -0.0179279353 ]
http://arxiv.org/abs/2406.17675v1
http://arxiv.org/pdf/2406.17675v1
Quantifying AI Psychology: A Psychometrics Benchmark for Large Language Models
2024-06-25T16:09:08
2024-06-25T16:09:08
Yuan Li, Yue Huang, Hongyi Wang, Xiangliang Zhang, James Zou, Lichao Sun
cs.CL
cs.CL
Large Language Models (LLMs) have demonstrated exceptional task-solving capabilities, increasingly adopting roles akin to human-like assistants. The broader integration of LLMs into society has sparked interest in whether they manifest psychological attributes, and whether these attributes are stable-inquiries that could deepen the understanding of their behaviors. Inspired by psychometrics, this paper presents a framework for investigating psychology in LLMs, including psychological dimension identification, assessment dataset curation, and assessment with results validation. Following this framework, we introduce a comprehensive psychometrics benchmark for LLMs that covers six psychological dimensions: personality, values, emotion, theory of mind, motivation, and intelligence. This benchmark includes thirteen datasets featuring diverse scenarios and item types. Our findings indicate that LLMs manifest a broad spectrum of psychological attributes. We also uncover discrepancies between LLMs' self-reported traits and their behaviors in real-world scenarios. This paper demonstrates a thorough psychometric assessment of LLMs, providing insights into reliable evaluation and potential applications in AI and social sciences.
Quantifying AI Psychology: A Psychometrics Benchmark for Large Language Models: Large Language Models (LLMs) have demonstrated exceptional task-solving capabilities, increasingly adopting roles akin to human-like assistants. The broader integration of LLMs into society has sparked interest in whether they manifest psychological attributes, and whether these attributes are stable-inquiries that could deepen the understanding of their behaviors. Inspired by psychometrics, this paper presents a framework for investigating psychology in LLMs, including psychological dimension identification, assessment dataset curation, and assessment with results validation. Following this framework, we introduce a comprehensive psychometrics benchmark for LLMs that covers six psychological dimensions: personality, values, emotion, theory of mind, motivation, and intelligence. This benchmark includes thirteen datasets featuring diverse scenarios and item types. Our findings indicate that LLMs manifest a broad spectrum of psychological attributes. We also uncover discrepancies between LLMs' self-reported traits and their behaviors in real-world scenarios. This paper demonstrates a thorough psychometric assessment of LLMs, providing insights into reliable evaluation and potential applications in AI and social sciences.
[ 0.0188376829, 0.0682794005, -0.0549192056, 0.0054753134, -0.0668005645, 0.0313297398, 0.0168988351, -0.0259659998, 0.031778533, 0.0036778103, -0.0636041462, -0.0316019319, -0.0074516293, 0.0451932997, 0.0032755774, -0.0282987058, 0.0187542401, -0.0011429776, 0.0543668233, 0.0052549741, -0.0359095149, 0.0400601849, -0.0093163149, 0.0200012121, -0.0027239136, -0.0212626364, 0.0079687191, -0.0319207422, 0.0359672941, -0.0366268679, 0.0033264449, 0.0271658357, 0.00861682, 0.0217129868, 0.0000019374, -0.0861850008, 0.0136568295, -0.0117822057, -0.0201844536, -0.025819812, 0.0583159029, 0.0296978168, -0.0015603751, 0.0352091156, -0.071577765, 0.0071731769, 0.0068808165, 0.0469818451, 0.0233230833, 0.060669709, -0.0166185126, -0.0017599144, 0.0409489386, 0.003466598, 0.0119263995, -0.0397579148, 0.0309807155, -0.0276461318, -0.030020697, -0.0433659852, 0.0297681559, -0.0154328635, 0.0231373254, 0.0168541502, -0.0290816985, -0.001590955, -0.056172464, -0.0231601484, 0.0031791085, 0.0506765358, 0.0327131599, 0.0017713383, -0.0091412282, -0.0075631728, 0.0182351861, -0.0029004773, -0.0319816209, 0.0300640017, 0.001707125, -0.0123768831, 0.0646202639, 0.0302181039, 0.0206522588, 0.0101796985, -0.0211366024, 0.0615188666, -0.0045302785, -0.0420655757, 0.0350857303, -0.0065416796, -0.0169478059, -0.0597178712, -0.0124851028, 0.0264140218, -0.039295584, 0.027449213, -0.0377973877, -0.015691936, 0.0218868293, -0.0474747531, 0.0662583634, 0.0273528174, -0.025813397, 0.0258389544, 0.0374331214, 0.0923016369, 0.038402658, 0.0319671258, -0.0333316028, -0.0169094838, -0.0722775012, 0.0088230884, 0.0236003902, 0.0172546618, 0.053911373, 0.0014065998, -0.0395429507, 0.0185039677, -0.0303148925, 0.0087828971, -0.0076632979, -0.0128241768, 0.0153452139, -0.0038279134, -0.0558357127, 0.0147903338, -0.0441084839, 0.0115996199, 0.0188892446, -0.0463980138, 0.0482561551, -0.0011531475, 0.0183448624, -0.0377508067, 0.0575547777, 0.0592981838, 0.0305856448, -0.0595479794, -0.0413034633, -0.0157209095, 0.0213517454, -0.0882495716, -0.0031143462, -0.0365425088, 0.0401210748, 0.0323088653, -0.0211844631, -0.058894366, -0.0098054027, 0.0273781437, 0.0003135357, 0.0359042771, 0.0276830383, 0.0215643756, 0.0211240035, 0.0172229446, -0.0082306871, 0.0294226967, 0.0280622188, 0.007888183, 0.0656713843, -0.0331558399, 0.0300430041, 0.0122880088, -0.0433081761, 0.0029725896, -0.0252571627, 0.0136388978, -0.0171026569, 0.0123590799, 0.0256499555, 0.0552342273, 0.0180507936, -0.0164122377, 0.0621292032, 0.0863106623, 0.0069460594, 0.0287920069, 0.0163364857, 0.0082024941, 0.0744948089, -0.0044374978, -0.0556252487, 0.0225505307, -0.0302228592, -0.0077565783, 0.0042881859, 0.0462410711, -0.020881623, -0.0611510277, -0.0210275277, -0.0185776185, 0.0247254577, 0.0394943245, 0.03495517, -0.0545621887, -0.0073014465, -0.0504014753, -0.0050343261, -0.0572161749, -0.027458312, -0.0743572861, 0.0125048179, 0.0797366351, 0.0238518156, -0.0249077305, -0.0256708562, 0.0069541433, -0.0314445458, 0.0145725543, 0.0579033233, -0.0228946805, -0.0096394988, -0.0210897941, 0.0138038602, -0.0044707111, 0.0083329529, 0.0108554037, -0.0174205229, 0.030376073, -0.0050078114, -0.036850594, -0.0312413592, 0.0296842083, -0.0829395726, -0.0369887389, -0.0126102632, 0.0133419596, 0.0590054654, 0.031067526, -0.0125941942, 0.0139673101, 0.02458735, -0.005036036, 0.0162201803, 0.0276842341, 0.0092219459, 0.06126865, 0.0430428758, 0.0251291748, -0.0089296512, 0.0579525977, 0.0294688214, -0.0224750135, 0.0027236138, -0.0192203056, 0.0222836882, -0.0069917492, 0.0512621105, -0.0571732745, 0.0514675863, 0.0004744215, 0.0584812127, -0.0170548018, 0.0329642445, 0.0193191916, -0.0074556149, 0.036028225, -0.0330524966, -0.0085062068, -0.0252306443, 0.0016065763, 0.0168650839, 0.0292824879, -0.0103134708, -0.0608037375, -0.0395763703, -0.0898556858, -0.0129352482, 0.0147470478, 0.0231220257, 0.0125614321, -0.019797232, -0.052049499, 0.0274453629, -0.0544890575, 0.0085249767, 0.0485477187, 0.0463723354, 0.0135665257, -0.0282219108, -0.0121163912, -0.0704337135, 0.0033004729, -0.0318738297, -0.0143820038, 0.0030142141, 0.0558945015, -0.0316339172, 0.0503299497, -0.0442415848, 0.0119649386, -0.0270672292, -0.0589866154, 0.0479677096, 0.007000824, 0.0320192948, 0.1233093292, -0.0010626682, 0.0282171164, 0.0636031255, 0.0175687559, -0.041122783, -0.0123544978, 0.0113971224, 0.0056726127, -0.0335663036, -0.0678236037, 0.0182233658, 0.0532616712, -0.0060890205, 0.0148147503, -0.0207009707, 0.0473425761, -0.0191905852, -0.0512557141, 0.0048077083, 0.0116618825, -0.0210710522, 0.1059692055, 0.003247801, -0.0500656739, 0.0545117557, 0.0039431201, 0.0235124398, -0.0177321509, -0.0476437919, -0.0359089263, 0.0844005942, -0.0018220939, -0.0173305236, 0.0126990844, 0.0187254753, -0.0535471588, -0.0165338293, -0.0361937098, -0.0569808036, 0.0128757833, -0.0184710193, -0.0336364992, 0.0239222236, 0.0061511132, -0.050244458, -0.039940998, -0.0503911488, 0.0034560938, 0.025385445, -0.064918533, -0.0374028198, -0.0513493232, 0.0651984662, -0.0117092403, -0.0049130232, -0.021483358, 0.0174817257, -0.0329125151, 0.0880290568, -0.0000494725, 0.0572115183, -0.0411589779, 0.0356282778, -0.0113550145, 0.0194196068, 0.0223866794, 0.0048443135, 0.0049707331, 0.0192509666, -0.0015087075, -0.0047763819, 0.0435683392, 0.0202009492, -0.0082863318, -0.0118929371, 0.1442836523, 0.0383723713, -0.0462394431, 0.0200268589, -0.0172129255, -0.0035922872, 0.0396333784, 0.0367093645, -0.0293124765, 0.0204934217, 0.0006664585, 0.0188948289, -0.0549142472, -0.0543462373, -0.0057549416, -0.0984215066, 0.016221175, -0.0220627766, -0.0056811911, -0.0277906116, -0.0532017648, 0.0296839569, -0.0072018807, 0.0173034407, -0.0341135934, 0.0071253078, -0.0298641082, -0.0043942989, -0.0402005054, -0.0518030189, 0.1095785797, -0.0354112312, 0.072129108, 0.0675022379, 0.0594542101, -0.0295667462, -0.0556334667, -0.0390499495, 0.0242338069, -0.0090929624, 0.0443409644, 0.0460974462, -0.0285363104, 0.0132324165, -0.033714965, -0.0019653363, -0.0545907989, 0.0037825864, 0.0150309037, 0.0789054781, 0.0070378319, -0.0379006192, -0.0016682722, -0.0278082602, 0.0016029548, 0.1022463292, -0.0016060503, -0.0673030764, 0.0064908173, -0.0022711542, -0.0039477935, -0.0198322143, -0.0625922978, 0.0247611329, 0.0096558677, 0.0085434346, -0.0338900089, -0.0851332694, 0.0397213139, -0.019551402, -0.0116766226, -0.0109167388, 0.0179886967, 0.0040349104, -0.0580516569, 0.0018309095, 0.0017469901, 0.027461797, -0.042248331, -0.0209505633, 0.0324231572, -0.0386116095, 0.0025473402, -0.0192641597, -0.0293510184, -0.0298524909, 0.0483219065, 0.0190509707, 0.0221724529, -0.0287244748, 0.001041838, -0.0532843433, -0.0068558259, -0.0304011125, -0.0421909764, 0.0382949375, -0.0072398996, 0.0521969981, 0.0086745648, 0.0315122269, -0.0503328629, 0.0222766455, -0.0078900363, -0.0308431685, -0.0084033282, -0.0793432742, 0.0183414351, 0.0423684493, -0.0021774897, -0.0108213024, -0.0080683194, -0.0190314129, -0.0505895466, -0.0638416633, -0.0221845284, -0.0249961726, -0.0665364042, 0.0396379568, 0.0150904292, -0.0156395547, -0.0411743373, -0.0257733092, -0.0134890536, -0.0304469932, 0.0242648032, -0.0417450815, 0.0362599753, -0.0053680963, 0.0191971362, -0.0161433015, -0.0352125056, -0.0147790937, 0.0468479171, 0.0235627461, 0.0766034499, -0.0136353858, -0.0008287685, 0.0167288315, 0.0475025587, -0.0551339462, -0.0454797447, 0.0519781671, -0.0044666384, 0.0016951944, -0.0145391999, 0.0569383092, -0.0442784317, 0.0287828874, -0.0377192609, -0.0083268182, 0.0045656343, -0.0106352763, 0.0044585085, 0.0147586968, -0.016406633, 0.002273584, 0.0590231195, 0.0058353469, 0.0375576392, -0.0160908736, -0.0175703894, -0.0358795226, 0.0275862813, -0.0198448449, 0.0391011685, -0.0320066288, -0.0087437397, -0.0797638446, -0.0825992227, 0.0337442309, 0.0079606911, 0.0070458725, -0.0407322273, -0.0022411477, 0.0739539415, -0.0448741131, -0.0085645504, -0.0271267127, -0.0362380892, -0.0654638261, -0.0127176195, -0.0370288715, -6.094006997e-33, -0.0008659495, -0.0159471687, 0.017342627, -0.000397565, -0.0092071211, -0.0256824959, -0.0276742242, 0.0320179388, -0.0558067039, -0.0086140269, -0.0478286296, 0.0196017381, 0.0041138213, -0.0085078049, 0.0295873806, 0.0025049024, 0.0027404306, -0.0561720245, -0.0231778473, 0.0019435047, 0.0064098928, -0.0210419334, 0.0603494383, -0.1006532833, 0.0425844863, -0.0130532067, 0.0174812227, 0.0143864919, 0.000164154, 0.0576640405, -0.0126800416, 0.0553973727, 0.0528015085, 0.0039397837, 0.0035457909, 0.0213676244, 0.0001251903, -0.0661022216, 0.0163044147, 0.0143801998, 0.0008083562, -0.0624884255, 0.0015846243, -0.0093092555, -0.0595912524, -0.0140292402, 0.0363631397, -0.0239283182, 0.0541363694, 0.0229340252, -0.0507730842, -0.0101824515, -0.0140966428, -0.033793401, -0.0066374801, 0.0343163088, -0.0301645435, -0.0128457611, -0.0585328564, 0.0286975373, -0.0094878087, -0.0089343041, 0.0610179491, -0.0159052722, 0.0188496411, 0.0546836294, 0.0861287788, 0.0555898547, -0.0127712023, -0.012937801, 0.041845981, 0.0582273528, 0.0418459661, 0.038534645, -0.0002678001, -0.017354209, -0.0062291822, 0.0303760916, 0.0810101405, -0.0103477668, 0.0440656506, 0.0000959875, -0.0449654311, -0.0202341639, -0.0266957637, -0.0072588795, -0.021128986, -0.0113951918, 0.0489711873, -0.0295317098, 0.033472091, -0.0428159535, 0.0436331443, -0.0329576172, 0.0151187088, 0.0321496166, -0.0485798568, 0.0661664084, 0.0188804567, -0.0079834107, -0.0663905665, -0.0002711247, -0.0076057091, 0.0480997376, 0.0672387481, 0.001447573, -0.0052706786, 0.0356653966, -0.0402552076, -0.0107645504, -0.0075629493, 0.0088105798, 0.0296100192, 0.0034634285, 0.0061657592, -0.0112984516, 0.0024285575, -0.0134809995, 0.0093828691, -0.0186878089, -0.0621935837, -0.0015671754, -0.0086017558, -0.0103451367, -0.0084186587, -0.014299727, -0.0674014986, 0.0093361298, 0.0345470496, -0.0031361822, 0.0138309514, -0.0212512296, 2.742e-7, -0.0218456164, 0.0769092068, -0.0273201689, 0.0582003295, -0.0064141164, 0.0544417724, -0.0252343249, -0.0061751571, 0.0323981233, -0.0169244166, 0.016054783, 0.0203158595, 0.0523779318, 0.0222432204, -0.0401504785, -0.0081101377, -0.0081667723, 0.0122647258, -0.0402798429, 0.0260430127, 0.0643899888, 0.0400401652, 0.0289019551, -0.0077565513, -0.0290513653, -0.0091282818, 0.0085327011, -0.041241698, 0.0054501812, 0.0100738993, 0.0214843769, 0.0231666248, 0.0139302611, 0.0486598909, 0.009233064, 0.0183050409, -0.0231437534, 0.0442724712, -0.0214534383, -0.0344703794, 0.0137911718, -0.0017562358, -0.0311295297, -0.046279341, 0.0617787018, -0.0302550439, -0.0868686363, -0.0465793684, -0.0081062764, -0.0198704973, 0.0345642492, 0.0036338498, -0.019712327, 0.0046093827, 0.0240686126, -0.0375384651, -0.0080000935, -0.0436089411, 0.0224651955, 0.0055554076, -0.0066833035, -0.059940923, -0.0407823585, 0.0193076171, 0.0162325241, -0.0657705367, 0.0245939605, 2.2513628e-34, -0.0118006207, 0.0214720536, 0.023687996, -0.000479621, 0.0752018243, 0.0055571916, -0.0435977727, 0.0305873062, 0.0185544323, -0.0956120417, 0.0060741492 ]
http://arxiv.org/abs/2406.17667v1
http://arxiv.org/pdf/2406.17667v1
This Paper Had the Smartest Reviewers -- Flattery Detection Utilising an Audio-Textual Transformer-Based Approach
2024-06-25T15:57:02
2024-06-25T15:57:02
Lukas Christ, Shahin Amiriparian, Friederike Hawighorst, Ann-Kathrin Schill, Angelo Boutalikakis, Lorenz Graf-Vlachy, Andreas König, Björn W. Schuller
cs.SD
cs.SD, cs.CL, eess.AS
Flattery is an important aspect of human communication that facilitates social bonding, shapes perceptions, and influences behavior through strategic compliments and praise, leveraging the power of speech to build rapport effectively. Its automatic detection can thus enhance the naturalness of human-AI interactions. To meet this need, we present a novel audio textual dataset comprising 20 hours of speech and train machine learning models for automatic flattery detection. In particular, we employ pretrained AST, Wav2Vec2, and Whisper models for the speech modality, and Whisper TTS models combined with a RoBERTa text classifier for the textual modality. Subsequently, we build a multimodal classifier by combining text and audio representations. Evaluation on unseen test data demonstrates promising results, with Unweighted Average Recall scores reaching 82.46% in audio-only experiments, 85.97% in text-only experiments, and 87.16% using a multimodal approach.
This Paper Had the Smartest Reviewers -- Flattery Detection Utilising an Audio-Textual Transformer-Based Approach: Flattery is an important aspect of human communication that facilitates social bonding, shapes perceptions, and influences behavior through strategic compliments and praise, leveraging the power of speech to build rapport effectively. Its automatic detection can thus enhance the naturalness of human-AI interactions. To meet this need, we present a novel audio textual dataset comprising 20 hours of speech and train machine learning models for automatic flattery detection. In particular, we employ pretrained AST, Wav2Vec2, and Whisper models for the speech modality, and Whisper TTS models combined with a RoBERTa text classifier for the textual modality. Subsequently, we build a multimodal classifier by combining text and audio representations. Evaluation on unseen test data demonstrates promising results, with Unweighted Average Recall scores reaching 82.46% in audio-only experiments, 85.97% in text-only experiments, and 87.16% using a multimodal approach.
[ 0.034656778, 0.0612476021, -0.027323192, 0.0237244722, -0.0117347278, 0.005310236, -0.0082539152, 0.0070892754, -0.0374903567, -0.0513358153, -0.0844535753, 0.0055534705, 0.0026811061, 0.0130112395, 0.0380510874, -0.0165573079, 0.0247216746, 0.0274254456, 0.0277221147, -0.0114603927, 0.0532845259, 0.0120180296, -0.0008517693, 0.0321612991, -0.0576925837, -0.004152216, 0.0071665687, 0.0324978456, -0.0200461708, 0.0024122836, 0.0171628986, 0.0479908027, 0.0304582231, 0.0490994826, 0.0000017853, -0.0130144469, 0.0162866265, 0.0043925988, 0.0064802854, -0.0146151194, 0.0847164541, 0.0108249439, -0.0128203817, 0.03103365, -0.0184742697, 0.0131998919, 0.0223175455, 0.0357416458, 0.1126413122, 0.0594555922, 0.0006896472, -0.007179602, 0.0079355277, -0.0104558188, 0.0330123082, -0.0156122008, -0.0112420246, -0.0360558107, -0.0063788034, -0.0039477278, -0.0033986981, 0.0255345497, -0.0385139324, -0.0234138947, 0.0421034284, 0.0276977029, -0.0190422218, -0.0345098861, -0.0230303314, 0.0436813608, 0.0249390025, -0.0260380656, -0.0074810651, 0.0330764502, 0.0216794424, -0.0299005099, -0.0172616057, -0.0251517873, 0.0233477633, -0.0159411076, -0.0815043375, 0.0187241081, 0.0004754295, 0.0286473706, 0.0505968928, -0.0119411191, -0.0391471945, -0.0465850309, 0.0413843915, 0.0031045598, 0.0256782398, -0.0475501791, 0.0272038262, 0.0216565635, 0.0270122793, -0.0568902045, -0.0343775302, -0.0535661466, 0.0282311514, -0.0135011841, 0.0506004356, -0.0129817082, -0.0832481906, 0.0101512978, -0.0044944584, 0.032646358, -0.001550398, -0.031516619, -0.0019697859, -0.0100634471, -0.038342908, 0.0133462511, -0.0146307684, 0.0157242492, -0.0053422912, -0.0297660232, -0.0133367619, -0.0151961958, -0.0612339266, 0.0200876258, 0.0012368249, -0.0004803751, -0.0168733075, 0.0209693797, 0.0284116529, 0.0358100757, -0.0314364769, 0.0312559046, -0.0374172032, -0.01351019, -0.0107491873, 0.0349663459, 0.0526159033, 0.0035064677, 0.0307147782, 0.0377818458, -0.0010047003, -0.0854148194, 0.0141344238, 0.0142129436, -0.0504761338, -0.0723424256, 0.0093768844, -0.0093856426, 0.0497713499, -0.0328722335, -0.0412810743, 0.0505698696, 0.0163538978, 0.0068106637, -0.0323478021, 0.0303633697, -0.0263924971, 0.0293016434, 0.0368035138, 0.0222274661, -0.0256675314, 0.0004289378, 0.0515821613, 0.0371742658, 0.0591336079, 0.0001539771, 0.0508286059, 0.0303586517, 0.0300932899, -0.0113409655, 0.0306729488, -0.0111775063, -0.0721366853, -0.0077794967, 0.0443341136, 0.0242826249, 0.0139177218, -0.0231749322, 0.0833118036, 0.0370218046, -0.0186140947, 0.0512506925, 0.0140866395, 0.0735871419, -0.0143025061, -0.0230634864, -0.0280210916, -0.0087529644, -0.0624841973, -0.0533588901, 0.0539978631, -0.0285279583, -0.0048915232, -0.0193991642, -0.0077849426, -0.0429431722, 0.0049561989, 0.0053734956, 0.0285561625, 0.0093300873, -0.0521655828, -0.0046583023, 0.0280009657, -0.080677703, 0.0068858028, -0.0239010137, 0.0431821942, 0.0459622256, -0.00097817, -0.0211009178, 0.0461181067, -0.0599498563, -0.0232750829, 0.0118163582, 0.0202841293, -0.0152033856, 0.0348104313, -0.0099131884, 0.0301936269, 0.0114336489, 0.007830522, 0.0088179829, -0.0105590979, 0.0508203171, -0.0264111962, -0.0106044225, 0.0084735733, 0.0036261133, 0.0728844702, -0.0098014977, 0.028049048, 0.023792563, 0.030977264, 0.0651343688, -0.0627781004, -0.0409167707, 0.0099525824, -0.0003088051, -0.0441909954, 0.0167911481, -0.0317334086, -0.0110584106, -0.0077458923, 0.0031359764, -0.0101635503, 0.0484530404, -0.0029989758, 0.0029021963, 0.0674158707, -0.0434384421, 0.0624521151, -0.0139819551, 0.0113312863, -0.0404214934, 0.0349740274, 0.0279468615, -0.0167390853, 0.0107807359, -0.0445234254, -0.0294297058, 0.0064817378, 0.0267776549, 0.0265276562, -0.0710620508, -0.017299911, 0.0117504671, 0.0155165512, -0.003994449, -0.0026008615, 0.0070589618, -0.0431971401, -0.1343955249, -0.0226410329, 0.0002777495, -0.0357557498, -0.0124383653, -0.0217290279, 0.0324077867, 0.0127154412, -0.0510341078, 0.0140414322, 0.0138455732, -0.0293697566, -0.0243292935, -0.0196608119, 0.0071506505, -0.0450208858, 0.0197476149, -0.0268044323, -0.000633129, 0.0036951932, -0.0016513894, -0.0034641011, -0.018639138, -0.0279017892, 0.0702793375, -0.0148783242, -0.0362485088, -0.0118849864, 0.0151184695, -0.0000210277, 0.0546087772, 0.0646405295, 0.0315547697, 0.0404205769, 0.0079636443, -0.0103114164, 0.0391867198, 0.0183795206, 0.0225414615, -0.0520406887, -0.0227578878, -0.0134883141, -0.0011945716, 0.0413786098, 0.0331585668, -0.0153995659, -0.013580801, 0.0162312984, -0.0262625962, 0.0089197578, -0.0284600202, 0.0022368559, 0.0031016907, 0.0264512207, -0.0521406941, 0.0240387488, -0.0631123558, -0.0766334832, 0.00036552, -0.0181443188, 0.0087706754, 0.0560607724, -0.0210449472, 0.0285352357, 0.0207880381, -0.0026008876, 0.0139692845, -0.0549430028, 0.0222909208, -0.0218939688, -0.0378445424, 0.0181715023, -0.0101329368, 0.0414085388, -0.0489812493, -0.0228468925, -0.0281004366, -0.0267392471, -0.024559252, -0.0344981253, -0.0330502912, -0.0265237149, -0.0028851421, -0.0075093051, -0.0130109498, -0.0196781587, 0.0090910429, -0.0095329005, 0.0260313973, 0.0504553653, 0.0252214782, -0.0199561529, -0.002539665, -0.0414280109, 0.0141014, 0.0019955651, 0.0510431454, -0.0174896829, -0.0377130844, 0.0256081298, 0.0358028151, 0.0248396453, -0.0312447455, 0.022389723, 0.0309144184, -0.0201262645, 0.0946215466, -0.0003396645, -0.0545745455, 0.016139213, 0.0496268496, 0.0051375846, 0.0203707945, 0.0038989463, -0.0733344853, -0.0273118354, 0.025100207, 0.0150597906, 0.0040237219, 0.0255877934, -0.0083873579, -0.0919764787, 0.0269542132, -0.0087239947, -0.0493086763, -0.0262599848, 0.0077868933, -0.0427260511, 0.008681722, 0.0049834778, -0.0082709724, -0.0038374383, -0.0207932517, -0.0251436941, -0.0377874896, 0.0006735132, 0.0641744509, -0.0409568809, 0.0686524585, 0.0503225736, 0.086146988, -0.0253268424, -0.0247116443, -0.0323927291, 0.0514810458, 0.0100215618, -0.0040877154, 0.0751825422, -0.035935808, 0.0266139228, 0.0108483313, -0.0107395602, 0.0045996383, 0.0572722778, 0.0437745713, 0.0141564952, -0.0303728525, -0.0516616255, -0.0244450979, 0.0776313171, 0.0301801208, 0.0613423139, -0.0017104198, -0.0707283467, 0.0073778322, 0.0492903851, 0.0122846495, 0.0043424196, -0.035553176, -0.0040322919, 0.0250812061, -0.0441082828, -0.080209814, -0.0243743509, -0.0663032457, -0.0068049426, 0.0146442195, -0.0147930728, -0.0055665253, 0.0297584161, -0.0448296703, 0.012888073, 0.0336806886, 0.0547623187, -0.0273988321, -0.0599507838, 0.0274219885, -0.0300355107, -0.0235344581, -0.0076318504, -0.0436675102, -0.0149405301, -0.0300909411, 0.059478581, 0.0150454957, 0.0690073669, -0.0109650064, -0.0493041575, 0.0259763822, 0.0026851494, -0.0733643472, 0.0316022113, 0.0589102618, -0.0059375186, -0.0202231109, 0.0216871146, -0.0320682749, 0.0462089591, 0.0033939213, -0.0691846907, 0.0536765717, -0.035413146, -0.0384643227, -0.0541238971, 0.0322218537, 0.0128915478, 0.0084776971, -0.0119103082, 0.0015918408, 0.0044004344, 0.0124278897, -0.070625551, -0.0149715096, -0.0273618586, 0.0233898684, -0.0324265063, -0.091909878, -0.0534453765, 0.0105131883, -0.0476111919, 0.0518327504, 0.0366875306, 0.0039015829, 0.0146740563, 0.0164743699, -0.0253991019, 0.0038099606, 0.0463468209, -0.0155400857, 0.0075956346, 0.0531346537, 0.0154553289, -0.0233390126, -0.0038606727, 0.0712936819, -0.012283735, 0.0207228009, -0.0446740054, 0.0047602509, 0.0259284265, 0.0047485721, 0.0659581944, 0.0015789912, 0.0280855577, 0.0108974632, 0.066853784, -0.0719483569, -0.044440072, -0.0043937718, -0.0074635055, -0.0230133981, 0.0329101384, 0.0081882309, -0.061639227, 0.003049776, 0.0089491848, -0.0090862596, -0.0413513146, -0.0142597156, 0.0445525087, 0.0470647253, 0.0118004903, 0.0295429938, -0.0366595052, -0.0674947649, 0.046811007, -0.020945888, 0.0193225425, 0.0005799762, 0.0503994711, 0.043380633, -0.0006087341, 0.013610594, -0.0129716685, -0.0685857087, -0.0305191465, 0.0008856133, -0.0275032036, -5.057805421e-33, 0.0279756915, -0.0051729209, -0.0354736969, 0.0849834532, -0.04772247, 0.0074919001, -0.0230368003, 0.007716794, 0.0176292155, -0.0010910435, -0.0162718669, 0.0320150778, -0.0120792035, -0.0084959334, -0.0037540528, 0.0242587086, 0.0550197177, -0.0321196578, -0.0311388224, 0.0513817407, 0.0480896421, 0.0335989967, 0.0622879788, -0.1261321753, -0.0136731239, 0.0030221415, -0.0507756621, -0.0212196037, 0.0187482797, 0.0314444341, -0.0311426632, 0.0740615353, -0.0162898265, 0.0432824716, 0.0330359563, 0.0557153523, -0.042366717, 0.0037964087, -0.015138397, 0.013405838, -0.0006654261, 0.0048354464, 0.0272497386, -0.0164497606, -0.0582583435, 0.018849507, 0.0621978976, 0.0118137896, 0.0620768443, -0.0375140682, -0.0819505379, -0.0261358451, -0.0142875612, -0.0660875142, -0.0052731507, 0.0433642343, 0.03388061, 0.0070885904, -0.0674664304, 0.055630967, -0.0198929571, -0.0073787305, -0.0241105761, -0.0347632915, 0.0229548812, 0.0887935758, 0.0704820678, -0.005744447, 0.0433750413, 0.0067534237, 0.0455398448, 0.0473174155, 0.0496052168, 0.0191991683, 0.0068619563, -0.0624972507, 0.0025004861, 0.0022005227, 0.0637353733, -0.0302368701, 0.0139228031, -0.0204752125, -0.059239123, -0.0103549231, 0.0009974982, -0.056024272, -0.0336978212, -0.0234748237, 0.0240188316, 0.017060142, 0.0197769813, -0.0005097008, 0.0252494514, -0.0603965372, -0.0328421891, -0.0015617716, -0.0197334681, 0.0017071912, 0.0050742552, -0.0485010408, -0.0153694404, -0.0330327451, 0.0379700363, 0.0161169227, 0.0407453217, -0.0093125748, 0.0456443764, 0.0644740835, -0.0518803485, 0.0043748962, 0.0069971075, 0.0310712997, -0.014837346, 0.0282131061, -0.0000825335, 0.026022654, -0.0259409454, 0.032154046, -0.0280225761, 0.0444862582, -0.0828767493, 0.00855383, -0.0020717534, -0.0117124366, -0.0705750957, -0.0372374617, -0.0066950819, -0.0676874816, 0.0129105337, -0.0450550206, 0.0118874237, -0.0606046915, 2.536e-7, -0.0543984622, 0.0575659163, 0.011030891, -0.0155103821, 0.0474034511, 0.0671140328, -0.0260569695, 0.0265400801, 0.001320114, 0.0735047013, 0.0179869756, -0.0501055308, 0.0522938669, 0.0424769036, -0.0795988441, -0.0288759824, -0.1009272635, -0.0208911709, -0.02262724, -0.0525134914, 0.044212278, 0.0677066371, 0.0488703065, -0.0474769548, 0.0059292763, -0.0111896116, 0.0146125797, -0.0413869806, 0.0317374617, -0.0466466919, -0.0291560069, -0.0457462668, -0.0277522914, 0.0331960469, -0.0183044709, 0.0186759774, 0.0003710386, 0.0748940781, -0.016224172, -0.0581858121, -0.0155447004, 0.0709402636, -0.0225386079, -0.0047587277, 0.0087218815, -0.0133437756, -0.0232633818, -0.0376148596, -0.0222902186, -0.0159911048, -0.0069234753, 0.0263922028, -0.007811653, 0.0411743559, -0.0024111476, 0.0078148795, 0.0003171864, -0.0817073733, 0.057299789, -0.0110549619, -0.0350479111, -0.0455179736, 0.0147855496, 0.0042042024, -0.0213296544, -0.0295781288, 0.029769171, 1.824293879e-34, 0.0071105589, -0.050416518, 0.0388724729, -0.0057345033, 0.0362524278, 0.0257872976, 0.0575264506, -0.047629904, 0.0224139132, -0.017222669, -0.0209356416 ]
http://arxiv.org/abs/2406.17663v1
http://arxiv.org/pdf/2406.17663v1
LLM-ARC: Enhancing LLMs with an Automated Reasoning Critic
2024-06-25T15:52:15
2024-06-25T15:52:15
Aditya Kalyanpur, Kailash Saravanakumar, Victor Barres, Jennifer Chu-Carroll, David Melville, David Ferrucci
cs.CL
cs.CL, cs.AI, cs.LO
We introduce LLM-ARC, a neuro-symbolic framework designed to enhance the logical reasoning capabilities of Large Language Models (LLMs), by combining them with an Automated Reasoning Critic (ARC). LLM-ARC employs an Actor-Critic method where the LLM Actor generates declarative logic programs along with tests for semantic correctness, while the Automated Reasoning Critic evaluates the code, runs the tests and provides feedback on test failures for iterative refinement. Implemented using Answer Set Programming (ASP), LLM-ARC achieves a new state-of-the-art accuracy of 88.32% on the FOLIO benchmark which tests complex logical reasoning capabilities. Our experiments demonstrate significant improvements over LLM-only baselines, highlighting the importance of logic test generation and iterative self-refinement. We achieve our best result using a fully automated self-supervised training loop where the Actor is trained on end-to-end dialog traces with Critic feedback. We discuss potential enhancements and provide a detailed error analysis, showcasing the robustness and efficacy of LLM-ARC for complex natural language reasoning tasks.
LLM-ARC: Enhancing LLMs with an Automated Reasoning Critic: We introduce LLM-ARC, a neuro-symbolic framework designed to enhance the logical reasoning capabilities of Large Language Models (LLMs), by combining them with an Automated Reasoning Critic (ARC). LLM-ARC employs an Actor-Critic method where the LLM Actor generates declarative logic programs along with tests for semantic correctness, while the Automated Reasoning Critic evaluates the code, runs the tests and provides feedback on test failures for iterative refinement. Implemented using Answer Set Programming (ASP), LLM-ARC achieves a new state-of-the-art accuracy of 88.32% on the FOLIO benchmark which tests complex logical reasoning capabilities. Our experiments demonstrate significant improvements over LLM-only baselines, highlighting the importance of logic test generation and iterative self-refinement. We achieve our best result using a fully automated self-supervised training loop where the Actor is trained on end-to-end dialog traces with Critic feedback. We discuss potential enhancements and provide a detailed error analysis, showcasing the robustness and efficacy of LLM-ARC for complex natural language reasoning tasks.
[ 0.0237939749, 0.0593535192, -0.0096962415, 0.0518842787, -0.0394435786, 0.0131506752, 0.0250257161, -0.0232799407, 0.0368763953, -0.0266831964, -0.0021258907, 0.0128378738, -0.0612236746, 0.0480502583, 0.0215275884, 0.0144498385, 0.068071112, -0.0413533896, 0.0419452712, -0.0144788111, -0.0166395754, 0.0328988358, -0.0151121952, 0.0038510603, -0.0223175157, -0.0208200626, -0.0213685408, -0.0049623786, -0.0062569482, -0.0473691821, -0.0118929045, 0.0370745473, 0.0401354767, 0.0255272314, 0.0000021432, -0.0506645851, 0.0097161597, 0.002178648, -0.0580819435, -0.0016658765, -0.0080970367, 0.0403855592, 0.0054859542, 0.0261460412, -0.033160869, 0.0125130564, -0.0065818788, 0.0468524881, 0.0842827037, 0.0870032981, 0.0137134567, -0.0314922333, 0.0213992056, -0.0102162426, 0.0204511508, -0.032838013, 0.0170442108, 0.0356874578, 0.029591877, -0.0105529204, 0.0000948083, 0.0116414977, 0.0815070719, -0.0303230193, -0.0362405702, 0.0317925289, -0.0370994173, 0.0042620935, -0.0178368259, 0.0659455061, 0.0306752641, -0.0282540489, -0.0414783694, -0.0047213137, 0.0002832698, -0.0413908549, -0.0040159803, 0.0391887091, -0.02115505, -0.0015951007, 0.0410050377, -0.0285870228, 0.0337206982, 0.0305598509, -0.0475301445, 0.0908560455, -0.0093489373, -0.037414927, 0.0504582413, -0.003105147, -0.0224896371, -0.0428983942, 0.0236603208, 0.0716747791, -0.0649792999, 0.0298812445, -0.004985502, 0.0013391154, 0.0245146118, -0.0142309796, 0.0499116629, 0.0587275401, 0.0189498495, 0.0550039746, -0.0470745601, 0.0589636341, -0.0024208408, 0.0811350569, -0.024186207, 0.0090380739, -0.0418754183, -0.0273810271, -0.0183106922, 0.0430319905, -0.0377566479, 0.008447499, -0.0144645385, 0.0239561014, -0.0162730664, 0.0405590944, -0.0041611986, 0.0264071729, 0.02745777, -0.0237452611, -0.0822956413, 0.0226452518, -0.0715905875, 0.0211015474, 0.0234668814, -0.0274731535, -0.0217163861, -0.0025149037, -0.0275758542, -0.0280565973, 0.0535941124, 0.0882518664, -0.0140860369, -0.0073708291, -0.0382504426, -0.0292537827, 0.0077034878, -0.0806385651, 0.0026642403, -0.0660256669, 0.0488545187, -0.0141264712, -0.0480648428, -0.0418291986, 0.0160960741, 0.0201077964, 0.0414309539, 0.0766749457, -0.0091025336, -0.0059403093, 0.0247410443, 0.0115326634, -0.0186591595, 0.052448757, 0.0357292816, 0.0030933812, 0.0034612147, 0.0025182609, -0.0100064762, 0.0185818486, 0.032831341, -0.0035952376, -0.0097049326, 0.0530481786, -0.0566132031, 0.0451469831, -0.0022090613, 0.0481527746, -0.0388024449, -0.0632359684, -0.0044950955, 0.0357948579, 0.0133792358, 0.0104878992, 0.0167113338, -0.0207254104, 0.026290819, -0.0497584343, -0.0792068839, 0.0438891537, 0.0017519982, -0.0063311523, -0.0236567706, 0.0008907609, -0.0254599024, -0.0676762164, 0.01854554, 0.0286574867, 0.0104824901, 0.0332379974, 0.0288317092, 0.0217524488, -0.0466673896, 0.0519597679, 0.0039736582, -0.0419712365, -0.0246499851, -0.0620875992, 0.06122756, 0.0425279737, 0.0469227731, -0.0389455482, -0.0627786815, -0.0268454943, -0.014045571, 0.0406193994, 0.0178946331, -0.0126517452, 0.0168842506, -0.0261144489, -0.0083278017, -0.0459528156, 0.0042952485, 0.0517332107, 0.0352074727, -0.0151249887, -0.0012236802, -0.0079216268, -0.0036560479, 0.0378234722, -0.0294428878, -0.0432857126, 0.0076870262, -0.0075542131, 0.03257294, 0.0082039395, 0.0046956008, 0.0162559468, 0.0132202152, -0.0106302127, -0.0551228784, -0.0235957652, -0.0648325756, -0.0504130311, 0.0813225731, 0.0076140934, -0.0462160185, 0.0139166722, -0.0707495809, -0.033563517, 0.0311949812, -0.010992323, -0.0302959923, -0.0470340848, 0.004141211, 0.031150911, -0.0272231381, -0.037327826, 0.0668175891, 0.0033969902, -0.0082104355, -0.0035175707, 0.0810733736, 0.0310281478, -0.0466486961, 0.019158965, -0.0034198638, -0.0306779221, -0.0046091755, -0.0012039426, 0.0297305826, -0.055167418, -0.0306936968, -0.0442420468, 0.0121199097, 0.0031574308, 0.0129096471, -0.0053898538, -0.0303073358, -0.0039226501, 0.0019363811, -0.035958644, -0.0130500821, 0.0464958809, -0.0068419771, 0.0242166594, -0.0065603298, -0.0194519963, -0.015231967, -0.0111552998, -0.041818127, 0.0060737166, 0.0267961826, 0.0493569225, -0.0228235591, -0.0228469186, -0.0559842885, 0.0300865881, 0.0063537573, -0.0268365499, 0.0526332073, -0.011237314, -0.0374537855, 0.0831261277, -0.0145309493, -0.0084603457, 0.0350903198, -0.0018959327, -0.0559090413, 0.0923991427, 0.0064264839, -0.0006622211, -0.0311369766, -0.0068654465, -0.026364699, 0.0523491763, -0.0051736399, 0.0204835385, -0.0190842822, 0.0370793231, 0.0195913315, -0.0343291573, -0.0176066775, -0.0256392956, -0.0078488495, 0.0438669883, -0.0330076627, -0.0087576052, 0.0952281207, -0.0024477828, -0.0107193431, -0.0475758128, -0.0079277353, -0.0637593344, 0.0854667202, -0.0046066195, -0.0171984248, -0.0120299533, -0.0092058862, -0.0184892919, -0.0622415878, 0.0120465886, -0.0406645611, 0.0706611052, -0.0355814919, 0.0047189468, 0.0131401485, -0.0475621335, -0.0459754206, -0.0169165861, -0.0666894838, -0.0085829245, -0.0395723246, -0.0102389241, 0.0212061778, -0.0267398674, -0.0165988281, 0.0232240818, -0.0157676563, -0.012687251, -0.0233613215, -0.0003582486, 0.008923348, 0.0407686383, 0.0228288099, -0.0074265446, 0.0532927625, 0.0481795073, 0.0122882817, -0.0137596764, -0.0697419569, 0.0439214297, -0.0262870509, -0.0287559722, 0.0355696306, 0.0045410395, 0.0472467691, -0.0973452404, -0.0061011394, 0.1136375368, 0.0377192535, 0.0135675911, 0.0055253757, -0.0435847305, 0.0421971269, 0.0423379131, 0.0650299489, -0.0905273706, -0.0480147526, -0.0025607932, -0.0018640985, 0.0034508996, -0.0487141982, -0.0122984899, -0.1097633168, 0.0313720815, 0.0068548932, -0.0222157035, 0.0100730127, -0.0376228094, 0.0051450687, -0.0197391901, 0.0007839164, 0.0524009131, 0.00480839, 0.0041125165, -0.0080721118, 0.0072760372, 0.042338077, 0.0842638463, -0.0318475887, 0.0512278266, 0.0368074737, 0.0025139714, 0.0025389178, -0.010278861, 0.0548583716, -0.0168812387, 0.0278513022, -0.054037746, 0.0359101333, 0.0073270625, -0.0068012956, -0.0240107775, 0.0262724664, -0.0337998942, -0.0307299793, 0.0337145962, 0.0698427036, -0.0163573381, -0.0126064252, -0.0264205411, 0.0173249952, -0.0273074098, 0.0410809629, 0.0137925195, 0.002571038, -0.0830768868, -0.0562538654, -0.0437442586, -0.0754078031, -0.0098836375, -0.000053835, 0.0195078366, 0.0265782848, -0.023006862, -0.036769364, 0.0287079643, 0.0199138708, -0.0221185498, -0.0336244293, -0.0114704808, -0.0116048176, -0.0273952447, 0.0205017645, -0.0472857207, 0.0457580164, 0.0012117584, 0.0379645601, -0.0102846865, -0.0278547816, -0.0215032827, 0.0292990915, -0.0547096282, -0.0494134873, -0.0296016745, 0.0557507165, 0.0211706404, 0.0450751297, -0.0230225399, -0.0106030693, 0.0235920493, -0.0026489329, -0.0347884931, -0.0105577493, 0.0463286713, 0.0196265839, -0.0058962656, 0.0203261524, -0.0273462497, 0.0286530666, 0.0075630248, 0.0695623234, -0.0163197294, -0.0242572594, 0.0462111235, 0.0367301032, -0.0299495291, 0.0309545714, 0.0263048001, -0.0245539974, 0.0295297839, -0.0330636874, -0.0107563417, -0.0103584463, 0.017245559, -0.0044610999, -0.0621830598, 0.0052136304, -0.0544038489, -0.0073370002, 0.0261353515, 0.0232953243, 0.0138149401, 0.0357824564, 0.0174124818, -0.0146911778, -0.0214363076, -0.0287551153, 0.016455492, 0.006578628, 0.053654667, -0.0106117344, 0.0632527843, 0.0489104316, 0.0062457495, -0.038779337, -0.0291983131, -0.052952759, -0.0453520715, 0.02760203, 0.0071972106, 0.0273446962, 0.0206393581, -0.0178531352, -0.0449142344, 0.0016903993, 0.0204240326, 0.0029880526, 0.0030964836, -0.0272097643, 0.0663806871, 0.0327802449, 0.0196204521, -0.0487795994, 0.0192182511, 0.0077149854, 0.0471569896, 0.0184427202, 0.025886571, -0.0641208813, -0.0615622066, -0.0514093116, -0.0250304714, -0.0103121689, -0.0058727548, -0.0235825274, -0.0958912596, -0.0032074847, 0.0543603636, -0.0043544215, 0.0041949097, 0.0056874887, 0.0423052646, 0.0323596969, -0.0574882589, 0.0104884626, 0.0136905778, -0.0319356993, -0.0177082475, -0.0356963724, -5.909314427e-33, -0.0240136031, -0.0342744403, 0.0861508027, -0.0150460871, -0.0244649947, -0.0305774435, -0.0599071234, -0.010870209, -0.0188932214, 0.035081014, -0.0229347255, 0.0375137143, 0.0005057336, 0.015613718, -0.0000222276, -0.0413553603, 0.0036016707, 0.0071503539, -0.0194866098, 0.0079266923, -0.0266578607, -0.0129058026, 0.0054387189, -0.0512193367, -0.026330648, 0.0208770409, -0.0204655603, 0.0554875545, -0.1006832719, 0.0479050502, 0.0288069975, 0.0206330568, 0.0164778531, -0.0239498466, 0.0066978293, -0.0050227349, -0.0884276554, -0.0695569366, 0.0401047766, -0.0299134273, -0.0721425489, -0.0221132934, 0.0292644706, 0.0046656807, -0.0026102082, -0.0272053406, -0.0135908611, -0.0188209787, 0.0159552526, 0.0085051488, -0.0144961542, 0.0161234438, -0.0292821862, -0.0150842546, -0.0278705712, 0.032490667, -0.0035153541, -0.0239087027, -0.0379816294, -0.0066713579, 0.0196481552, -0.0190256648, 0.052115716, 0.026699435, -0.0148628755, 0.0308359489, 0.0261713024, 0.0028587379, -0.0744841248, 0.0092272907, 0.0041521466, 0.069046855, 0.0575262792, 0.0227978304, -0.0064648436, 0.0220603757, -0.0694405213, 0.0019599916, 0.0237372909, 0.0919936523, 0.0127491634, 0.0054682693, -0.0222672913, -0.001056074, -0.067285791, 0.039334286, -0.0365063809, -0.0403572842, 0.0331210904, 0.0028056568, -0.0135878157, -0.0295301341, -0.0182482321, -0.0244377479, 0.0109220613, 0.0390374959, 0.0002700079, -0.0063593308, -0.0030618128, 0.0425726064, -0.0433116294, 0.0146248452, 0.0058301473, -0.0065444335, 0.0078103999, -0.0021122592, 0.0180758387, 0.0257047787, -0.0487170331, -0.0086517902, -0.0407811254, 0.029949788, -0.0033399304, 0.0210187677, 0.0108955614, 0.0355296321, -0.0197857544, 0.0371944532, -0.0215191171, -0.0062053395, -0.0367096625, 0.0302395001, 0.0252264068, 0.0448310301, -0.0216464438, 0.0053431182, -0.045799125, 0.0122158192, -0.0214745477, -0.0932767168, 0.0066491044, 0.0287544262, 2.752e-7, 0.039696794, 0.0457961112, 0.0459674448, 0.0931871831, 0.0295904167, 0.0328971259, -0.0137670171, -0.0250344966, 0.0295181256, -0.0530381054, 0.0037350589, 0.0896766707, 0.0118344668, 0.0429147966, -0.0167473592, 0.026270736, -0.0486807674, 0.0052905018, -0.054623697, 0.0226029176, 0.0846205577, 0.0790662691, 0.0306748636, -0.0468197241, 0.0014565928, -0.057682734, 0.0422333553, 0.0086366544, 0.0062460112, 0.0196862165, -0.0016033418, -0.0022636147, 0.0301222708, 0.0620620325, 0.0126541005, 0.0074059274, -0.0572095066, 0.0569161475, -0.0082016289, -0.0401793495, -0.0261757616, -0.0352849215, 0.0195668228, -0.0508653335, 0.0772218108, -0.0446220301, -0.0346604176, -0.0111268368, -0.0145793203, -0.0362062789, 0.0075221993, -0.0037407016, 0.0033737768, 0.0065793502, 0.0088864025, -0.0143983029, 0.0001979961, -0.0720245764, -0.0506649651, -0.0014696107, -0.0230796393, 0.0155602964, -0.0426952429, 0.0250486266, 0.0275782291, -0.0655626506, -0.0088137211, 2.590949802e-34, 0.0354650207, -0.0279021692, 0.0215025451, 0.0422590822, 0.0773367509, 0.0145078534, -0.0234273393, 0.0090012094, 0.0298096053, -0.0574819185, -0.0272742677 ]
http://arxiv.org/abs/2406.17650v1
http://arxiv.org/pdf/2406.17650v1
ELIZA Reinterpreted: The world's first chatbot was not intended as a chatbot at all
2024-06-25T15:41:40
2024-06-25T15:41:40
Jeff Shrager
cs.AI
cs.AI, cs.CL, cs.CY, cs.HC
ELIZA, often considered the world's first chatbot, was written by Joseph Weizenbaum in the early 1960s. Weizenbaum did not intend to invent the chatbot, but rather to build a platform for research into human-machine conversation and the important cognitive processes of interpretation and misinterpretation. His purpose was obscured by ELIZA's fame, resulting in large part from the fortuitous timing of it's creation, and it's escape into the wild. In this paper I provide a rich historical context for ELIZA's creation, demonstrating that ELIZA arose from the intersection of some of the central threads in the technical history of AI. I also briefly discuss how ELIZA escaped into the world, and how its accidental escape, along with several coincidental turns of the programming language screws, led both to the misapprehension that ELIZA was intended as a chatbot, and to the loss of the original ELIZA to history for over 50 years.
ELIZA Reinterpreted: The world's first chatbot was not intended as a chatbot at all: ELIZA, often considered the world's first chatbot, was written by Joseph Weizenbaum in the early 1960s. Weizenbaum did not intend to invent the chatbot, but rather to build a platform for research into human-machine conversation and the important cognitive processes of interpretation and misinterpretation. His purpose was obscured by ELIZA's fame, resulting in large part from the fortuitous timing of it's creation, and it's escape into the wild. In this paper I provide a rich historical context for ELIZA's creation, demonstrating that ELIZA arose from the intersection of some of the central threads in the technical history of AI. I also briefly discuss how ELIZA escaped into the world, and how its accidental escape, along with several coincidental turns of the programming language screws, led both to the misapprehension that ELIZA was intended as a chatbot, and to the loss of the original ELIZA to history for over 50 years.
[ 0.0888182893, -0.0070734401, -0.0291803163, -0.0199835598, -0.0197573379, -0.0003069343, 0.0046123802, 0.0191156752, 0.0572040826, -0.0238992274, 0.0013818785, 0.0083404323, 0.0065135355, 0.0213212743, -0.0206557363, -0.0360484533, -0.0072699795, 0.0196642857, -0.0036677395, 0.0302545596, 0.0118348403, 0.0305656809, -0.0300253648, 0.0825071782, -0.0500980318, -0.0405422151, -0.047125861, 0.0079449583, -0.0197997335, 0.027595466, -0.0109831588, -0.016199477, 0.0004314744, 0.0135388933, 0.0000018336, -0.0337222889, 0.0111815725, -0.001469353, -0.0762031749, -0.0667979345, 0.0124924434, 0.0320892185, -0.0158392936, 0.007724863, -0.0392715968, 0.0680469796, 0.043534331, -0.0001346433, 0.0348382182, 0.0636758432, -0.0145490449, 0.0547722876, 0.020911172, 0.0265628267, 0.0593302809, -0.0771634579, -0.0111224279, 0.0210460052, 0.0502226092, -0.014634734, -0.0140171032, -0.0045289081, 0.0304576978, -0.0217310321, -0.0225095656, 0.0194235947, -0.0010518258, -0.07828179, -0.0218767188, 0.0125892665, 0.0614685379, -0.0041532228, 0.0322114304, 0.0555459931, 0.0209339913, 0.0010396198, 0.0074724816, 0.0064857458, -0.0069756797, -0.0339758284, 0.0063109202, -0.0513132289, -0.0017989354, 0.0215848833, -0.0222924575, 0.0264604148, -0.019571783, 0.025175089, -0.0154929142, -0.0191945322, 0.0293122213, -0.0561218895, 0.0168169215, 0.0582649708, -0.0151919024, -0.022180995, -0.0193631817, 0.0191362668, -0.0373174511, -0.0521319471, -0.0445962586, 0.0476782545, 0.0191543009, 0.0459530577, 0.054703299, 0.0768170655, -0.0094398716, 0.0277754944, -0.0263926182, -0.0097664176, -0.0649436861, -0.0196878389, -0.0481027737, 0.0459012166, 0.0419589207, -0.0210553594, -0.0056198756, 0.0580989346, 0.0166080184, -0.0157383252, -0.0674842224, -0.0111051779, -0.0118407132, -0.0013989261, -0.0289957337, -0.0310568381, -0.0433499515, 0.0079125138, 0.0253800377, -0.0837153047, -0.0285388865, -0.0059397086, -0.032281395, -0.014435878, -0.0116994372, 0.0802351609, 0.0122120138, -0.0560678802, 0.0150965853, -0.0096926028, -0.014692165, -0.0736380965, 0.0067330473, -0.0152590005, -0.01046173, 0.0292367879, 0.0206739269, -0.048315078, 0.0090965992, -0.0160060246, -0.0230116434, 0.0002213668, -0.0289678778, 0.004083727, 0.0601454005, 0.0263109244, -0.0556666814, 0.0284923054, 0.0326062068, 0.0256868117, 0.0133304661, 0.024177894, 0.0838552564, 0.013683999, -0.0307561066, -0.0514675565, 0.0307231769, 0.0006229611, -0.0277959444, -0.001301376, -0.0186752845, 0.0019338136, -0.0276062004, 0.0495568924, 0.0064560049, 0.0768327862, -0.0465292372, 0.0428945497, 0.0430448949, -0.004848728, 0.0342738144, 0.0123435762, -0.0307062231, 0.0041832938, -0.0423229448, -0.0507313795, -0.0239055865, -0.0296631251, -0.0068135937, -0.0423677862, 0.0421604328, -0.0243876111, 0.0221311189, -0.0088051744, 0.0274969172, 0.0123603269, 0.0055661211, -0.0147072626, -0.0249343663, -0.0350558087, -0.0510504059, -0.0833599791, -0.0297483671, 0.1099418998, 0.0450396724, -0.0058230455, 0.0147668328, 0.0059462823, -0.0049504964, 0.036217615, 0.0325581431, 0.016593596, -0.0078023816, 0.0105975047, 0.0060168337, -0.0118704392, -0.009569034, 0.0130548645, -0.0587452352, 0.0232340582, -0.0166883375, 0.0699493736, 0.0073062661, 0.0058254581, -0.0100487713, -0.0290496256, -0.0680754334, 0.0054184152, -0.0029420743, 0.043342609, -0.017938504, 0.0228866152, 0.0260643791, 0.0065196496, -0.0005810889, 0.0369556323, 0.0085309967, 0.0246324651, 0.0249395166, -0.0954691991, -0.0148810539, 0.0801263899, 0.0167981293, -0.0055693975, 0.0410287119, -0.074442625, 0.0610819794, -0.0096499864, 0.0436020605, -0.0428607911, 0.0003537349, -0.0225765109, 0.0260002762, -0.0043209046, -0.0455341376, 0.025695052, -0.0011854111, -0.0142165385, -0.0210967492, -0.0322669595, 0.0236553438, 0.0188525263, 0.0168000422, 0.0262101237, -0.0322925523, -0.0488788001, 0.0012873275, -0.0851641074, -0.0253168605, 0.0163165443, -0.0288797691, 0.0159116071, 0.0087815979, 0.0009501367, -0.0504590869, 0.0058675124, 0.0562807843, -0.0366148204, 0.0083290692, 0.0577369332, 0.0093946261, -0.0546824709, -0.0455018729, 0.0212545376, -0.0298493803, 0.008933858, 0.0586789213, -0.0441381708, -0.0371799693, 0.0403164215, -0.0158737916, 0.0174310282, -0.026775999, -0.0134306001, 0.0415732674, -0.0141520659, 0.003504354, 0.0390091985, -0.0002316865, 0.1051971391, 0.0309669096, 0.0316287838, -0.0121744163, -0.0005012767, 0.0191378631, 0.0060110753, -0.0010545321, -0.0420465469, -0.0114798946, 0.0268670917, -0.0292973202, 0.0297329091, -0.0151932621, 0.0722380355, -0.0047219009, -0.0397282131, -0.026343694, -0.0331414491, -0.0234281011, 0.0647395402, -0.0099287303, -0.0319113806, 0.0601968355, 0.0103278887, -0.0102749681, -0.0045965677, -0.0275072381, 0.024778489, 0.0827722698, -0.0532323979, 0.0099871578, 0.0122773126, -0.0046779858, -0.0321461521, 0.0031425436, 0.0093965866, -0.0596259162, -0.0252602044, -0.0192612, 0.0166497398, 0.019277269, -0.0161613151, -0.0365900286, -0.0692882985, -0.0485093892, -0.0226295143, 0.0362408087, -0.0063420068, -0.0646349713, -0.0121686282, 0.006179614, 0.0420231558, 0.0455735885, -0.02043649, -0.0100553408, -0.0895815268, 0.0101169776, 0.0341245607, 0.0111070583, 0.0245522391, 0.056295678, 0.0282018371, 0.0143716745, -0.0100213606, 0.0211306605, -0.0441143326, 0.0010938308, -0.0053158016, -0.0484656394, 0.0118845636, 0.0236376803, -0.0470008813, -0.0214003064, 0.0634113476, 0.0214673113, -0.0590408035, -0.0021198015, -0.0237867441, -0.0465842783, 0.0209896564, 0.0535979234, -0.0853899866, 0.0183822811, 0.0277044326, -0.0097481199, -0.0339739546, -0.0325338319, -0.0107396236, -0.0779208913, -0.001300416, -0.0188204255, -0.0417315364, -0.0339031331, -0.0190242417, -0.0410566814, 0.0262766443, 0.0142769488, 0.0255091116, 0.007832882, 0.008941927, -0.0235491022, -0.0264800396, 0.0633795336, 0.0285960566, -0.0158070419, 0.0752830058, 0.0563876219, 0.0588972531, 0.0077751591, 0.0673372149, -0.0179978851, 0.0737306252, 0.0247323737, 0.0167568829, 0.045320075, 0.0017017593, -0.0255146492, -0.0278766826, -0.0217538122, 0.00983107, 0.0296039265, 0.0495076925, -0.0107184704, -0.0579611249, 0.0041834116, -0.0277182143, -0.0089244563, 0.028683573, 0.0069275443, 0.026287606, -0.0347120278, -0.0741492957, -0.0402064808, 0.0295445789, 0.0264917426, -0.0934232101, 0.0120706279, 0.0030725929, -0.0734496787, -0.0921212211, -0.0417195484, 0.0042310492, 0.0433637537, 0.0682950988, -0.0280854478, 0.005267208, 0.0453993641, -0.0264346376, -0.0021844234, 0.010550906, -0.0276646186, -0.0155941881, -0.0048352196, -0.0043089897, -0.0672971085, 0.0045046187, 0.0157257859, 0.025367979, -0.0031435834, 0.0227126032, 0.0307031292, 0.0359033272, 0.0003026737, -0.0226186458, 0.006286961, -0.0059293984, -0.004194329, -0.0144908298, 0.0247869231, 0.0292322747, 0.0665143132, 0.0023269677, 0.030681856, -0.0726237223, -0.0242084898, -0.0273399577, -0.0193471555, -0.0083113583, -0.0254477803, 0.0492032543, 0.0430410467, -0.0008983595, 0.0129669011, 0.0383186303, 0.0468453765, 0.000673514, -0.0356721021, 0.000673759, -0.0168098658, -0.0112948176, -0.0352386199, -0.0129901571, 0.0633360445, 0.0122080734, -0.0332402736, 0.0129423141, 0.0065147323, -0.0012858255, -0.0100046759, 0.0949237272, -0.0098225186, -0.0002018104, 0.0019345448, 0.0577733181, 0.0022599117, 0.0521666482, 0.0354036801, 0.0535974316, -0.0038178507, -0.0163231865, -0.0129197566, 0.0279776677, -0.1191446111, -0.1191550195, 0.0593911894, 0.0717958584, 0.0329326726, 0.0350687653, 0.0441464745, -0.0134030608, 0.0106759639, -0.0163495354, 0.0211346708, 0.021035688, -0.007371265, -0.0656151399, 0.0087052733, -0.0001630935, -0.0032731371, 0.0054545733, -0.0397017077, 0.0770922601, 0.0162853245, -0.0162953921, -0.037832357, -0.0632820576, -0.0079873195, -0.0380125605, 0.0372319371, -0.0211093631, -0.0197241865, -0.0267197266, -0.0180619527, 0.0232693572, 0.011231577, -0.0290351566, 0.0517122708, 0.0230505187, -0.0027202519, -0.0330218635, -0.0511985831, 0.0103323329, -0.0247641299, 0.018558925, -0.0006754386, -5.455179449e-33, 0.0149231935, -0.0274733678, 0.0434405729, 0.045122046, -0.0308301672, -0.0243879538, -0.0481502302, 0.0442650244, -0.0675293058, 0.0028168755, -0.0546453521, 0.0537495762, 0.0094619589, -0.0568944588, 0.0118021592, 0.00453887, 0.0258968975, 0.00187471, -0.0287906155, 0.0196053796, 0.0321452543, 0.0064808442, 0.0006252484, -0.0300503448, 0.0673742294, -0.0185010489, -0.0295363665, -0.0095847659, 0.0044968859, 0.0108585805, 0.0068419906, -0.0017790321, 0.0112333829, 0.0274387468, 0.0025422412, 0.0366650261, 0.0008539427, -0.0356118605, -0.0157293286, 0.0183388088, -0.0326932222, -0.0446989574, 0.0182342939, -0.0115566356, -0.0309516285, 0.0289684217, 0.0307577159, -0.0100208009, 0.0305424742, 0.0647943765, -0.0571051911, -0.0377157256, -0.0721902698, -0.0868159905, 0.0296652559, 0.0248540845, -0.0133990878, -0.0400987379, -0.0091387928, 0.0638730228, -0.0504511558, -0.0412930958, 0.0215111393, -0.0062737153, 0.0330960043, 0.0702103078, 0.1075878143, 0.044404339, -0.0408697575, -0.0350401737, -0.0044243662, 0.0245405976, 0.0038839169, 0.0228432175, -0.0237847604, -0.0233208984, -0.0581538305, 0.0085149091, 0.0465201065, 0.0041203271, 0.0142196668, -0.0105650555, -0.1152542979, 0.0022211673, 0.0090769585, 0.0277820211, -0.021005867, -0.0365710966, -0.0074943956, -0.0219990313, -0.0118724192, -0.0449898615, -0.01413712, -0.0911441967, 0.0134153571, -0.024950834, -0.0225137658, -0.0034831029, 0.024301311, -0.0073990854, 0.0257990137, -0.0359591059, 0.0039168941, 0.0468691438, 0.0679186136, 0.0174420848, -0.0342595875, 0.0355861895, 0.0384376459, -0.0445190854, -0.0161121637, 0.0000349135, 0.016253924, -0.0622873567, 0.0012250367, 0.0073318495, -0.0194625109, 0.0535416566, 0.0369975418, 0.0286443233, -0.0383800678, 0.0320257619, -0.0105377166, -0.0169422794, -0.033927761, -0.0412728079, -0.0468156226, 0.0877517238, 0.0221436191, -0.0595715232, -0.0351612233, -0.0103525873, 2.623e-7, -0.0119256405, 0.0351881832, 0.0663904399, 0.0165696424, -0.0014298392, 0.0198225416, -0.0294401478, 0.0390429534, 0.0087059168, 0.0280188508, 0.008455947, 0.0238481555, 0.014190427, -0.021037221, -0.0597985722, -0.0467435606, -0.0246574022, -0.0337421745, -0.0253981482, 0.0316066369, 0.0635207668, 0.0285414569, 0.0190543178, -0.0211962871, 0.0224222448, 0.011430718, -0.0165588409, -0.0334101059, 0.0256469753, 0.00919847, -0.0671070442, 0.0494672246, -0.0142031899, 0.0162114091, 0.0272765439, -0.0178844556, 0.016873654, 0.0364914574, -0.0091378326, 0.0024745828, 0.0208063424, 0.0180063173, -0.0033485934, -0.0035070542, 0.0472288094, -0.0335435122, -0.0665096417, -0.0079505341, 0.0079211174, 0.014779333, 0.0397281647, 0.0081020193, -0.0176643934, 0.0114167901, -0.0279066414, 0.001386915, -0.0044358969, -0.0209321082, 0.0202112198, -0.0130852778, -0.0461454913, -0.1013464704, -0.0197092928, -0.0095706489, 0.0375376977, -0.0169218909, 0.0327012464, 1.253696294e-34, -0.0175665766, -0.0207649656, 0.0229218062, 0.0789197832, 0.0263353307, 0.0321657546, 0.0339728519, -0.0006001741, 0.0553039014, 0.0162634123, -0.0385346413 ]
http://arxiv.org/abs/2406.17647v1
http://arxiv.org/pdf/2406.17647v1
Variationist: Exploring Multifaceted Variation and Bias in Written Language Data
2024-06-25T15:41:07
2024-06-25T15:41:07
Alan Ramponi, Camilla Casula, Stefano Menini
cs.CL
cs.CL
Exploring and understanding language data is a fundamental stage in all areas dealing with human language. It allows NLP practitioners to uncover quality concerns and harmful biases in data before training, and helps linguists and social scientists to gain insight into language use and human behavior. Yet, there is currently a lack of a unified, customizable tool to seamlessly inspect and visualize language variation and bias across multiple variables, language units, and diverse metrics that go beyond descriptive statistics. In this paper, we introduce Variationist, a highly-modular, extensible, and task-agnostic tool that fills this gap. Variationist handles at once a potentially unlimited combination of variable types and semantics across diversity and association metrics with regards to the language unit of choice, and orchestrates the creation of up to five-dimensional interactive charts for over 30 variable type-semantics combinations. Through our case studies on computational dialectology, human label variation, and text generation, we show how Variationist enables researchers from different disciplines to effortlessly answer specific research questions or unveil undesired associations in language data. A Python library, code, documentation, and tutorials are made publicly available to the research community.
Variationist: Exploring Multifaceted Variation and Bias in Written Language Data: Exploring and understanding language data is a fundamental stage in all areas dealing with human language. It allows NLP practitioners to uncover quality concerns and harmful biases in data before training, and helps linguists and social scientists to gain insight into language use and human behavior. Yet, there is currently a lack of a unified, customizable tool to seamlessly inspect and visualize language variation and bias across multiple variables, language units, and diverse metrics that go beyond descriptive statistics. In this paper, we introduce Variationist, a highly-modular, extensible, and task-agnostic tool that fills this gap. Variationist handles at once a potentially unlimited combination of variable types and semantics across diversity and association metrics with regards to the language unit of choice, and orchestrates the creation of up to five-dimensional interactive charts for over 30 variable type-semantics combinations. Through our case studies on computational dialectology, human label variation, and text generation, we show how Variationist enables researchers from different disciplines to effortlessly answer specific research questions or unveil undesired associations in language data. A Python library, code, documentation, and tutorials are made publicly available to the research community.
[ 0.0099772671, 0.1126901284, -0.051938761, 0.0076595419, -0.0335743688, 0.0143539738, 0.0160572361, -0.0157431755, -0.006177533, -0.0265450738, 0.0033157058, -0.0197177865, 0.0396965407, 0.0560765639, 0.0286747701, -0.0642423704, 0.0418245196, 0.0294444468, 0.0290063638, 0.0229830295, 0.026499439, 0.0397438966, -0.0169566255, 0.0445907936, -0.0587831475, -0.0271158591, -0.0087941997, 0.009272676, 0.0413098037, -0.0254123528, 0.043821156, 0.0711932182, -0.0201794673, 0.0308029838, 0.0000021604, -0.0527968258, -0.0040788008, -0.0254290886, 0.0037992476, 0.0137398867, 0.1156984121, -0.0053891344, -0.0034531753, 0.0198233332, -0.030915983, -0.0029120946, 0.07809183, 0.0031633209, 0.0245995726, 0.0046649072, -0.0096471049, 0.0119525986, -0.066743508, -0.0058189598, 0.0037291618, 0.0559029877, -0.0001634337, -0.0437219515, 0.0062819342, -0.0021608633, -0.0174390413, 0.0049812212, -0.0093860794, -0.0064049512, -0.0123079419, 0.0014669133, -0.0390342586, -0.0502580926, 0.0149610722, 0.0315552019, 0.0405504555, -0.0350428484, 0.0124172904, -0.0307777729, 0.0157679673, 0.0271962415, -0.039121706, 0.0096365279, 0.0044807447, -0.0175318457, 0.0159860831, -0.0024304911, -0.0154968267, 0.0253232177, -0.0025814148, 0.022747241, 0.0347137824, 0.0090336725, 0.0159334503, 0.0346879549, -0.0120419106, -0.0629331246, 0.0089671379, -0.0129887164, -0.0700021535, 0.0039552012, -0.0205526352, -0.0908234119, 0.0529415719, -0.0090053817, 0.0527805649, 0.0022359607, -0.0379363261, -0.0331751741, 0.0678007007, 0.025766965, 0.0105719613, -0.0115342084, -0.003006069, -0.020131072, 0.0038116204, -0.0353927575, -0.0048333006, -0.0618591681, 0.0218615029, 0.050472863, -0.0360060968, -0.0083380789, -0.0676568151, -0.009948276, -0.0042764167, 0.0469008125, -0.0317598432, 0.021561211, -0.0378011055, 0.0378518663, -0.0899940208, 0.0295214858, -0.0127085503, -0.0437125079, 0.0010460981, 0.0431031026, 0.0028010109, -0.0449240729, -0.014857267, 0.0015298364, 0.0095351087, 0.0060631465, -0.007612709, -0.0341514312, -0.0737481043, -0.0190549791, -0.0330300927, -0.0963981599, 0.020451149, 0.0368605219, -0.000042781, -0.0484731793, -0.0180586874, 0.0266925097, -0.0348273441, 0.0219310336, -0.0204965752, -0.0074621085, 0.0285265632, -0.0171126574, 0.0030977717, 0.0552779362, 0.0271108113, -0.0117576728, 0.0417736694, -0.0382655971, -0.023958141, 0.02332622, -0.0041269027, 0.0069237435, -0.0713358819, -0.0381513201, -0.0254869442, 0.0145464335, 0.0390028134, 0.0527332798, -0.0089296885, -0.0364202857, 0.0514973029, 0.052628804, 0.0253313016, -0.0121203754, 0.0251254998, 0.0453856736, -0.0038548454, -0.0008163406, 0.0161027685, 0.0113632903, 0.0453315303, -0.0042995703, -0.0253138356, 0.0730856881, 0.0170302913, -0.0480890498, -0.0422183871, 0.0144705586, 0.0292054657, -0.0357250012, -0.0022764872, 0.0267667044, 0.037210457, 0.0510705635, 0.0592323653, -0.0661348179, 0.0031249039, 0.0130792372, 0.0217155889, 0.0497686379, 0.0056077372, -0.0199558455, 0.0302893464, -0.0254432075, 0.0188030954, 0.0826788396, 0.0605859235, -0.0098426417, -0.0072517223, -0.0449796729, 0.0013448931, -0.029701177, 0.028219454, 0.0355054997, 0.0057035387, 0.0285078082, -0.0083457623, -0.0818261057, -0.0243947506, 0.032293994, 0.0105008772, 0.015134817, 0.0335051343, -0.009575977, 0.0094998479, 0.0466239378, -0.0237388648, 0.0340303443, -0.0124045806, -0.0151191661, 0.0060178079, 0.0556180924, 0.0213365778, -0.0061358018, 0.0234271288, -0.0831537247, -0.0358041748, -0.0029453011, -0.0036691735, -0.0339149162, 0.0388031676, -0.0629651845, 0.0255148988, -0.0184109993, -0.0336393528, 0.0002166394, 0.0441139452, -0.0481088087, 0.0547108725, 0.0100433128, 0.01676297, 0.0176894329, 0.0123169096, 0.0418244451, 0.021188302, 0.0115378443, 0.0059290295, -0.0126764625, 0.0027988586, 0.0424813144, -0.0122896452, -0.0681779757, -0.0188824069, -0.1253567189, -0.0008616679, -0.0203504022, 0.0051729977, 0.0132250329, 0.0017387042, -0.0107897045, 0.0092143025, -0.0242141467, -0.0046266927, 0.0108660404, 0.0341410488, 0.0603421517, -0.0464342013, 0.0290929731, -0.0289675742, 0.0091657797, -0.0317377076, 0.0333168171, 0.0233497154, 0.0331865586, -0.0120272748, -0.0338958502, -0.0578644313, -0.0567121543, 0.0215900578, -0.0239840895, 0.0528462976, 0.0334640704, 0.1148019508, 0.023047151, 0.0447097719, 0.0577564649, 0.0633214563, -0.0319248959, -0.0385580584, -0.0251023788, -0.0275716074, 0.0394297577, 0.0363854878, -0.0681824312, 0.0057446095, 0.0896073729, 0.0633188188, -0.0115524307, -0.0461249873, 0.0544050187, -0.0372587927, -0.0163771696, -0.016052749, 0.0148646887, -0.0406564064, 0.0484002829, 0.0061819316, -0.0510834195, 0.0460806116, 0.0123338848, 0.0169820935, -0.0327182375, -0.0421728045, -0.0296843983, 0.0532877669, 0.0030749319, 0.0122912517, 0.001857915, -0.0378912836, -0.0567217954, 0.0262996405, 0.0398302115, -0.0157724172, 0.0351737998, -0.0853289366, -0.0194976665, 0.0057017221, -0.0125960959, -0.0257222503, -0.0305062253, -0.0139705865, -0.0347442031, 0.061849732, -0.0438920744, 0.014597659, -0.0496926159, 0.0305068158, 0.016677402, -0.0302560758, 0.0075361929, -0.0012809536, 0.006271448, 0.0465724021, 0.0360671915, 0.0100854998, -0.0192986913, -0.0287402347, 0.0436217301, -0.0505202711, 0.0548345558, -0.008844859, 0.0235327613, 0.0007190346, -0.018690411, -0.0141419219, 0.0248035695, 0.0577858984, -0.019043101, -0.0032970756, 0.0805718303, 0.0320821963, -0.0253733061, 0.0387313738, -0.0686283484, 0.1109674796, 0.0246534329, 0.0142563144, -0.104306832, 0.0029010777, -0.0297429804, 0.0203600023, -0.0374646224, -0.0435431786, -0.0225078557, -0.0829179212, 0.0328940265, 0.0112682218, -0.0527309999, 0.0424183682, -0.0432502888, -0.0004832455, -0.0187154301, 0.0156175513, -0.0264020488, 0.0170317292, 0.0290121753, 0.0385963842, -0.0174175594, -0.0509816892, -0.0307245553, -0.0311838295, -0.0016885503, -0.0091249272, 0.0528111868, -0.0360529572, -0.0167529285, 0.0068871793, -0.0032096102, -0.023871541, -0.0073064463, 0.0729421899, 0.0274074897, -0.0284480359, -0.0651622415, 0.0245587751, -0.0116173169, -0.036875546, -0.0042035556, 0.043530263, -0.0076878956, -0.007001224, -0.0480861254, -0.0036924488, 0.0088928891, 0.0577376038, 0.053000506, -0.0564286821, -0.0041930024, -0.0421687514, 0.0292036999, -0.0383852609, -0.0477768444, 0.0439976007, 0.0018198701, -0.0768123642, -0.0664748773, -0.0454331972, 0.0239130463, 0.0072701843, 0.0115451384, -0.0305059385, 0.0053958953, -0.0405241102, -0.0921955854, 0.0431968383, 0.0207328219, 0.0272661168, 0.0104258321, -0.0632235706, 0.0039342209, -0.0125879636, -0.0101234112, 0.0429240204, -0.0982782841, -0.068448253, 0.0560512617, -0.0078879828, 0.0152371041, 0.0285516959, 0.0160017498, -0.0535585135, 0.0018164462, 0.0519730859, 0.032349892, 0.0101838065, -0.0434647836, -0.008696218, 0.000776578, -0.0092179123, -0.009167823, 0.0012529814, -0.0268450603, 0.0133789517, -0.0385486297, -0.0144529808, 0.0144773554, -0.0408110321, 0.0033356063, -0.040460784, -0.0220735185, -0.064427115, -0.0651125088, -0.0360304117, -0.0520465635, -0.0078950552, -0.034831088, 0.0114685846, -0.0086097317, -0.040677622, 0.0224117301, -0.0368252844, 0.0455205925, -0.0070393044, 0.0435381904, -0.0039719576, 0.0591140091, 0.0154847093, 0.0078863623, -0.0053010392, 0.0442634337, -0.0370611325, 0.0842681676, 0.01594452, 0.0371500403, 0.0040663676, -0.0167303868, 0.0342434645, -0.0034226805, -0.0282298718, -0.0390065685, 0.0113462983, -0.0365843959, -0.0018933057, -0.0456389971, 0.0921281353, -0.0301720109, -0.0088699115, -0.0128177218, 0.034660019, 0.0120114246, -0.008727584, 0.01583956, -0.0025628866, -0.0433763824, -0.0282844491, 0.0637251735, -0.0019670837, 0.0136227291, -0.0351977386, -0.0315185487, -0.0181444138, -0.0147051737, 0.0296645481, 0.0322775356, 0.0095133595, -0.0073135104, -0.0967418626, -0.0700723976, 0.0059249601, -0.0147907818, 0.0294783972, -0.0007841883, 0.0172313638, 0.052012004, -0.0231697895, -0.0131980861, -0.0208023563, 0.024388114, 0.0030398478, -0.0333985686, -0.0361024998, -6.696482749e-33, -0.0304057356, -0.1019663215, -0.0279526021, -0.0750388056, -0.0033561059, -0.0144951418, -0.0241751485, 0.0482840538, -0.01129713, 0.0414833054, 0.0160102788, 0.0111723822, 0.0169498306, 0.0195808578, 0.0669104755, 0.0087084165, 0.0255604107, 0.0189480968, -0.011347048, -0.0163206682, 0.0492277332, -0.0162406527, 0.0750032738, -0.0726421699, 0.0660718605, 0.0152343269, -0.0041637602, -0.0128081692, 0.0011611474, 0.0142397685, -0.00005773, 0.0764534771, 0.0232406333, 0.0025651429, -0.0277060475, -0.0028217346, -0.004501319, -0.0953772441, 0.012444593, 0.0338345207, -0.0043410803, -0.0253680032, 0.0400316156, -0.0066419421, -0.0307019986, 0.0097831059, 0.0525618494, 0.0072803604, 0.0252144374, 0.0259736758, -0.0455825031, 0.0316604003, -0.045071885, 0.0385822169, -0.0278825276, 0.0078312429, -0.0102340989, 0.016554419, -0.038586814, 0.0503799841, 0.0308540519, -0.0032036533, 0.0511492491, 0.0133162057, -0.0081964312, 0.0454665571, 0.025690956, 0.0277833715, 0.0102525726, 0.0149671379, 0.0317824855, 0.017621126, 0.0232139863, 0.0948631465, 0.0180540085, -0.0349539071, 0.0003340993, -0.0014807107, 0.0467628166, -0.0191369411, -0.0339218937, -0.0082646711, -0.0395480134, -0.048616074, -0.0233048499, -0.0465040468, -0.0217778999, -0.0270943847, 0.0369358025, -0.0583141558, -0.0001069311, -0.0099451384, -0.0019729852, 0.0082152067, -0.039190907, 0.0281022452, -0.047548756, 0.017572213, 0.0178668573, 0.0060328045, -0.0273861047, 0.0087091904, 0.0068779252, 0.0238657519, 0.0543930493, 0.019860778, -0.0020830699, -0.0187847763, -0.0587567873, 0.0171909221, 0.0081066769, -0.021769898, -0.0247542467, 0.0088402797, 0.0330494978, 0.0089921076, -0.0187505409, 0.0342309736, 0.0566395819, 0.0085914303, -0.0437444448, -0.0472757965, -0.0313764252, 0.0165666696, -0.0100682089, -0.015665004, -0.0131292492, -0.0255528949, 0.0662664697, -0.0283234864, 0.0011727177, -0.0569250584, 2.8e-7, 0.0525891371, 0.082961008, -0.009590731, 0.0502250306, 0.0212496985, 0.020128632, 0.0280167498, 0.0128301932, -0.0210848469, -0.0242606793, -0.0178712346, -0.0071375538, 0.0593620054, 0.0789876878, -0.0232700408, -0.0080173537, -0.0384259894, -0.0046463436, -0.0790308565, 0.0201041847, 0.076182425, 0.0448438078, 0.0154010458, 0.0125550795, -0.0027681419, 0.0183667187, 0.0304865483, -0.0255171638, 0.0003364876, 0.0113560529, 0.0134364143, -0.0443874709, 0.0311809499, 0.0070975679, -0.0314763747, -0.0144502232, -0.0052188323, 0.0213218182, 0.0053301868, -0.0069697397, -0.0062230132, -0.0071690534, -0.0110056512, -0.0801481754, 0.0328545533, 0.0303200502, -0.0161372852, -0.0317195579, 0.0109423185, 0.0117676528, 0.0518501103, -0.0463701524, -0.0293667614, 0.0036835559, 0.0165659878, 0.0056857266, -0.032031104, -0.0155395158, -0.0624818541, 0.0247769654, -0.0234865844, -0.0329165086, -0.0089146569, 0.0386096947, 0.0190301854, -0.0173811298, 0.0241864268, 2.711625547e-34, 0.0339373723, 0.0119184246, 0.0352354981, -0.0149547765, 0.0609359555, -0.0097238058, -0.0065266103, 0.0321616605, 0.0317477062, -0.0189143661, -0.0261431132 ]
http://arxiv.org/abs/2406.17642v1
http://arxiv.org/pdf/2406.17642v1
Banishing LLM Hallucinations Requires Rethinking Generalization
2024-06-25T15:31:01
2024-06-25T15:31:01
Johnny Li, Saksham Consul, Eda Zhou, James Wong, Naila Farooqui, Yuxin Ye, Nithyashree Manohar, Zhuxiaona Wei, Tian Wu, Ben Echols, Sharon Zhou, Gregory Diamos
cs.CL
cs.CL, cs.AI
Despite their powerful chat, coding, and reasoning abilities, Large Language Models (LLMs) frequently hallucinate. Conventional wisdom suggests that hallucinations are a consequence of a balance between creativity and factuality, which can be mitigated, but not eliminated, by grounding the LLM in external knowledge sources. Through extensive systematic experiments, we show that these traditional approaches fail to explain why LLMs hallucinate in practice. Specifically, we show that LLMs augmented with a massive Mixture of Memory Experts (MoME) can easily memorize large datasets of random numbers. We corroborate these experimental findings with a theoretical construction showing that simple neural networks trained to predict the next token hallucinate when the training loss is above a threshold as it usually does in practice when training on internet scale data. We interpret our findings by comparing against traditional retrieval methods for mitigating hallucinations. We use our findings to design a first generation model for removing hallucinations -- Lamini-1 -- that stores facts in a massive mixture of millions of memory experts that are retrieved dynamically.
Banishing LLM Hallucinations Requires Rethinking Generalization: Despite their powerful chat, coding, and reasoning abilities, Large Language Models (LLMs) frequently hallucinate. Conventional wisdom suggests that hallucinations are a consequence of a balance between creativity and factuality, which can be mitigated, but not eliminated, by grounding the LLM in external knowledge sources. Through extensive systematic experiments, we show that these traditional approaches fail to explain why LLMs hallucinate in practice. Specifically, we show that LLMs augmented with a massive Mixture of Memory Experts (MoME) can easily memorize large datasets of random numbers. We corroborate these experimental findings with a theoretical construction showing that simple neural networks trained to predict the next token hallucinate when the training loss is above a threshold as it usually does in practice when training on internet scale data. We interpret our findings by comparing against traditional retrieval methods for mitigating hallucinations. We use our findings to design a first generation model for removing hallucinations -- Lamini-1 -- that stores facts in a massive mixture of millions of memory experts that are retrieved dynamically.
[ 0.0063720159, 0.0764822364, -0.0311836004, 0.0300194677, -0.0202681292, 0.0020635303, -0.0304014347, 0.0218647812, -0.0078918338, -0.0131092211, -0.0401335992, -0.06183419, 0.0037447521, -0.0091330139, 0.0227004662, -0.0131437564, 0.0405968241, -0.0522313304, 0.0236879922, -0.0233793445, -0.0123402607, -0.0069627105, 0.0060313395, 0.0364275463, -0.0097687133, -0.0123493215, 0.0017315334, -0.0157541614, -0.0191846583, 0.0000513005, 0.009628444, 0.0430173799, 0.0182079338, 0.0406882092, 0.000002103, -0.0322698615, 0.0146176266, -0.0357910916, -0.0368012451, -0.0029955369, -0.0082230531, 0.003756095, -0.0001373773, -0.004779147, -0.039848648, -0.0130322427, 0.0391225219, 0.0684159622, 0.0593387112, 0.0778365359, 0.0132788857, -0.0457053557, 0.0805977061, 0.0185497645, 0.0762622803, 0.0107165687, 0.0347536653, 0.0295711067, -0.0197993573, -0.0437987, 0.0000016911, 0.0297588073, 0.0211757254, 0.0329441354, 0.0206365921, 0.0433188863, -0.0935419947, -0.0299676452, -0.0172203686, 0.058963567, -0.0331532992, -0.0524008162, 0.0203876235, 0.0334028266, 0.0200800207, -0.0092948405, -0.0636300147, 0.0515563861, -0.0186292268, 0.0001997667, -0.0217601061, 0.0000244138, 0.081193924, -0.0018643251, -0.0470726788, 0.0401201509, 0.0014172086, -0.0404692255, 0.0179784205, -0.0040771202, -0.0070266002, -0.0304956548, 0.0159413181, 0.0219861679, -0.0263880678, 0.0138087319, -0.0207953416, 0.0037812022, 0.024873754, -0.0984134823, 0.0048951246, 0.0224183947, -0.024924539, 0.043923784, -0.042953603, 0.006229355, -0.0288142432, 0.0668304786, 0.0257171653, 0.0024870522, -0.0676332563, -0.02930074, -0.0156109473, 0.0499192588, 0.0479853526, -0.0274832174, -0.0142826457, 0.0121173179, -0.0348506942, 0.0308136847, -0.0441202484, 0.0385339297, 0.0335646905, -0.0307769291, -0.0276790187, 0.0061514457, -0.0070438315, -0.0104276147, 0.0397920571, -0.082787402, -0.0263351798, 0.0033096629, -0.0117001161, 0.0164009277, 0.0140153291, 0.0708847493, 0.0057671415, -0.0210114345, -0.0632883236, -0.0343425833, -0.0014233709, -0.0599677637, 0.0269972626, -0.0502266921, 0.0337185152, -0.0021686328, -0.010645452, -0.0483012907, 0.0183842685, 0.0244100261, -0.0599135607, 0.0604479164, -0.0023839022, -0.0038358558, 0.0541948415, 0.0448743552, -0.0219299886, 0.0535503216, 0.0177850835, 0.0087739415, 0.0194150899, -0.0072886846, 0.0568822026, 0.0184770525, 0.0294223893, -0.0214038156, -0.043563351, 0.0009369812, 0.032810349, 0.0218837466, 0.0159836765, 0.0679231659, -0.0390800238, -0.035520032, 0.0950049534, 0.0275764503, -0.0338072665, 0.0537208915, 0.013426492, -0.0137083242, 0.023584608, 0.0202439353, -0.0718837529, 0.0362590775, -0.0261556841, 0.0143840164, 0.0699488074, -0.0024498287, -0.0474876203, -0.0192935001, -0.0063082138, -0.0099161081, 0.0063422029, 0.0242447704, 0.0189812873, 0.0120723005, -0.0407415405, 0.1053100452, 0.0212524887, -0.0196178909, 0.000369932, -0.0570217967, 0.0107480055, 0.0658193976, 0.0509282425, -0.018833112, -0.0023683661, -0.0288277585, 0.02617562, 0.0260947347, 0.0313750952, 0.0128275556, 0.025757188, -0.0655530542, 0.0416023359, -0.0077195158, 0.0433982722, 0.00204722, 0.0005024176, -0.0038963943, -0.0038531953, -0.0189273506, -0.0314890482, 0.0191849992, -0.0388205722, -0.0337856412, -0.0208454281, -0.0211571511, 0.0143575026, -0.0110102957, -0.0183158014, 0.0394586921, 0.0206828751, -0.0407965109, -0.0592876486, -0.009795676, -0.0611070953, 0.0022647753, 0.0304797143, 0.0005525344, 0.016660545, -0.0194695704, 0.0044881576, -0.0279177465, 0.0465907827, -0.0643122792, 0.0846927688, -0.033313565, 0.0352570899, 0.0065799141, 0.0446545817, 0.0157230552, 0.0784094557, -0.0410397127, 0.0072103371, 0.0201493241, 0.0050397804, -0.0024757828, -0.0377877951, -0.0056523019, 0.0168094765, -0.0074791904, 0.0483580232, -0.0067873239, -0.067072019, -0.0582961701, -0.0280416682, -0.0764869004, 0.003463432, -0.0232588612, -0.0242041219, 0.0126357926, -0.0318095833, -0.0026355975, -0.0123471292, -0.0382237919, 0.0142648071, -0.0136865396, 0.0481746979, 0.0106150936, 0.009074891, -0.0300384872, -0.0214438513, -0.0064707082, -0.0034700662, 0.0597648025, -0.0335462838, -0.0245262608, -0.0474763289, 0.013290788, -0.0649649277, 0.0064917346, -0.0244076438, 0.0011065368, 0.0176026244, -0.043526046, 0.042550005, 0.0908655897, 0.0169977732, -0.0224251803, 0.0553336814, -0.0302093644, -0.0457296297, -0.0148540577, 0.0057611461, -0.0233442206, -0.0026216549, -0.0180960484, -0.0110800667, 0.0483584329, 0.027080873, 0.0315032117, -0.0489598922, 0.0473325811, -0.0016837863, -0.0437611304, -0.021504065, -0.0146159679, -0.0310031027, 0.0376654901, -0.0240842346, -0.0622854978, 0.0796478838, 0.0063930382, -0.0025035243, 0.0113782585, -0.0148665868, -0.0048137619, 0.0145742288, -0.0091415616, -0.0041901888, -0.0066719335, -0.0189910736, -0.0630320013, -0.036175523, -0.0345622115, -0.0572038256, 0.034796048, -0.0216428116, 0.002227006, 0.0256109256, -0.0444639288, -0.066969052, -0.0183998682, -0.0229914393, -0.0044102799, -0.0183086656, -0.0000899051, -0.0468350723, -0.0310406294, 0.0376213007, 0.012448376, -0.0125807561, -0.0337857604, 0.0222697929, 0.0086144907, 0.0402321257, -0.0082658567, 0.0398098379, -0.036272753, 0.0349646285, 0.0814492181, 0.0322442353, 0.0172573682, -0.0221349187, -0.0018532611, 0.0056625591, -0.0040295385, -0.0215879343, 0.0350817367, 0.0246479101, -0.0573704466, -0.0312258285, 0.0969337299, 0.0242004283, -0.0672437772, -0.0108114332, -0.0291473009, 0.0518224426, 0.0059051649, -0.0094970968, -0.0512273125, -0.0106437821, 0.0051464611, -0.0029623711, -0.0151170604, -0.0256250128, -0.0044997139, -0.0814094469, -0.010526292, 0.040999569, -0.0220601819, -0.0094041731, -0.0211490709, -0.0345061719, 0.01550656, 0.0485139191, 0.0129643539, 0.0096656149, -0.0255383048, -0.0034143261, -0.0432223603, -0.0078303507, 0.0503267795, -0.0211986788, 0.0775025412, 0.0645428449, 0.0196008701, -0.0052459901, -0.0775699541, 0.0306005143, 0.0009483126, 0.0064827083, 0.0166689083, 0.0681425035, 0.0160847753, 0.0019291709, 0.0085292561, 0.0287129506, -0.0480141416, -0.0213075932, 0.0048641125, 0.0612315685, -0.0330143906, -0.0581526384, -0.0045922403, 0.0350471288, -0.0028797609, 0.072588563, 0.0254148617, -0.0200612918, -0.0391421095, -0.0398751795, 0.0249524862, -0.0176912844, -0.0543431826, 0.0340532102, 0.0761759579, -0.0198341701, -0.065635331, -0.0294722319, -0.0040633078, 0.0183661021, 0.0070050899, 0.0123677356, -0.0128046395, 0.0069715097, -0.013177366, 0.0249349456, -0.046331685, -0.0280973651, -0.0338386074, 0.0240466651, 0.0191289894, -0.0289437305, 0.0243926495, 0.0153866122, -0.0928306356, -0.0797463879, 0.0265390221, 0.0582476035, 0.0328693651, 0.044128105, 0.008734365, -0.0657137185, 0.0696107447, 0.0492607318, -0.0541605912, 0.0106115686, 0.0221121814, 0.0247619171, 0.0255812332, 0.0397332907, -0.0259965155, 0.0063604163, 0.0199027676, 0.0015883613, -0.0234633833, 0.0413610153, 0.0401523039, -0.0154937468, -0.0069593317, 0.00499749, 0.0097954068, -0.0013335105, 0.0149066029, -0.0013632163, 0.0109292278, -0.006831063, -0.0013802118, 0.0370048136, -0.0131516391, -0.0020997152, -0.034880612, 0.013455783, 0.0045760218, 0.0174900498, -0.0509447195, 0.0084113879, -0.0021252823, 0.0538268276, -0.0177691355, -0.0572176017, 0.0377033539, -0.0182502978, 0.0511984192, -0.0098920316, 0.0594879724, -0.0071079405, -0.0163847134, -0.0255243126, 0.007843107, -0.0681001693, -0.128442511, 0.0250555426, 0.0423422493, 0.056278266, 0.0030130588, 0.0042153238, -0.0341399014, 0.0203901418, -0.0304699577, -0.011709773, 0.0242372025, -0.0188307669, -0.0259039402, 0.0264550634, -0.0500280149, -0.006679575, 0.0613240525, -0.0323241539, 0.034007214, -0.0040211612, -0.0068802461, -0.0267484747, -0.0376031063, -0.0123125073, 0.0217561591, -0.0184189584, -0.0340188816, -0.0512883179, -0.0663126186, -0.0012163839, 0.0048897518, 0.0237059742, -0.0233896747, -0.00121783, 0.0566454306, -0.0205345862, 0.0180275477, -0.003155112, 0.0180481877, 0.0196882132, 0.0099274535, -0.0425179601, -6.878929024e-33, 0.0387589447, 0.0026133414, 0.0332155935, -0.0279619042, 0.0084056035, -0.0154017778, -0.0251771118, 0.0248589236, -0.0383043699, -0.0139509831, -0.0230400451, 0.0400062799, 0.0123206722, 0.0193478726, 0.0285590682, -0.0464617051, 0.0317545868, -0.0100185461, -0.0124599738, 0.0369621888, 0.0625103042, 0.0337067991, 0.0474417359, -0.0662762374, 0.0364034213, -0.0189443585, 0.0015644076, 0.0151676126, -0.0115184337, 0.0236853678, -0.0412992686, 0.0469923094, 0.0116021158, -0.0196455363, -0.0367390327, 0.0001158764, -0.0119969612, -0.0235942211, -0.004816175, 0.0145922005, 0.0158337727, -0.0272416212, 0.0727774426, -0.0388151817, 0.0061278865, -0.0074375439, 0.0204335321, -0.0328808837, -0.0633593276, 0.01456689, -0.0177183505, -0.0098689096, 0.0134292664, 0.0068220589, -0.0197876282, -0.0005504072, -0.0196837131, -0.0119084679, -0.1136138067, 0.0330453739, -0.0337107405, 0.0023231304, 0.0297131613, -0.0040905443, 0.0190405659, 0.0972537324, 0.0903108567, 0.0549101196, -0.0481310301, -0.0257932618, 0.0337670259, 0.0497307181, 0.0407070294, -0.0055050887, 0.0933358595, -0.0369349085, -0.0370230302, -0.0367952287, -0.0052305399, 0.0133227445, 0.0783360526, -0.0039742882, -0.0546515472, -0.0551751927, 0.0028439106, 0.0351752229, -0.0477394499, -0.0353906751, 0.0198827125, 0.0210852772, 0.0155597692, -0.0405648053, 0.0293396357, -0.0557875037, 0.0271224156, 0.040854875, -0.0144074196, -0.0023687552, -0.0121506229, -0.0111936182, -0.0362255685, -0.0158410445, -0.0195240397, 0.0129786432, 0.0208220016, 0.0437664427, 0.0435926057, 0.0113237053, -0.0557710007, -0.0194412619, -0.0532589741, 0.0166813694, 0.0263190027, 0.0036451276, 0.0452560931, 0.0087331347, -0.0115310708, -0.015896203, 0.0150730778, 0.0128100608, -0.0108525362, 0.0267448556, -0.0201224927, 0.0078684781, -0.0135755669, -0.0217626188, -0.0612862557, -0.0078934692, 0.0317859203, -0.0712701529, 0.0044173542, -0.0136973634, 2.872e-7, -0.031346947, 0.0406859629, -0.0063335267, 0.0152073475, 0.0293070618, 0.0159911104, -0.0164426286, 0.0122725628, 0.022945486, -0.0805295035, -0.0147137232, 0.0518012457, 0.0564617701, 0.0622196086, -0.1100768074, -0.025837725, -0.0660861805, -0.0010490202, -0.0495676845, -0.0431121811, 0.0196739417, 0.0778786987, 0.0404218361, -0.0248570405, 0.0229567327, -0.0957703367, 0.0982067958, -0.0138162728, 0.0156286452, 0.0180199984, 0.0625414476, 0.0427555442, -0.0203947835, 0.0257668663, -0.0033516297, -0.0218093935, -0.0267420281, 0.0306424107, 0.0081166923, -0.0639264882, -0.0185489487, -0.0134553779, 0.0017478698, -0.0288604815, 0.05107297, -0.045621112, -0.0422733203, 0.0090607256, 0.0047845314, -0.0211422481, 0.0473972596, 0.0065145572, -0.0068052593, 0.0119855693, -0.0183024257, 0.0375009701, 0.035218928, -0.0822376683, -0.0245981161, 0.0202004015, -0.0556203648, -0.0529842675, -0.0245088991, 0.0010124976, 0.0132074477, -0.0494430177, 0.0346402153, 2.685309627e-34, 0.0400599018, -0.0167173482, 0.0394766405, 0.0539950579, 0.0257029254, -0.0066775135, 0.0557594746, 0.0187180918, 0.0277404785, -0.0414204933, -0.0117425956 ]
http://arxiv.org/abs/2406.17639v2
http://arxiv.org/pdf/2406.17639v2
Mitigate the Gap: Investigating Approaches for Improving Cross-Modal Alignment in CLIP
2024-06-25T15:24:02
2024-06-26T10:58:48
Sedigheh Eslami, Gerard de Melo
cs.CV
cs.CV, cs.AI, cs.CL, cs.LG
Contrastive Language--Image Pre-training (CLIP) has manifested remarkable improvements in zero-shot classification and cross-modal vision-language tasks. Yet, from a geometrical point of view, the CLIP embedding space has been found to have a pronounced modality gap. This gap renders the embedding space overly sparse and disconnected, with different modalities being densely distributed in distinct subregions of the hypersphere. In this work, we aim at answering two main questions: 1. Does sharing the parameter space between the multi-modal encoders reduce the modality gap? 2. Can the gap be mitigated by pushing apart the uni-modal embeddings via intra-modality separation? We design AlignCLIP, in order to answer these questions and show that answers to both questions are positive. Through extensive experiments, we show that AlignCLIP achieves noticeable enhancements in the cross-modal alignment of the embeddings, and thereby, reduces the modality gap, while maintaining the performance across several downstream evaluations, such as zero-shot image classification, zero-shot multi-modal retrieval and zero-shot semantic text similarity.
Mitigate the Gap: Investigating Approaches for Improving Cross-Modal Alignment in CLIP: Contrastive Language--Image Pre-training (CLIP) has manifested remarkable improvements in zero-shot classification and cross-modal vision-language tasks. Yet, from a geometrical point of view, the CLIP embedding space has been found to have a pronounced modality gap. This gap renders the embedding space overly sparse and disconnected, with different modalities being densely distributed in distinct subregions of the hypersphere. In this work, we aim at answering two main questions: 1. Does sharing the parameter space between the multi-modal encoders reduce the modality gap? 2. Can the gap be mitigated by pushing apart the uni-modal embeddings via intra-modality separation? We design AlignCLIP, in order to answer these questions and show that answers to both questions are positive. Through extensive experiments, we show that AlignCLIP achieves noticeable enhancements in the cross-modal alignment of the embeddings, and thereby, reduces the modality gap, while maintaining the performance across several downstream evaluations, such as zero-shot image classification, zero-shot multi-modal retrieval and zero-shot semantic text similarity.
[ -0.0385551266, 0.0596274659, 0.0174322352, 0.0479889587, -0.040716894, -0.01166711, 0.0372344814, -0.0024257374, -0.0847809985, -0.0456418172, -0.0974812359, -0.0301979017, 0.0268390309, 0.0430643298, 0.0218666736, -0.0384461023, 0.0313609466, 0.0069858688, -0.0323454179, -0.0017686086, 0.0290400609, -0.0052950014, -0.0257307477, 0.019968491, 0.0036294907, -0.0019730909, -0.029474292, -0.0015691731, 0.0352342203, -0.012128341, -0.0164365452, 0.0390502065, 0.0491584577, 0.0193275642, 0.0000017254, 0.0013124341, 0.0312345866, 0.0222930722, 0.0078272307, -0.007323259, 0.0506880283, -0.070423305, 0.0219613872, -0.0158485658, -0.0071820845, -0.0727281347, 0.0836183131, 0.042020075, 0.0283044185, 0.0684457719, -0.0474121198, -0.0069144405, 0.0114890551, -0.0317668654, 0.0133610126, 0.014140157, -0.0063610449, -0.0131583977, -0.0112451455, 0.0295590125, -0.0584094264, 0.0416524038, 0.007756527, 0.0232001189, -0.0259166267, 0.0500517525, -0.0241084322, -0.0267689694, -0.0290658828, 0.0949219093, -0.0106873941, -0.0114302058, -0.0418375023, 0.0007325916, -0.0008618569, 0.0300708041, -0.0265162904, 0.0320459008, 0.0244021583, 0.0446959026, -0.0343453512, 0.0306258928, 0.0163354184, -0.0359617099, 0.0056366669, -0.0304381568, 0.0406714119, -0.0007257715, -0.0274103377, 0.0194088817, -0.0019978629, -0.0250917822, -0.0079758819, 0.0008486839, 0.0000735911, -0.0245760288, -0.0759448931, -0.0078827906, -0.0059187743, 0.0268387832, 0.0522572361, 0.052849628, -0.0591201968, 0.0538148656, -0.0289977808, 0.0237785317, -0.0240088757, 0.0776508674, -0.0357540362, -0.0180679634, -0.0118763382, -0.0069658705, 0.0074811741, -0.0131243942, -0.0502358824, -0.0504870601, -0.0082780095, 0.0008043865, -0.0221452471, 0.0082875816, -0.0088310875, 0.0179732945, -0.0401922837, -0.0284525063, 0.0100283427, 0.0730513036, 0.0012608739, 0.0551451109, -0.0351434238, -0.0372608602, -0.0203444976, 0.0526568815, 0.020349618, -0.0257803202, -0.0142847998, 0.0728635639, -0.0495614558, -0.0442570969, 0.0022620447, 0.0090734465, -0.0357888415, -0.0346338376, 0.0362764075, -0.0294065457, 0.0469269976, 0.0117988437, -0.0309870038, -0.0091933021, -0.0106329303, -0.0166995712, -0.0081102997, 0.1020955071, 0.0064513502, -0.0211410951, -0.0070400331, 0.0012797209, -0.0069025038, 0.059057273, 0.0035462086, -0.0096334415, 0.0123512736, 0.0293249283, -0.0115101179, 0.0330004841, 0.0211375989, 0.0202157237, -0.0728412271, -0.0245802272, 0.0125097129, 0.0204258915, 0.0602327809, 0.0049933158, -0.0549679026, -0.0399544798, 0.0316788778, 0.0326724499, 0.0173130874, -0.0046902313, 0.0219339188, 0.0175266825, -0.0337174982, -0.0027797152, -0.0384936184, 0.0174345486, -0.0151312258, 0.0160891004, 0.0196626671, 0.0434931926, -0.0560308471, -0.0057721492, 0.0201951899, -0.0269327387, 0.0030975547, -0.015914306, 0.0701473802, -0.030731149, -0.0292179659, 0.0559089184, 0.0428476892, -0.0490898229, 0.0182474274, 0.0062592477, -0.0244459305, -0.0185294375, -0.0161262099, -0.0125969537, -0.0509348959, 0.0226364993, -0.0040295138, 0.0463774092, -0.0073227943, -0.0104336226, 0.0514021926, -0.0491175056, -0.0851776525, -0.0558604896, -0.0056714853, 0.0283110552, -0.0253865942, 0.0323616602, 0.0193104427, -0.0172058009, 0.0750669092, 0.0211886596, 0.0214527845, 0.022209486, -0.042844791, 0.0468594804, 0.0251131747, -0.0140446313, -0.043603465, -0.009048623, 0.010255238, -0.058346618, 0.0207595266, 0.0376248732, -0.0592223331, -0.067961812, 0.0487529896, -0.0928893909, 0.0204165597, 0.0251019634, -0.0211431514, -0.0437157638, 0.0173821915, 0.0180427786, 0.0578181297, -0.0407846086, 0.0014284025, -0.0215418022, 0.0051032789, 0.0001615301, 0.0247909371, -0.0009529009, 0.00409543, 0.0582801029, 0.0420164466, 0.0536888875, 0.0304152351, -0.037239369, 0.0139024965, -0.0084945215, 0.0217859, 0.0199216865, 0.0007993897, -0.0256661419, -0.0081371451, -0.1193065941, -0.0080174049, -0.0148731833, 0.0263969004, 0.0354379751, -0.004609481, -0.0142148025, -0.0518507063, -0.039051868, 0.0149865849, 0.0131800901, 0.0118245929, 0.0169000346, 0.0326341651, 0.0074252877, 0.0033035625, 0.0001011155, 0.0221269727, 0.0517623238, -0.0245807506, -0.0016546389, -0.0221003201, 0.0615608506, -0.0246052444, 0.0209080968, -0.027069157, -0.0287440401, -0.0003372665, -0.0079600615, 0.0474515334, 0.0250381343, -0.0291349422, 0.005656661, 0.0448757075, -0.0112827895, -0.0442839302, -0.0214473847, 0.0051852507, -0.0036950654, -0.0232610572, 0.040503215, -0.0327294841, 0.0441114828, -0.0272429101, 0.0233328808, -0.0650217086, 0.0082023218, 0.0113350144, -0.0186860152, -0.0440514535, 0.0213731714, -0.0132436138, 0.0232977867, -0.0454928949, -0.0442666374, 0.0339690261, -0.041044604, -0.0206199437, -0.0134049989, 0.0014202218, -0.0078383321, 0.0399076901, -0.0175280776, 0.0289367251, 0.0126097295, -0.0454452261, -0.0679962561, 0.0004919155, 0.0098423595, 0.0109295724, 0.0479493067, 0.0122736171, 0.004682967, -0.063791953, -0.0532379411, -0.0136397444, -0.0689248443, -0.0136089586, 0.0086246906, -0.030324379, -0.0047433064, -0.0459596254, 0.0127592906, 0.0131558254, 0.0306392275, -0.0393132493, 0.0139670325, -0.0181556679, 0.082877636, 0.0118026342, 0.0117809093, -0.0350550972, -0.0265528914, -0.0437973328, 0.0370589904, 0.0225244612, 0.0295129344, -0.0249308385, 0.0222718567, 0.0513255298, 0.0511684008, -0.0250913911, 0.020801058, 0.0780705288, -0.0008559256, 0.0024786536, 0.0622095689, -0.0089674862, -0.0859280974, -0.0517675541, 0.0052966285, 0.0490082167, 0.0148045607, -0.0032947939, -0.0584728532, -0.0629200265, -0.0175173283, -0.0392427444, 0.0473558083, 0.0431267545, -0.0220088325, -0.0889828354, 0.0085395146, 0.0117562562, -0.0431505106, -0.0544931479, -0.0192243867, 0.0288482644, 0.001288739, 0.0104733519, -0.0407094173, -0.0213281605, -0.0194426682, 0.0071254103, 0.0743589401, -0.0901874527, -0.0319769047, -0.0335274264, 0.0169722084, -0.0072969054, 0.0233054329, -0.0204657577, -0.0822349563, 0.0046938118, -0.0191198513, 0.0021391816, 0.0183577612, 0.0551617518, -0.0355530009, -0.0238939822, 0.0010454095, 0.0251303911, -0.0252778698, -0.018076973, -0.0527519695, 0.0484943874, 0.0000927934, 0.0411734916, -0.0182747655, 0.0191109926, 0.0060213222, 0.0380194858, 0.0001809757, -0.0032545084, -0.0049557383, 0.0156904366, 0.0194636863, -0.0166371502, -0.0186323319, 0.0067690481, 0.0061337641, 0.0132288327, -0.0135298027, -0.0228270497, -0.0046009542, 0.0581945442, -0.0084138326, -0.0571496971, -0.04480391, 0.0085004559, -0.0541474409, 0.0583553128, -0.039125815, 0.0062084561, -0.0169656165, 0.0304273274, 0.0325197726, -0.0191924274, -0.0280560236, -0.0253190231, -0.0730733126, -0.0642375723, -0.0563618056, 0.0208327491, 0.0026956236, 0.081773296, 0.0297749098, 0.0080920896, 0.1050246656, 0.0215157997, 0.0019388287, 0.0158339422, 0.0423012078, -0.0620292164, 0.0026105475, 0.0194229539, -0.0368654057, -0.0059942049, 0.0305879116, 0.0030766509, 0.0282508228, -0.0419930704, 0.0156906433, -0.0443435796, 0.0723844841, -0.0238592904, 0.0582260452, -0.0477235951, -0.0137036666, -0.0300151184, -0.0000162471, -0.0325527303, -0.0138027063, 0.0038915293, -0.0195966996, 0.0222706497, 0.0791600719, -0.0022304, 0.0300996453, -0.0339520276, 0.0311019029, 0.0346437842, -0.0190042034, -0.0210639387, 0.0442092009, 0.0018870056, 0.0079767136, 0.0007066087, 0.0188424215, -0.0554482266, 0.0814388916, 0.0337239243, -0.0120972786, -0.0577831157, 0.0759395063, -0.0158049315, 0.0299266595, -0.0152445938, 0.0054162527, 0.0488435738, -0.0293325167, 0.0181934368, 0.0079637477, -0.0323745087, 0.00214079, 0.042896226, -0.0241623726, -0.0050471597, 0.0578787476, 0.0009349035, -0.0096720094, 0.0163510554, 0.0631263182, -0.0388230346, 0.005499843, 0.0011775128, 0.0156902168, -0.0180570539, -0.0005009493, 0.0854855776, 0.0233423673, -0.027691897, -0.0136380261, -0.0120773995, -0.0479865521, 0.0122420285, 0.0185360685, -0.0521467663, 0.0384397581, 0.0328463539, 0.03670872, 0.0018959283, -0.0137434248, 0.0110870171, -0.0251569916, 0.0176851414, 0.023903504, -0.0085762655, -5.573221857e-33, -0.0256085638, 0.0368373431, 0.0115708066, -0.0387414284, -0.0423778445, 0.0537800118, -0.0181444231, 0.0154226953, -0.0113003589, 0.0184387937, 0.0332004391, -0.0062822015, 0.0160430875, 0.0390450731, -0.0008171732, 0.0064430437, 0.0413918234, -0.0102685653, -0.0290128719, -0.010879932, 0.0245962758, 0.005573954, 0.0539902337, -0.0428803042, 0.0018428252, 0.0149367889, 0.0036762683, 0.0130965766, 0.0671936497, 0.0649371073, -0.0245161448, 0.0259664878, 0.0112867728, -0.010108823, -0.0176005997, 0.033913631, -0.0356823169, -0.0283250902, 0.0017419882, 0.0276473183, -0.0602493323, -0.0200153645, 0.0257274322, -0.0333458669, -0.0159992334, -0.0589568578, 0.0464580096, 0.0110184448, 0.002354525, 0.0205121599, -0.0528300256, 0.0128830932, -0.0126599651, -0.0499222204, 0.0616904013, 0.0610179342, 0.0610297993, 0.0186339207, -0.0497220308, 0.0347156376, -0.03404947, 0.0198141281, 0.0733154863, 0.0579641536, 0.0177070145, 0.0739013702, 0.0074229292, -0.0693867728, 0.0065039191, -0.0305771846, 0.0485739075, -0.0042903377, -0.0186992958, 0.1028881893, 0.0623428188, 0.016159106, -0.03111808, 0.0262424592, -0.050021138, 0.0105764316, 0.0005444636, 0.045857206, 0.0080700805, -0.0095743416, -0.0293612927, -0.0703542531, -0.0357648917, -0.0496190637, 0.0173471719, 0.0216391087, 0.0013194935, 0.0376476645, 0.0980772525, -0.0192112178, 0.0060804291, 0.0366580337, -0.0278525464, 0.0500884503, -0.014229496, 0.0066288887, -0.0444367304, -0.0343645439, 0.0574803539, -0.0099755879, -0.0011121548, -0.0056890338, 0.0495224446, 0.0309867654, -0.0445392579, -0.0221862867, -0.0197772887, -0.0572133549, 0.0074664704, 0.0623515248, 0.0595589727, 0.0296105687, -0.0242033508, 0.0445410721, 0.0168413762, 0.018812079, -0.0181368291, -0.0565546788, 0.0065363962, 0.0142197041, 0.0084845303, 0.0232500415, -0.0742770135, -0.0303730667, 0.0178647432, -0.053245645, -0.028006969, 0.0083239488, 2.459e-7, 0.0130660255, -0.008477076, 0.0316102915, -0.0511874296, 0.014791688, 0.0444585755, -0.0191127229, 0.0388957448, -0.0145226391, -0.0447655767, 0.0132945245, 0.0263134614, 0.0243445393, 0.0110545075, -0.00700119, 0.0379628055, -0.0456370078, -0.0182140023, -0.0179987289, 0.0313441046, 0.0799404457, 0.120052807, -0.0261917505, -0.0001989045, 0.0165534988, -0.0268511642, 0.0870765224, -0.0261701196, -0.0175312925, -0.0034413682, -0.0580840558, 0.0162428655, -0.0175801143, 0.0062654344, 0.005820828, 0.0294847284, 0.0311988723, 0.0279564336, -0.0228269212, -0.0281509943, -0.0306715705, 0.0514018349, -0.0450681709, -0.0187798236, 0.0672516972, -0.0864263102, -0.0422942936, -0.0839109793, -0.046379067, 0.0356538929, 0.0255331676, -0.0083712162, -0.0319680348, 0.0169736929, 0.0314646251, -0.0157443155, -0.0449555628, -0.0625921115, -0.0232636463, 0.0357444733, -0.0345050991, 0.0147421891, -0.0229547787, -0.0629999116, 0.0754748657, -0.0446875058, -0.0189995505, 1.802983796e-34, -0.0259150676, -0.019212326, 0.0230517089, -0.0387871265, -0.0115303099, -0.0286990367, -0.0113611454, -0.0314237177, -0.0090776132, -0.0300694108, -0.0067888102 ]
http://arxiv.org/abs/2406.17633v1
http://arxiv.org/pdf/2406.17633v1
Knowledge Distillation in Automated Annotation: Supervised Text Classification with LLM-Generated Training Labels
2024-06-25T15:20:25
2024-06-25T15:20:25
Nicholas Pangakis, Samuel Wolken
cs.CL
cs.CL, cs.LG
Computational social science (CSS) practitioners often rely on human-labeled data to fine-tune supervised text classifiers. We assess the potential for researchers to augment or replace human-generated training data with surrogate training labels from generative large language models (LLMs). We introduce a recommended workflow and test this LLM application by replicating 14 classification tasks and measuring performance. We employ a novel corpus of English-language text classification data sets from recent CSS articles in high-impact journals. Because these data sets are stored in password-protected archives, our analyses are less prone to issues of contamination. For each task, we compare supervised classifiers fine-tuned using GPT-4 labels against classifiers fine-tuned with human annotations and against labels from GPT-4 and Mistral-7B with few-shot in-context learning. Our findings indicate that supervised classification models fine-tuned on LLM-generated labels perform comparably to models fine-tuned with labels from human annotators. Fine-tuning models using LLM-generated labels can be a fast, efficient and cost-effective method of building supervised text classifiers.
Knowledge Distillation in Automated Annotation: Supervised Text Classification with LLM-Generated Training Labels: Computational social science (CSS) practitioners often rely on human-labeled data to fine-tune supervised text classifiers. We assess the potential for researchers to augment or replace human-generated training data with surrogate training labels from generative large language models (LLMs). We introduce a recommended workflow and test this LLM application by replicating 14 classification tasks and measuring performance. We employ a novel corpus of English-language text classification data sets from recent CSS articles in high-impact journals. Because these data sets are stored in password-protected archives, our analyses are less prone to issues of contamination. For each task, we compare supervised classifiers fine-tuned using GPT-4 labels against classifiers fine-tuned with human annotations and against labels from GPT-4 and Mistral-7B with few-shot in-context learning. Our findings indicate that supervised classification models fine-tuned on LLM-generated labels perform comparably to models fine-tuned with labels from human annotators. Fine-tuning models using LLM-generated labels can be a fast, efficient and cost-effective method of building supervised text classifiers.
[ 0.0465122163, 0.0571201332, 0.0077259089, 0.000175944, -0.028061593, 0.0340343527, 0.0706859306, -0.0116423415, 0.0471397899, -0.1074874476, -0.0166809522, 0.0051544346, -0.0466479808, -0.0170024242, 0.031637378, -0.0112535274, 0.0275900308, -0.0109264357, 0.0079099797, -0.009094758, -0.0248459727, 0.0083295638, 0.0451833494, 0.0209162775, 0.0148313688, -0.0376415215, 0.0031630183, -0.0110277096, 0.0189660564, -0.0330047272, 0.0418825857, 0.0886411816, 0.0367598645, 0.0562503263, 0.0000020025, -0.0481911749, -0.0214555003, 0.0436819308, -0.0309915133, 0.0448551029, 0.0470426157, -0.0342157818, 0.0119302301, 0.0063522318, -0.0545496754, -0.0345804468, 0.028772369, 0.0312328432, 0.0489580929, 0.0696472749, 0.0079361461, -0.0631700382, 0.0454625897, -0.0489745997, 0.0282765403, -0.0665379912, 0.0667912886, 0.0130281243, -0.0281841867, 0.07390102, 0.004346699, 0.0172850452, 0.0035835828, -0.0138927251, 0.0378796682, 0.0299833976, -0.0527727827, -0.0005836997, 0.0112083675, 0.0248775724, 0.0317540877, -0.0391694047, 0.0393267013, -0.0056765713, 0.006344202, 0.0376281515, -0.0143907145, 0.0430698618, -0.027160719, 0.0113138035, 0.0398393571, -0.0223867521, 0.038951166, 0.0320612267, 0.0291704405, 0.045599144, -0.0164393168, -0.0280167721, 0.0410777815, -0.0023766349, -0.0394845642, -0.0576022863, -0.0020085596, -0.0137917502, 0.031328097, 0.0094476892, -0.0326545089, -0.0267717615, 0.034595944, -0.0380373038, 0.0035501067, -0.0053852783, -0.00644481, 0.0362989344, 0.0082563134, -0.007250783, 0.009051322, 0.0172594097, 0.0193361081, -0.0008767621, -0.093989253, -0.023042839, -0.0170834977, 0.0337460302, -0.027211966, 0.0097616278, -0.0631773472, 0.018931631, -0.0115888249, 0.0106053585, 0.0233849455, -0.0161815453, -0.01789877, 0.0146062756, -0.009966189, -0.0349549577, -0.003433889, 0.0277938005, -0.009923324, -0.0359133109, 0.0148664918, 0.015030453, 0.0039725653, -0.0329899564, 0.016740784, 0.0252552498, 0.0058765616, -0.0466761626, -0.0585462376, -0.010795177, -0.0390789732, -0.0445786938, 0.0093759177, -0.059419658, -0.0061535649, 0.006682232, 0.0055485838, -0.0818926543, 0.0014383348, 0.0128174275, -0.0240499303, 0.0676328987, -0.0204614978, -0.0200954806, 0.059344586, 0.0199837368, -0.0004074882, 0.0500468835, 0.0191533584, 0.0508405827, 0.039153669, -0.0339620039, 0.0542678796, 0.0234842245, 0.0003528683, 0.0132002467, -0.0084403818, -0.0031146198, 0.0118597541, -0.027113419, 0.0176637769, 0.032728646, -0.0169011038, 0.0105586536, 0.0812616646, 0.0638249442, -0.0044248085, 0.033922717, 0.0248701852, 0.0359456949, 0.0062651439, -0.026009243, -0.0285398383, -0.0152014839, -0.0328092501, -0.0221075509, 0.0214042868, -0.0005767003, -0.0045477976, -0.0231011789, -0.0148519995, 0.0025890248, 0.0052129561, 0.0156472698, 0.0734201372, 0.0071842405, -0.0039782766, 0.0061882059, 0.0419809297, -0.009694132, -0.0056055272, -0.0471742488, 0.0211453643, 0.044048842, 0.0151866022, -0.0208366681, 0.0389954187, -0.0508827157, 0.0482964665, 0.0123106744, 0.0242749359, -0.0242706295, -0.0216159988, -0.0256124083, -0.0352979042, -0.0273739509, 0.0422968641, 0.0131578501, 0.0330662243, -0.004352977, 0.0344836973, -0.0535943843, -0.0653041378, 0.068968527, -0.0196877941, -0.0572150722, 0.0126866177, -0.0059413384, 0.052597411, 0.0219954886, -0.0221048538, -0.0155576877, 0.0273840483, -0.0099534513, -0.0050845924, -0.0504737012, 0.0354535468, -0.0137978662, -0.0044792253, -0.0085837692, -0.0068073254, 0.0026119486, -0.0524448641, -0.0118935108, -0.0292208437, 0.0056594135, 0.0211575702, -0.0867067277, 0.0054840068, -0.0399364233, 0.0040991143, -0.0622528642, 0.0726017058, 0.0182952844, 0.0090693738, -0.0522071719, -0.0442060195, 0.0049097883, 0.0065092454, -0.0185802858, -0.0239256304, 0.0058328183, 0.0153254662, 0.0290654153, -0.0042730356, -0.0298485644, -0.0312928148, -0.0799603015, 0.0239350609, -0.0201500319, 0.0171419103, -0.0116362106, -0.0289258044, -0.0114279669, 0.0292900726, -0.0212575831, -0.0420408733, -0.0415460356, 0.0267364122, 0.0119237062, -0.0150380936, 0.0154271517, -0.0074543203, 0.0079723727, -0.0038919484, 0.0450146198, 0.0151316049, -0.0091555119, -0.034838818, 0.0177220274, 0.0100437161, 0.0088398121, -0.0153065594, -0.0188060757, -0.0038934911, 0.0308709443, -0.0202815477, 0.0512626581, 0.0105617745, 0.0483615659, 0.0307530575, 0.037698552, -0.0326893702, -0.0021741679, -0.0255237054, -0.0363709703, -0.0330053046, -0.0149715049, 0.0009117819, 0.0391837545, 0.0058688191, -0.0295788348, -0.0267714281, 0.0342976339, -0.0066628531, -0.0363633521, -0.0386585407, -0.0128638837, -0.0176960304, 0.0334963538, -0.0077158203, -0.0461153127, 0.0250236113, -0.009822268, 0.0355741419, -0.0325242095, -0.0445953831, -0.0180740394, 0.0598524734, 0.0125651127, 0.0022036517, 0.0276070163, 0.0088595282, -0.0165873282, 0.0067509264, 0.0367418826, -0.025282206, 0.0026881169, -0.0319982395, 0.0162834097, -0.0189359728, -0.048734732, -0.0338391885, -0.083440043, -0.0384255797, 0.0352707505, 0.0447607897, -0.0118642133, -0.0083756503, 0.018560892, 0.0132753989, 0.0523034148, -0.0115944659, -0.0298403632, -0.012641116, 0.0134533616, 0.0231525768, -0.0506061763, -0.0315555297, -0.011132584, -0.0029056617, 0.0321871415, 0.0030874165, 0.0247014835, -0.0147936018, 0.0025032342, 0.0292866472, -0.0590331517, -0.0333109573, 0.0090185627, 0.0398993269, -0.0573739596, -0.0034245604, 0.1094403118, 0.0292256232, -0.0557742566, -0.0126972357, 0.0301299468, 0.0114154592, 0.0249705017, 0.0076396344, -0.102135241, -0.0189325847, 0.0111726578, 0.0445082858, -0.0377389342, -0.0517733023, -0.015032107, -0.1083165407, 0.0570128523, -0.0107210493, -0.0492438003, 0.0296292584, -0.0718442053, -0.0061179716, -0.0304905977, -0.0321220197, -0.0118064173, -0.0093150148, 0.0274214428, 0.0503100902, -0.0354731232, -0.0774137825, 0.0267490186, -0.0199725237, 0.0578104667, 0.0363713056, 0.0413370989, -0.0232815258, -0.0288699586, 0.0613985173, -0.0187508482, 0.0009845822, 0.0062159034, 0.0635435879, -0.0285684951, 0.0078133876, 0.014653692, 0.0806675255, -0.0122331819, -0.0347559676, 0.0401863717, 0.0530175492, -0.0115613444, -0.0127140945, 0.009514289, -0.0361111425, 0.0060831597, 0.0754059181, 0.0121306647, -0.0592175089, -0.0094015626, -0.0187321156, 0.0407448672, -0.021157993, -0.0245192498, 0.0453901254, -0.0594643764, -0.0642217025, -0.0385400355, -0.0289694667, 0.030523004, 0.0252715386, 0.0290347636, 0.0403009914, 0.013231812, 0.0020462777, -0.0451012999, 0.0172453709, -0.0724897757, -0.0217944086, -0.0117871258, -0.0325416923, 0.0489110462, -0.0265064407, -0.0432582945, 0.0090678213, -0.0707792938, -0.0655828416, 0.0211400278, 0.001373945, 0.0223363396, 0.0275199413, -0.0044040862, -0.0820372477, 0.0788381249, 0.1034948379, -0.0087443488, -0.0125958929, -0.0041743559, -0.0042928271, -0.0206620079, -0.0008037836, -0.0425185226, 0.033864297, 0.038351126, -0.025664553, -0.0570451096, 0.0039663208, -0.0037654843, 0.0135031184, 0.0835413784, -0.0651226491, -0.0124014663, -0.0224181581, -0.0067150323, 0.0017540725, -0.0194899049, -0.0101954564, -0.0094456198, 0.033382684, -0.0075030741, -0.0408499353, 0.0540857352, 0.0059164721, 0.0651044026, -0.068368338, -0.0066630766, 0.0243481398, 0.0329612866, 0.0057201996, 0.0003430673, -0.014830417, 0.0141904633, 0.0138718355, 0.0440548323, -0.0050568916, 0.0276842695, 0.0093551567, 0.0326279737, 0.009928056, -0.0032143786, -0.0736090094, -0.023099333, -0.0472258031, 0.0590159483, 0.0522629656, -0.0088233585, 0.0148659898, -0.0359358341, -0.0456302539, -0.0264704656, 0.0213806089, 0.0127611076, -0.0186941326, 0.0023928082, 0.0049393298, -0.0753917322, 0.0227785371, 0.0732867941, 0.0720528662, 0.0390678011, 0.0187251437, -0.0689429864, -0.0337129794, -0.0136706121, 0.0007141979, 0.0301619247, 0.0173443425, -0.0119483918, -0.0556100421, -0.0468319021, -0.000919894, -0.0277505331, -0.0066806534, -0.0104956822, 0.0134516573, 0.0613367744, -0.0315298848, -0.0266477205, -0.0804135352, -0.0053580543, 0.0111138541, -0.0285441149, -0.0076958714, -6.880133905e-33, 0.02086775, -0.0447402187, -0.0002580621, 0.0212800317, -0.0039385576, -0.0044964957, 0.0144062992, -0.0329219401, 0.0069599207, 0.0378020629, -0.0200987589, -0.0082827592, 0.0048432811, 0.0490807667, 0.0342044383, 0.0050112098, 0.0359438583, -0.011438028, -0.0168784261, 0.0380073972, 0.0454746075, 0.0069563747, 0.0441167913, -0.1046309322, 0.0675067678, 0.0005055683, -0.0185995195, -0.0067273318, 0.0186182093, 0.0275019333, -0.0262798276, 0.0511711761, 0.0328505188, -0.0576267801, 0.0431025438, 0.0323343426, -0.0777548328, -0.0086611751, 0.0139465416, -0.0278621018, 0.0296200253, -0.0611190498, 0.0403753445, -0.010919475, -0.0163125023, 0.0222935099, 0.05567744, -0.0140306214, -0.053601779, -0.1230171844, -0.0357833691, 0.0286772531, -0.021939259, 0.0148897255, 0.0235092472, 0.0069944058, 0.0009836031, 0.0646994784, -0.0672383383, 0.0024035648, 0.02818349, -0.0129896831, 0.0175570808, 0.0372568034, 0.0194902308, 0.0539234653, 0.0150311068, 0.0253044143, -0.0103644421, -0.0127662746, 0.0225728452, 0.0058632051, 0.0016743864, 0.0943515226, 0.0481503941, -0.0584196784, -0.0820355341, 0.0177206788, 0.0012869649, -0.0426096097, 0.0565675385, 0.0046209092, -0.0314375088, -0.0427813195, -0.0721071437, -0.0330547765, -0.0121737365, 0.0129842786, -0.0025249389, 0.000208644, -0.051733721, 0.0322811268, 0.0165157616, -0.0265532024, -0.0541842729, 0.0070822597, -0.0068762191, 0.0132173188, -0.0050267489, -0.016488893, -0.0689053908, -0.0223009977, 0.0153618585, 0.0330438986, 0.037804354, -0.0095147043, 0.0897523388, -0.0051342156, -0.0727668256, 0.0104124052, -0.0109700691, 0.0498068966, 0.0169484019, -0.0004328645, 0.0312203839, 0.009936302, 0.0271278378, 0.0146946134, 0.0756371468, 0.0189400222, 0.0248400103, 0.0036879962, -0.006617988, -0.012028316, 0.0160844624, 0.0045824265, 0.0470863879, 0.015996689, 0.0441454686, -0.0313747972, 0.0144802472, -0.0132637415, 2.839e-7, 0.0610301122, 0.0704688355, -0.0365063138, -0.0515829697, 0.0247435048, 0.0503662266, 0.0329140089, 0.0124999601, 0.0013376505, -0.061699871, 0.0211705193, 0.005981178, 0.0743765309, 0.0231404901, -0.0847544074, -0.0155923124, -0.0621509291, 0.0025278332, -0.0645913258, 0.0103032198, 0.0347770005, 0.0678824484, 0.0459001623, -0.0126910703, -0.0195465311, -0.0462584868, 0.0574705526, -0.0257468801, -0.0073176264, 0.0386499017, 0.0123901321, 0.0399971791, -0.0027154849, 0.0416021831, -0.0436285883, -0.0001178838, -0.0137977777, 0.0715053231, 0.0026421524, 0.0164278597, -0.004300769, -0.0089496719, 0.013254948, -0.054495357, 0.0904177651, -0.0729797781, -0.0529326908, -0.0305253733, -0.0289854277, -0.0532480441, 0.0360408388, -0.0233121961, 0.0396176055, 0.0265647434, -0.0233411063, -0.0071117105, -0.0174724292, -0.0694988668, -0.0035536611, 0.0240768883, -0.0364629701, -0.0513084345, 0.0243199673, 0.0089146756, 0.0114036957, -0.0427058712, 0.0525861233, 2.618552568e-34, 0.0291768182, -0.0072799795, -0.0080590788, 0.002427167, 0.0341612101, 0.0391906127, -0.0125968261, 0.0162350386, 0.0015188075, -0.0635468066, 0.0305667836 ]
http://arxiv.org/abs/2406.17626v1
http://arxiv.org/pdf/2406.17626v1
CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference
2024-06-25T15:13:02
2024-06-25T15:13:02
Erxin Yu, Jing Li, Ming Liao, Siqi Wang, Zuchen Gao, Fei Mi, Lanqing Hong
cs.CL
cs.CL, cs.AI
As large language models (LLMs) constantly evolve, ensuring their safety remains a critical research problem. Previous red-teaming approaches for LLM safety have primarily focused on single prompt attacks or goal hijacking. To the best of our knowledge, we are the first to study LLM safety in multi-turn dialogue coreference. We created a dataset of 1,400 questions across 14 categories, each featuring multi-turn coreference safety attacks. We then conducted detailed evaluations on five widely used open-source LLMs. The results indicated that under multi-turn coreference safety attacks, the highest attack success rate was 56% with the LLaMA2-Chat-7b model, while the lowest was 13.9% with the Mistral-7B-Instruct model. These findings highlight the safety vulnerabilities in LLMs during dialogue coreference interactions.
CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference: As large language models (LLMs) constantly evolve, ensuring their safety remains a critical research problem. Previous red-teaming approaches for LLM safety have primarily focused on single prompt attacks or goal hijacking. To the best of our knowledge, we are the first to study LLM safety in multi-turn dialogue coreference. We created a dataset of 1,400 questions across 14 categories, each featuring multi-turn coreference safety attacks. We then conducted detailed evaluations on five widely used open-source LLMs. The results indicated that under multi-turn coreference safety attacks, the highest attack success rate was 56% with the LLaMA2-Chat-7b model, while the lowest was 13.9% with the Mistral-7B-Instruct model. These findings highlight the safety vulnerabilities in LLMs during dialogue coreference interactions.
[ 0.0425913073, 0.0203616433, 0.0000658428, 0.0682321936, -0.0604173392, -0.0240807254, 0.0278404988, -0.0208182037, 0.0473102406, -0.0556697622, -0.037393596, -0.0043935729, -0.0089428304, 0.0283086095, 0.0452695861, 0.0037747344, 0.0826611072, -0.0772488192, -0.0167935155, 0.0021311885, 0.0198444091, -0.0175725278, -0.0479325019, 0.0565008558, -0.0060564512, -0.0276388042, -0.0427769087, 0.0224389341, 0.0161089562, -0.0151159652, -0.0089931488, 0.0415951423, -0.0148037309, 0.0640776232, 0.0000020777, -0.0768809617, -0.0361884944, 0.0325749442, -0.0549935438, 0.0188633706, -0.0089934943, 0.0306453425, -0.0286506638, 0.0469034277, -0.0076724091, 0.0306208804, 0.0283772945, 0.0658689961, 0.1071688682, 0.0671061277, -0.000545509, 0.0166334622, 0.035873577, -0.0199530683, -0.0517949648, -0.0802627876, 0.0171649735, -0.0053163739, 0.0072124237, 0.0486062914, -0.0078540724, -0.0053293626, 0.0174732786, -0.037229348, -0.0763223544, 0.0031660644, -0.0836030096, -0.0264532771, -0.0154823065, 0.0703656673, -0.0081238225, 0.0298666824, 0.0007348029, 0.020319514, -0.0029055681, 0.0038275768, -0.0145240361, 0.0096900035, -0.0221232548, -0.0447762907, 0.0035552713, -0.0182352811, 0.0051629534, -0.0051640011, -0.0043173935, 0.0743569657, 0.0232753512, 0.0034674052, 0.0190351866, -0.0479950607, 0.0021506958, -0.0647435784, -0.0129339229, 0.0003069236, -0.063357383, 0.023565501, -0.0455989949, 0.0333523229, 0.0654324815, -0.0682930648, 0.0120032299, 0.0390785448, -0.0219151825, 0.0192504656, 0.0242549051, 0.0310969595, -0.0130245173, 0.0286903549, 0.0261739474, -0.0072046705, -0.0023727408, 0.0080404459, -0.0096110087, 0.0294865835, -0.0091386475, 0.0052813063, -0.0381892323, 0.0586608611, 0.0084414827, 0.0668217167, 0.0084951418, -0.0207193233, -0.0071198377, -0.0018901505, 0.0207420774, -0.0001062259, 0.015614111, -0.0015570384, 0.0032757334, -0.1201345101, -0.0193928294, 0.0292229783, -0.0482582673, -0.0386570692, -0.0414880402, 0.1207010597, 0.0150955878, -0.0210749302, -0.0127168437, -0.1132553443, -0.0360564291, -0.0170962699, -0.0118295839, -0.0605115108, -0.0024060633, -0.0030249485, -0.0114255929, -0.0358560495, 0.0379371457, 0.0190221034, -0.0316825509, 0.042716518, 0.0048151221, -0.0134747168, 0.0057613598, 0.0547861382, -0.0425330587, 0.065981999, 0.0054472904, 0.0175758284, 0.0554991737, -0.01019355, 0.0319720805, 0.0684650764, -0.0113517009, 0.02698089, 0.0309697855, 0.0496021807, 0.0152353877, -0.0262344442, -0.0183837675, 0.0816227645, -0.0050565032, -0.0620995536, 0.101676546, 0.060649395, -0.039721366, -0.0032664824, -0.012916049, 0.0001021337, 0.0078832563, 0.0104555804, -0.0519065335, 0.004398372, -0.0678570122, -0.0439435318, 0.0251799244, 0.0198341999, -0.0194405653, 0.0066357995, -0.0128754303, 0.0064720092, -0.017686639, 0.001070661, -0.0031568529, 0.0021788138, -0.0198543742, 0.0486197807, 0.0011516864, -0.047789339, -0.0534198508, -0.0498288125, -0.0435872115, 0.0260906182, 0.0654532239, 0.0058047776, -0.0912823603, -0.0559950471, 0.0103097884, 0.0269738343, 0.0395541377, 0.0189983882, 0.0251749959, -0.0171801411, -0.021730952, 0.0166144185, -0.001906656, 0.0014434276, 0.0079088314, 0.0026939816, -0.0087592285, 0.0332623497, -0.0134519264, 0.0049713645, -0.0014522927, -0.0261010695, -0.0066538528, 0.0198327769, 0.0390417278, 0.0066552558, 0.0042177001, 0.0094907098, 0.0457467251, 0.0318689458, -0.0051598288, 0.0286036562, -0.0430312864, -0.0126209548, 0.0267589781, -0.0064209546, -0.0085415263, 0.0040900805, -0.021240091, -0.0046981084, 0.0354493447, -0.051769387, 0.0240034927, -0.0069663329, 0.0207812302, 0.0129483696, 0.0050030109, -0.0209843554, -0.0133890873, -0.0292993374, -0.0051141744, -0.0113557912, -0.0043680705, 0.0185001399, -0.0476352684, -0.0288103558, 0.0283421911, -0.0335269235, 0.0236416832, -0.0053064926, 0.0145154791, -0.0498575494, -0.0151243219, -0.0572470613, -0.0149638327, -0.0161766782, 0.0419564806, 0.0177865848, 0.0301569253, -0.0178368744, -0.0066626635, -0.0024709492, -0.0088461377, 0.0117417984, 0.0337425955, 0.0084089153, -0.0011119262, -0.0742120072, -0.0310946163, 0.0133341011, -0.0169716477, 0.0481778122, -0.0246970374, -0.0083467839, -0.0338698141, 0.0324283727, -0.0094825486, 0.0301827788, -0.0226505678, -0.0399714969, 0.0317327194, 0.0149714677, 0.0172261372, 0.0337609798, 0.0233636722, 0.0340200253, 0.0283930376, -0.0527367331, -0.0638106838, 0.0009805236, 0.0369489789, -0.0859868601, -0.0360298902, 0.0070803985, 0.0226915125, 0.0507003181, 0.0150330551, -0.0034943647, 0.0067388727, -0.0241859797, -0.0135005098, -0.0306000877, -0.0002494855, -0.0505548976, -0.0391330495, 0.0309582688, 0.0049834736, -0.0474271439, 0.0452408902, -0.0577011146, 0.0076394654, -0.0327084661, -0.0399142168, -0.0100494195, 0.051144205, -0.0539496914, 0.005185517, 0.0753074065, -0.0599066988, -0.0460051522, -0.0474723168, 0.036079362, -0.0386893339, 0.0508873984, -0.0554201566, -0.0034702376, -0.0183475316, -0.0181685518, -0.077259995, -0.0277892072, -0.0398028418, -0.0155236833, -0.0056131827, -0.0110196173, -0.0161536671, -0.004278596, 0.0258382503, -0.018358564, -0.0084632179, -0.0028380719, -0.0230705328, 0.0271067433, 0.030131828, -0.0469541401, -0.0016170074, -0.0188174639, 0.0697158054, 0.1089668348, 0.032724537, -0.0171087179, -0.0092307348, -0.0449327789, 0.0429041572, -0.0328151882, -0.0167510677, 0.0274434686, 0.0607725121, -0.0444099344, -0.0318725146, 0.0786883906, 0.0299675371, -0.0254694056, -0.0007136902, -0.0574380681, 0.010528137, 0.0040660514, 0.0062447763, -0.0559000708, -0.020018829, 0.0305990707, 0.0340373628, 0.0051047737, -0.0285504144, -0.0136320461, -0.1095084548, 0.023945339, -0.0089614429, -0.0436734855, -0.0086646248, -0.0163391009, 0.0166279711, 0.0239866674, 0.0787612274, 0.0062144287, -0.0012064909, -0.0166817289, 0.0126598524, -0.0576187633, -0.0071070623, 0.0313429236, -0.0491914004, 0.0506070629, 0.0698753297, -0.0007142385, -0.0134145133, 0.0020067645, 0.0722547472, 0.0011447041, -0.0155098001, -0.0068191364, 0.0276699346, 0.0052637099, -0.0077013373, 0.0155006796, 0.0257063042, 0.0163591485, -0.0061454326, 0.062510483, 0.0163079146, -0.052305501, 0.0271456279, -0.021150643, 0.0333474837, 0.0416111015, 0.0562366135, 0.0342543833, -0.0271813162, 0.0192627292, -0.0555758253, 0.0072811572, -0.0179565027, -0.0584236979, -0.0082600173, 0.0394764878, -0.0153671559, -0.0596318208, -0.009442905, -0.004150535, 0.017395556, 0.0134514505, -0.0056628063, 0.0024730163, 0.040409822, 0.0102417273, -0.0288816504, -0.0707923397, 0.0532021709, -0.0902446657, 0.0804683864, 0.0341530442, 0.025146015, -0.0057414863, 0.0664648935, -0.0465597883, -0.081236124, 0.0021076242, 0.0613991357, 0.0265851691, -0.0076227635, -0.0304375906, -0.0417037681, 0.0128274374, 0.0168668218, -0.0527098998, 0.0030380283, 0.0653131679, 0.0208004918, 0.0148065761, 0.0283003822, -0.0392127931, -0.0289680976, 0.0102670966, 0.0138733247, -0.0556704663, -0.1173601151, -0.0308425892, 0.0523366816, 0.0456532352, 0.0209492892, 0.074765712, 0.0222829636, 0.0530362017, -0.045587685, -0.0298544336, 0.0058515472, 0.0503136329, -0.0323597752, -0.0320547894, 0.0497497767, 0.0036922605, -0.0443091281, 0.0542087555, -0.0001660332, 0.0330533795, 0.0347468629, 0.0283291955, 0.01847215, 0.0003864594, -0.0120569495, 0.0256938487, 0.0329408273, 0.0374242924, 0.0125966379, 0.0717416033, -0.029979242, -0.0179179031, 0.0210715402, 0.0470430516, -0.0520959869, -0.0748482719, 0.0422763564, 0.0248251595, 0.0266438592, 0.0435738675, 0.0393377244, 0.0002385185, 0.0132349087, -0.0344689749, 0.0102430647, -0.0135939438, -0.044352971, 0.0336850435, 0.0087821921, -0.0029714429, 0.0366603658, 0.077657491, -0.0039278795, 0.0281746611, 0.0161124766, -0.0275108833, -0.0524480157, -0.0328268632, 0.0412500985, 0.0186010133, -0.0395597629, -0.0489517413, -0.0781446323, -0.0643326566, 0.0422347337, -0.0129881436, 0.0193478018, 0.0350891054, 0.0067590671, 0.0249080621, 0.0369892977, 0.0064706206, -0.0556530058, 0.0032174003, 0.0112884818, -0.0444383174, 0.0125417868, -6.08479957e-33, -0.0007972004, -0.0423662104, -0.003334316, 0.0149205169, -0.0102679534, 0.0084670233, -0.0235803779, 0.0329720601, -0.0484628305, -0.0172138661, -0.0096040051, 0.0134578599, 0.0237086825, 0.0122129871, -0.0184106994, -0.0259463005, -0.0044279774, -0.0011842657, -0.0069147456, 0.0352462046, 0.063971214, -0.0191761423, 0.0197948553, -0.1256084144, 0.0486930236, 0.0089858118, -0.0307885129, 0.0277841464, -0.0439356379, 0.0412535444, -0.0028561442, 0.0352229849, -0.0220633484, -0.0462759249, 0.0093448404, 0.0250726938, 0.0003588962, -0.0391834341, 0.0123692323, -0.0062984056, -0.001991672, -0.0410876721, -0.006103449, -0.0083639063, 0.007232442, -0.0169948488, -0.0143270008, 0.0145372069, -0.0084934141, -0.0035764838, -0.0118904468, 0.0135270339, -0.0280911345, -0.0387599058, -0.0237802118, 0.0542271659, 0.0407515168, -0.0445928238, -0.0346957929, 0.0433941893, -0.0379989743, -0.0019846417, 0.0583897382, 0.008701154, -0.004587864, 0.0432808585, -0.030383328, 0.0095731551, 0.0068756221, -0.0022134264, -0.0238602255, 0.0657630414, 0.0236724708, 0.0754872486, -0.053956639, -0.0064740586, -0.0761276633, 0.029443251, 0.0294222236, 0.0688368604, 0.0239864103, -0.0088600433, -0.0817769915, -0.0059795198, -0.0224751234, 0.0005473988, -0.0346440412, -0.0335213952, 0.0204038508, -0.0116145611, -0.0099412519, 0.0133440951, 0.0469822437, -0.0202373471, 0.0022542863, 0.0121934768, 0.0050188857, -0.0261735227, -0.0182223767, 0.0290661491, 0.0028211277, 0.0213097427, 0.021397965, -0.0496325828, 0.0299001187, -0.0106461309, 0.0038154649, 0.0284152813, -0.0754739493, -0.0204143431, -0.0160289556, 0.0469546951, -0.0213676281, 0.016688684, 0.0309440419, -0.0234693177, -0.0106679965, 0.0379475802, 0.0408072956, -0.0140695125, -0.0536986813, 0.0033824355, -0.0179364774, 0.011205717, -0.0073811091, 0.0100960825, 0.0145674739, 0.0058138473, 0.0310299192, -0.0213917848, -0.0076067839, -0.0602001771, 2.857e-7, 0.0137084704, 0.0267162099, 0.0299259908, 0.03570823, 0.0208392441, 0.0710989833, -0.0078930603, 0.0603781268, 0.0385156199, -0.0410165302, -0.0220420491, -0.0018595051, 0.0355024636, 0.0620078631, -0.0569007285, -0.0400716923, -0.0584899597, -0.0008128046, -0.0468982682, -0.0325802714, 0.0596086122, 0.0672671422, 0.009871413, -0.0262228232, -0.032522358, -0.0288416501, 0.0010671709, -0.015520989, -0.0368277766, -0.0000135124, 0.0348379873, 0.0263016131, 0.0159439556, 0.0463612862, -0.0066345059, -0.0180696584, -0.0163077209, 0.0603494681, 0.0197347067, -0.0546133369, -0.0242815297, -0.0562540703, 0.007535744, -0.0361934006, 0.0556757674, -0.0294871032, -0.0098910239, -0.0142495958, -0.0307367705, -0.0203436892, 0.031251736, -0.0308909621, -0.0075999419, 0.0367669463, -0.010559462, 0.0121995928, 0.0044860509, -0.0321351103, -0.0331925601, -0.0398122557, -0.0010354011, -0.0732648745, 0.0264973398, -0.0030374622, -0.0519277304, -0.0104679251, 0.0129913697, 2.647622864e-34, -0.0128077734, -0.011917307, 0.0085001029, 0.0266995318, 0.046567522, 0.0086420914, 0.0046564471, 0.0105221607, 0.0135382144, -0.0110995788, -0.0394475162 ]
http://arxiv.org/abs/2406.17624v1
http://arxiv.org/pdf/2406.17624v1
Self-assessment, Exhibition, and Recognition: a Review of Personality in Large Language Models
2024-06-25T15:08:44
2024-06-25T15:08:44
Zhiyuan Wen, Yu Yang, Jiannong Cao, Haoming Sun, Ruosong Yang, Shuaiqi Liu
cs.CL
cs.CL, cs.AI
As large language models (LLMs) appear to behave increasingly human-like in text-based interactions, more and more researchers become interested in investigating personality in LLMs. However, the diversity of psychological personality research and the rapid development of LLMs have led to a broad yet fragmented landscape of studies in this interdisciplinary field. Extensive studies across different research focuses, different personality psychometrics, and different LLMs make it challenging to have a holistic overview and further pose difficulties in applying findings to real-world applications. In this paper, we present a comprehensive review by categorizing current studies into three research problems: self-assessment, exhibition, and recognition, based on the intrinsic characteristics and external manifestations of personality in LLMs. For each problem, we provide a thorough analysis and conduct in-depth comparisons of their corresponding solutions. Besides, we summarize research findings and open challenges from current studies and further discuss their underlying causes. We also collect extensive publicly available resources to facilitate interested researchers and developers. Lastly, we discuss the potential future research directions and application scenarios. Our paper is the first comprehensive survey of up-to-date literature on personality in LLMs. By presenting a clear taxonomy, in-depth analysis, promising future directions, and extensive resource collections, we aim to provide a better understanding and facilitate further advancements in this emerging field.
Self-assessment, Exhibition, and Recognition: a Review of Personality in Large Language Models: As large language models (LLMs) appear to behave increasingly human-like in text-based interactions, more and more researchers become interested in investigating personality in LLMs. However, the diversity of psychological personality research and the rapid development of LLMs have led to a broad yet fragmented landscape of studies in this interdisciplinary field. Extensive studies across different research focuses, different personality psychometrics, and different LLMs make it challenging to have a holistic overview and further pose difficulties in applying findings to real-world applications. In this paper, we present a comprehensive review by categorizing current studies into three research problems: self-assessment, exhibition, and recognition, based on the intrinsic characteristics and external manifestations of personality in LLMs. For each problem, we provide a thorough analysis and conduct in-depth comparisons of their corresponding solutions. Besides, we summarize research findings and open challenges from current studies and further discuss their underlying causes. We also collect extensive publicly available resources to facilitate interested researchers and developers. Lastly, we discuss the potential future research directions and application scenarios. Our paper is the first comprehensive survey of up-to-date literature on personality in LLMs. By presenting a clear taxonomy, in-depth analysis, promising future directions, and extensive resource collections, we aim to provide a better understanding and facilitate further advancements in this emerging field.
[ 0.0759177878, 0.0330460258, -0.0718101189, 0.0355531946, -0.0304182712, 0.0176601261, 0.026431324, -0.0057423725, 0.0387227871, -0.0131422896, -0.0748141706, -0.0093603227, -0.0012417767, 0.0085774409, 0.0266640149, -0.0721409395, 0.0270870104, -0.0183594357, 0.0398286022, -0.0104236351, 0.0295521077, 0.051314082, 0.0218071975, 0.0314039998, 0.0057136156, -0.0487742908, 0.0285249986, -0.0279498268, 0.0204859376, -0.0319156758, 0.0139395688, 0.0785391107, 0.004117189, 0.022744365, 0.0000022138, -0.0288038328, -0.0006149836, -0.0423504226, -0.00910962, 0.0180808231, 0.079066731, 0.0404632837, 0.0185714625, 0.0245595183, -0.0690362528, -0.0393895134, 0.0070653679, 0.0340218917, 0.0547429062, 0.0015506571, 0.0053565903, -0.0056035086, -0.0123168146, 0.0169035606, -0.0059883446, 0.003653045, 0.0114724366, -0.0237401277, -0.0168426596, 0.0187668093, 0.0284461137, 0.0167230871, -0.0024460603, 0.0020207709, -0.0000512081, -0.00611478, -0.0259010028, -0.0218705628, -0.0113310777, 0.0462103449, 0.0691910014, -0.0238639843, -0.0178754907, -0.0087938942, 0.0240627415, 0.0359670594, -0.0083341207, 0.0326458849, 0.014571798, -0.0412210636, 0.0059785079, 0.0473767631, -0.018273063, 0.0157067254, -0.0053756353, 0.0146926213, -0.0098904269, -0.0094954027, -0.017682327, -0.0064463289, 0.0039100512, -0.0892639682, 0.012178787, 0.0342989005, -0.0060840733, 0.0261734929, -0.0385022201, -0.0301657952, 0.0314617455, -0.0737307668, 0.0069752452, 0.0143698696, -0.016395377, -0.0002530128, 0.0560540743, 0.0537739471, 0.0294847414, 0.0367746763, 0.0015941139, 0.0033474925, -0.0501834415, -0.0142915072, 0.0208244734, 0.0044492781, 0.0584483519, -0.0064234878, -0.054552827, 0.0343944244, -0.0231668688, -0.0086488845, -0.0176050588, -0.0173667874, 0.0192656703, 0.0183845982, -0.0560007989, 0.0013780775, -0.0517489426, -0.0330216102, -0.0235073213, -0.050364498, 0.0624199323, -0.006050711, 0.0080670249, -0.0193069018, 0.030488288, 0.0608531833, 0.0316126123, -0.0704910383, -0.0192000866, -0.0504026935, -0.0302929562, -0.070495978, -0.0233980902, -0.0396170132, 0.0182741284, 0.0354502276, 0.0050891722, -0.0513488017, -0.0056462721, 0.0278415382, 0.0256476309, 0.0419115722, -0.0076203542, 0.0306484923, 0.0199032426, 0.0397074372, 0.024399044, 0.0177388098, 0.0098845046, 0.0065381406, 0.0426952094, -0.0419534445, 0.0350283682, 0.0395352878, -0.0298143234, -0.0264288019, -0.0123031391, 0.0045695207, 0.024987461, -0.0012486756, 0.012004545, 0.0552266613, -0.0237576533, -0.0082447948, 0.1075075567, 0.071429722, 0.0018008457, 0.0377761163, -0.0227766279, 0.0547190197, 0.0543199591, -0.0088617243, -0.0153462719, 0.0355530158, 0.0103709456, -0.0136028696, -0.0015128363, 0.0204018243, 0.0000107675, -0.0524732471, -0.0315233171, -0.0003523058, 0.0289275516, 0.0533196628, -0.0232526623, -0.0042654984, 0.0102626858, -0.0234744661, 0.0349378288, -0.0551802181, -0.0115356799, -0.0628599375, 0.010010873, 0.1259111762, 0.0350792855, 0.0083640767, 0.0079494715, -0.0233882144, -0.031885922, 0.0493974574, 0.035686411, -0.0219357349, -0.0227446929, -0.017092986, 0.0344242863, -0.0056619272, -0.0094552245, 0.0189017635, 0.0090916436, 0.0331546962, -0.0258441586, 0.0214653295, -0.0207476169, -0.0106319338, -0.0680665225, -0.0624996088, -0.0080946973, 0.0365652218, 0.0161306337, 0.0343863592, 0.0056585912, -0.0184623245, -0.0069693453, 0.0265154839, -0.0004235948, 0.0346547477, -0.0048514297, 0.0157279689, 0.0084147556, 0.0074668904, -0.0422702581, 0.0079029221, 0.0567565188, 0.0118173519, -0.0054300525, -0.0366003066, 0.0459735133, -0.0147658968, 0.0129234903, -0.0703433752, 0.0629007891, 0.0000744514, 0.0351250395, -0.0015443483, 0.0146105951, 0.0405232012, -0.0212442856, -0.0311382953, -0.0469099432, -0.0396695137, -0.0352829285, 0.004784958, 0.0254069623, 0.0067013563, -0.0109302597, -0.0503708348, -0.0454659872, -0.0922911689, -0.0293484833, 0.0195074826, 0.0136794373, 0.0437664241, -0.0249679442, -0.0520597473, 0.0072942623, -0.0334644765, -0.0203919336, 0.0234193075, 0.0233875476, 0.0352851786, -0.0229045264, 0.0004751935, -0.0518292375, 0.0034628019, -0.0333636887, 0.0225681644, -0.0049702851, 0.0298695192, -0.0189771205, 0.0119184442, -0.0550454669, -0.0263664592, -0.0067057661, -0.0499168932, 0.0393485576, -0.0015934006, 0.029985154, 0.044900503, 0.0543893538, 0.0518251806, 0.0289133228, -0.0350099802, -0.0451075621, -0.026439514, 0.0034508777, 0.0298539307, -0.0056745992, -0.0966644362, -0.0041640047, 0.0120301265, 0.0116399648, 0.0361766629, -0.0283899326, 0.04949544, 0.0007706881, -0.0575339124, 0.0223608427, 0.0154515086, -0.0124539305, 0.0490827858, 0.0339898057, -0.0367217213, 0.0675095096, -0.0189440381, 0.0115673579, 0.0094280476, -0.0533865504, -0.0254307743, 0.0917823985, -0.0090406444, -0.00791586, 0.0418510735, -0.0187068544, -0.0267899167, -0.0533489808, 0.008315444, -0.0375683345, 0.0152813597, 0.0022809061, -0.0573326983, 0.0595069975, -0.0188085064, -0.0317503139, -0.0702359453, -0.039155785, -0.0400959887, 0.0356241018, -0.0212106388, -0.0264483914, -0.0031900357, 0.0416152887, -0.0483528636, -0.0167456642, -0.0171299223, -0.0030797245, -0.0130936988, 0.049042128, 0.0248087607, 0.0494939834, -0.0557055622, -0.0010963021, 0.0178751871, -0.0218206663, -0.0089831101, 0.0084323734, -0.00420313, 0.0434590876, -0.0014229275, -0.0199315362, 0.0206201989, 0.041130282, -0.0250310823, 0.0165750962, 0.1282985061, 0.0323430561, -0.0187853128, 0.0106544429, -0.0194855519, 0.0182013288, 0.0144162541, 0.042970214, -0.0044779046, 0.0013587122, -0.0324849971, -0.0003066144, -0.064940922, -0.0185990501, -0.0225413069, -0.0651799962, 0.013353399, 0.0112303179, -0.0330564566, 0.032748878, -0.0059730965, 0.0185053162, 0.0127011752, 0.0120129948, -0.0402280688, -0.032055486, -0.0215745568, -0.017336715, -0.0691246837, -0.0246479958, 0.0760875195, -0.0446058512, 0.0340683274, -0.006232576, 0.0393307395, 0.0073596276, -0.0563931875, -0.0792280883, 0.0182137843, 0.0234938562, 0.0242428482, 0.0447036065, -0.0508593135, 0.0305419359, -0.1125479117, -0.0230846871, 0.0041451515, 0.003298477, 0.0225576777, 0.0412845761, -0.0027206219, -0.033027593, -0.0247973632, 0.0035738158, 0.0340228714, 0.0359984487, -0.0009262016, -0.0935707539, -0.0012738849, 0.0052295881, 0.0189162884, 0.0237503741, -0.0909099206, -0.0231091585, 0.0438409448, -0.0144966356, -0.0348953865, -0.1043621153, -0.0158024877, 0.0081414906, 0.0055354116, -0.0032878958, -0.0170034636, 0.0109858681, -0.0197936464, 0.0097768968, -0.0381841995, 0.0499187186, -0.0298260953, -0.0286232848, 0.0320190862, -0.0579087548, 0.023823332, 0.0152075812, -0.0369658917, -0.0312305298, 0.0193911362, 0.008822741, 0.0171673577, -0.010139836, 0.0407968685, -0.0869734585, 0.0273983348, -0.0063203806, -0.0282762088, 0.0534919687, -0.0251509864, 0.022328997, -0.0117639238, 0.0169741642, 0.0104418965, 0.0199928321, -0.0114726136, -0.0405936651, -0.0202957895, -0.0974629149, -0.0029802273, 0.0175762568, -0.0043704538, 0.0007387624, -0.0273025315, 0.0018792683, -0.0429477319, -0.0391462818, -0.0243470315, 0.0303831138, -0.0790168047, -0.0186090823, 0.0097062346, -0.0223314017, -0.0171862897, -0.0286184605, 0.0210944023, -0.0353632532, 0.0305286255, 0.0059703044, 0.0077300384, -0.0269975513, 0.0222582836, -0.0116293021, -0.0046184617, 0.0062730559, 0.034691602, 0.0072404989, 0.0869661048, -0.0331856608, -0.024408631, -0.008429681, 0.0568143986, -0.0319724195, -0.0304306075, 0.0324118398, 0.0100893946, -0.0270134732, 0.0007322484, 0.0560632087, -0.0414315835, 0.0154801365, -0.0305306669, 0.0160961617, -0.0252120476, -0.0548771881, 0.0167319104, 0.0050187139, -0.0101411389, -0.0250324868, 0.007234381, 0.0190493129, -0.010391999, -0.0240606293, -0.0433254689, -0.0358861387, 0.0248264838, 0.0192674976, 0.0179630797, -0.0254967213, -0.030435957, -0.1035016775, -0.0935059786, 0.0509764068, 0.0237858798, 0.0050719227, -0.0575234406, 0.0193374753, 0.0514884144, -0.0069177034, -0.0259602834, -0.0481461994, 0.0228220392, -0.04057936, -0.0052389237, -0.0159261972, -5.893426519e-33, -0.0147955138, -0.0607268065, 0.0285705701, -0.010449525, -0.0095076887, -0.0544572808, -0.0177438892, 0.0509663559, -0.0086771194, -0.027918661, -0.0382041, 0.0024715266, 0.0091115283, -0.0372461863, 0.0095678456, 0.0120402928, 0.0484024957, -0.0108588785, -0.0294948928, 0.0220789984, 0.0201159436, -0.030784253, 0.0439225435, -0.0792063251, 0.0560141318, -0.0730682388, 0.0052409377, -0.0284717288, 0.0333616696, 0.0753232986, -0.0205031335, 0.0452043191, 0.0305164568, 0.0150826694, -0.0237797517, 0.0768360496, -0.0389436968, -0.0461664349, 0.0078109102, 0.0456002764, -0.0043237945, -0.043691542, 0.0139593128, -0.0210683215, -0.0832714438, -0.0091675008, 0.0555435829, 0.0161278825, 0.0504891351, 0.0772153884, -0.0186201688, 0.0073440918, -0.0115308315, -0.0120982938, 0.009763225, 0.015498017, -0.0159085579, 0.0018833865, -0.0837999955, 0.0718208253, -0.0418598242, -0.0254018232, 0.0500697382, 0.0275144409, 0.0125506315, 0.0553311035, 0.0616321936, 0.0921938047, 0.022397358, -0.0301498212, 0.0614302792, 0.0260848645, 0.0261727534, 0.0782231018, -0.0198466964, -0.0384630039, 0.0296549201, 0.0040039015, 0.0741862357, -0.0608596951, 0.0598563589, 0.0006029178, -0.0604307391, -0.0376713164, -0.0070160269, 0.0050947876, -0.0321407169, -0.0330171622, 0.024391856, -0.0358010419, -0.0231370721, -0.0226317085, 0.0433049537, -0.0590523146, -0.0713657588, 0.0167448632, -0.059879411, 0.0666495413, 0.0003820082, -0.0071742376, -0.0177236143, 0.0074946447, -0.0115337716, -0.0003419053, 0.0436213911, 0.0325408205, -0.0167086292, 0.0260524694, -0.0569669195, -0.0069287037, 0.0390769728, 0.009522208, -0.0098857898, -0.0061859591, 0.0368015058, 0.0003587898, -0.0100790253, -0.000434155, 0.0031849374, 0.0010179055, -0.0556384698, -0.0008408843, -0.0133392345, -0.0182833467, 0.0029625678, 0.0007847068, -0.0719532445, 0.018998038, 0.0497610047, -0.0221004207, 0.0314232744, -0.0143410424, 2.778e-7, -0.0572994798, 0.0659524426, -0.0421139151, 0.0446595326, 0.0344117917, 0.0584162101, 0.0126460968, 0.0231238995, 0.0572682396, -0.0343160219, -0.016574882, 0.0244178958, 0.049092304, 0.0508197732, -0.0466485843, 0.0309789274, -0.0196097307, 0.0493255593, -0.0301303864, 0.0223955363, 0.0591326393, 0.0207346808, 0.0407492816, -0.000966927, -0.0311049782, 0.0101505443, 0.0353515223, -0.0134169599, -0.0075552985, 0.0255990643, -0.0176687371, -0.0055124778, 0.0099088401, 0.0575343035, 0.0060715592, 0.0064176372, -0.0137279462, 0.0484123193, -0.0039263302, 0.0300840568, 0.0287082382, 0.0155053297, -0.02056369, -0.0454822704, 0.029804958, -0.0260196459, -0.0452951938, -0.0274553224, 0.0200113021, -0.0654494315, 0.046186883, -0.0014033617, -0.0402355492, 0.0475113764, 0.0223613214, -0.0457520373, -0.0076142382, -0.0457869358, 0.0257546082, -0.0402298942, -0.0282055568, -0.0494446717, -0.0425454192, 0.0162971802, 0.0235253312, -0.0564121008, 0.0427341275, 2.261985871e-34, -0.0333797224, 0.0402150154, -0.0128676072, -0.0083437031, 0.0600803159, 0.0323539339, 0.0172983669, 0.0186676644, 0.0569714159, -0.0681179017, -0.0068110479 ]
http://arxiv.org/abs/2406.17618v1
http://arxiv.org/pdf/2406.17618v1
Towards Building an End-to-End Multilingual Automatic Lyrics Transcription Model
2024-06-25T15:02:32
2024-06-25T15:02:32
Jiawen Huang, Emmanouil Benetos
eess.AS
eess.AS, cs.CL, cs.SD
Multilingual automatic lyrics transcription (ALT) is a challenging task due to the limited availability of labelled data and the challenges introduced by singing, compared to multilingual automatic speech recognition. Although some multilingual singing datasets have been released recently, English continues to dominate these collections. Multilingual ALT remains underexplored due to the scale of data and annotation quality. In this paper, we aim to create a multilingual ALT system with available datasets. Inspired by architectures that have been proven effective for English ALT, we adapt these techniques to the multilingual scenario by expanding the target vocabulary set. We then evaluate the performance of the multilingual model in comparison to its monolingual counterparts. Additionally, we explore various conditioning methods to incorporate language information into the model. We apply analysis by language and combine it with the language classification performance. Our findings reveal that the multilingual model performs consistently better than the monolingual models trained on the language subsets. Furthermore, we demonstrate that incorporating language information significantly enhances performance.
Towards Building an End-to-End Multilingual Automatic Lyrics Transcription Model: Multilingual automatic lyrics transcription (ALT) is a challenging task due to the limited availability of labelled data and the challenges introduced by singing, compared to multilingual automatic speech recognition. Although some multilingual singing datasets have been released recently, English continues to dominate these collections. Multilingual ALT remains underexplored due to the scale of data and annotation quality. In this paper, we aim to create a multilingual ALT system with available datasets. Inspired by architectures that have been proven effective for English ALT, we adapt these techniques to the multilingual scenario by expanding the target vocabulary set. We then evaluate the performance of the multilingual model in comparison to its monolingual counterparts. Additionally, we explore various conditioning methods to incorporate language information into the model. We apply analysis by language and combine it with the language classification performance. Our findings reveal that the multilingual model performs consistently better than the monolingual models trained on the language subsets. Furthermore, we demonstrate that incorporating language information significantly enhances performance.
[ 0.0027447152, 0.0500481501, -0.0136723453, 0.0649301186, -0.0281420033, -0.0051537356, 0.0532878451, -0.0430490747, -0.0271094199, 0.045981314, -0.0843025893, -0.0440554321, 0.0236617234, 0.0006069824, 0.0253492016, -0.0176939126, 0.0101845535, -0.0620078184, 0.0408104472, -0.0292242356, -0.0321289003, 0.0245579928, -0.0027712074, 0.0344678201, -0.0345545709, 0.0272170845, 0.0029505165, 0.0212086886, -0.0092170732, -0.0103973774, -0.0419213548, 0.0167883541, -0.0021691355, 0.0206095669, 0.0000017302, -0.0476323403, -0.0246505905, 0.0149333645, -0.0053120623, -0.0502613783, 0.0241217054, -0.0164062344, -0.0227308646, 0.0141939279, -0.0005586116, -0.0456005745, 0.0691910535, -0.0049958341, 0.0515368171, 0.0743800253, 0.0070573865, -0.0669000521, 0.0221720096, -0.0555575006, -0.0127462484, -0.0533339903, -0.0038944371, 0.0129203806, 0.0211227369, -0.0382047072, -0.0096067097, 0.0965586454, -0.0046368935, 0.0099371932, -0.0094244713, 0.0174302384, -0.073251009, -0.0549620241, 0.0281245112, 0.0121313231, -0.000945947, 0.003147451, -0.0110992696, 0.0121083083, -0.0046423599, -0.0281531643, -0.0151271373, 0.0054097171, 0.0245399196, 0.0079324013, 0.0315563269, 0.0141575998, 0.0498138368, -0.0218864791, 0.0031107746, 0.0056079682, 0.0305709857, -0.0607932806, 0.0466615185, 0.0221025702, 0.0386314131, -0.0387113728, 0.0307590533, -0.0229890551, -0.0082518114, 0.0035655205, -0.032094229, -0.0563980974, -0.0147158317, -0.0549179167, 0.0560922287, 0.0228184983, -0.0366083309, 0.0635801554, -0.028648071, 0.0139567452, -0.0222801343, 0.005463358, -0.039513085, -0.0560379289, 0.0071903793, -0.0084548378, -0.0506129526, -0.0393142216, 0.0156454891, -0.0005741414, -0.033444088, -0.0247655436, -0.008455785, 0.0297378879, 0.0190184209, -0.0660557747, -0.0355095491, 0.0028587419, 0.0637427345, 0.0198544022, 0.0072748512, 0.0316058584, -0.0221177936, -0.0351988003, 0.0118853636, 0.0110457232, 0.0315887481, 0.0045050061, 0.0749029592, 0.0762347579, 0.0513221398, -0.0554136746, -0.0299939942, 0.013911563, -0.052964516, -0.0667829365, -0.0017375586, -0.0925994441, 0.0972236395, -0.0208257232, 0.0071065086, -0.0173687637, 0.0032576784, -0.0284016188, -0.011336985, 0.060491886, 0.0177216586, 0.0320592485, 0.0080760829, -0.0192399435, -0.0156713631, 0.0531441271, 0.074324958, 0.0007729947, 0.0713888332, -0.043470405, 0.0179555323, 0.0988670364, -0.0008665696, 0.0111911371, -0.0464926921, -0.0174423791, 0.0085325334, 0.0322111994, 0.0202129874, 0.0276586022, 0.0320429206, -0.0328691229, 0.0214618444, 0.0194801819, 0.0179226901, 0.059472166, 0.0323370025, -0.0130302394, 0.0100495331, -0.0114493025, 0.0196155962, 0.0428426154, 0.0666599348, 0.0256463606, 0.0860086232, 0.0436681882, -0.0176511407, 0.0185154304, -0.025160199, -0.0274396557, -0.0168595966, 0.0355429947, 0.0607739612, 0.0103176227, -0.0149668818, 0.0151932314, 0.0290874243, -0.0658993647, 0.0125194276, -0.0254653599, 0.001265162, 0.0517739616, 0.031024063, -0.0060546915, 0.0351341404, -0.0190372989, -0.0284990892, 0.0507145934, -0.0045235036, -0.0304643437, 0.0561281256, -0.0160430204, 0.0070918305, 0.0113809248, -0.0122238575, 0.0466068871, -0.0301990658, 0.0312870853, -0.0263691805, -0.0265298467, -0.0244379938, -0.0267719124, -0.0224878509, -0.0009617429, 0.0348495692, 0.0406450033, 0.0546562485, 0.0605596863, -0.0621772707, 0.0177090392, 0.0037490579, -0.0404345095, -0.0000256392, 0.0249179155, -0.0213836916, -0.063358061, 0.0525980853, -0.0215277281, -0.004048279, 0.0026918463, 0.0161671657, -0.0728992, 0.0602188967, -0.0119616417, 0.0411975533, -0.0060148453, 0.0109221647, -0.0286250748, -0.012606184, 0.0420214236, -0.0273047592, 0.0345393047, 0.0270971581, -0.0218495298, -0.0622369461, 0.0576580763, -0.0287939999, -0.0470744185, 0.0780901909, 0.0202527214, 0.0085922768, -0.0001044802, -0.0047678617, 0.0320817679, -0.0396693051, -0.0825815424, 0.003613991, -0.0284103211, -0.0088330936, -0.0015564633, 0.0229513161, -0.0216608085, -0.0314163826, -0.0247015934, -0.0317119993, 0.0125867911, 0.0375931673, -0.0571245961, -0.004979379, -0.0513389744, -0.0265274644, -0.0287670419, 0.0071322569, -0.0121973045, -0.0316294283, -0.0379096866, -0.0211432762, -0.0256738141, -0.0514236465, 0.0035389145, 0.0097526945, -0.0035747248, -0.0268124547, -0.0295934472, 0.0411286429, 0.024391884, -0.0676976219, -0.0447738618, 0.0969569832, -0.0059717922, 0.0107443482, 0.025411129, 0.0307264179, 0.0144180013, -0.0065585673, -0.0369886979, -0.0378102288, 0.0381543264, 0.0622191392, -0.0346245989, -0.0079273237, -0.007222916, -0.0673659295, 0.028092226, -0.0304050632, 0.0409638807, -0.0255524032, 0.0477170236, -0.0110091092, -0.0241614375, 0.0575583875, -0.0543410853, -0.048303429, -0.0538713001, -0.0463183895, -0.0331095569, 0.0191626605, -0.0494033583, 0.0323279649, -0.0320100896, -0.012144383, -0.0141724143, 0.0432290919, -0.005326841, -0.044811029, 0.0513908751, -0.0142982509, -0.0023246859, -0.0191832576, -0.0490588099, 0.0120830666, -0.0162866563, -0.0297601484, -0.0086084334, -0.0180185381, -0.0380871557, 0.0221053623, -0.0131460996, 0.0286903474, 0.0537379794, -0.0084787635, 0.0326235853, 0.020058997, 0.0276074111, -0.0020823833, -0.0424302891, -0.0398316421, -0.0399116278, 0.0095292237, -0.0424499661, -0.0020538031, 0.0135678574, 0.0027289088, -0.0121446569, 0.0437926129, 0.0475187637, 0.0059657679, -0.0118739894, 0.0054783467, -0.0419987924, 0.0236910656, 0.0795392022, -0.0038362679, -0.0290599409, -0.0279124323, -0.0043396722, 0.0220966898, 0.0287072435, 0.0180476084, -0.0549244173, -0.0041418374, 0.0318586864, 0.0468131751, 0.0468271002, 0.0303614903, -0.0266215019, -0.0492063798, -0.011344987, 0.0049503352, -0.0817475393, -0.023283897, -0.0460679419, -0.034580376, -0.0440824218, -0.0154562145, -0.0185161494, 0.0190460458, 0.0068891337, -0.0225355905, -0.0025638961, -0.0480021089, 0.0416126251, -0.0586879253, 0.0772290304, 0.0573529527, 0.0556462891, -0.0464717411, -0.0366931595, 0.0019499224, 0.0268354956, -0.0172591079, 0.0182211269, 0.0259776637, 0.0511928536, -0.0232753772, 0.0391194038, 0.0036914921, -0.0293690339, 0.0374373049, 0.0294869579, 0.0531864129, 0.0142479315, 0.0171805881, 0.0245302152, 0.0280096903, 0.0259650145, 0.0794310942, 0.0293522608, -0.0682886019, 0.0241112374, -0.0038903558, -0.0058004791, -0.0922333747, 0.0141765177, 0.0161357932, -0.0322866961, -0.0545720607, -0.0151669821, -0.0212670974, -0.0072353967, -0.0190995373, 0.0581968315, -0.0090138735, -0.0246682502, -0.0165339019, -0.0213789642, 0.0036114273, -0.0175516102, 0.0639716536, -0.0271482524, -0.0544607826, -0.0025507712, -0.02387725, -0.0145240603, -0.013541053, -0.0673867017, -0.0421550721, 0.0190227553, 0.0213090666, 0.0125939911, 0.0624367148, -0.0100912703, -0.0097199241, -0.0296643693, -0.0303264111, 0.0370949693, -0.0019355885, 0.0589946769, -0.0094586294, -0.0046480373, 0.0351367965, -0.0477374308, 0.0475903712, 0.0285381489, 0.022229122, 0.0128920609, -0.0449649394, -0.006064747, 0.0101680383, 0.0028725401, -0.0297138896, -0.0350575522, -0.0766262412, -0.0145389568, -0.0265732147, -0.00420056, -0.0230438579, -0.0074366145, 0.0913606435, 0.0280400794, -0.0205506142, -0.0206810813, 0.0030415426, -0.0416716635, -0.032487575, 0.047684513, 0.0811596587, -0.0459588356, 0.021542795, 0.0266538598, -0.0239993129, 0.0399289988, 0.0412775353, -0.0061600911, 0.0001965234, -0.0325267278, -0.0097231325, -0.0448148884, 0.0135740144, 0.0113051357, 0.0052533736, 0.0086194407, 0.025654953, 0.0135335065, 0.0800170228, -0.0152076492, 0.0536103323, 0.0179600045, -0.0130000329, -0.0100050187, 0.0279551297, 0.007023674, -0.0428546295, -0.0207191166, -0.0003063953, 0.0076479902, 0.002366619, 0.0718962774, -0.0186961368, 0.0194500219, -0.0200756248, -0.0020132931, 0.0196146537, 0.0246195253, -0.0214208644, -0.0318323709, -0.0175451729, 0.0179050975, -0.0314679667, -0.0696814954, 0.0015480078, 0.019094808, 0.0106598074, 0.0153187932, -0.0185133554, 0.0483795702, -0.0054887095, 0.0138254669, -0.0324503072, -0.0539127402, 0.0147148734, -0.0074490025, 0.0025996123, -5.301295851e-33, -0.012751949, 0.0552662872, -0.0192268975, 0.0163715258, -0.0086051151, -0.0033657262, -0.0631378293, 0.0137191219, -0.0189561378, -0.0182381198, 0.006814593, 0.0024738873, 0.0079381298, 0.0032103902, 0.0540994145, 0.0205373392, 0.0548207574, 0.0064538904, -0.0437703431, 0.0315804072, 0.0764844865, 0.0635910556, 0.0453939512, -0.0697662756, 0.0399788953, -0.010405913, 0.0300000329, 0.0336996578, 0.0188136958, 0.0367595293, -0.0115800314, 0.023293145, 0.0150899608, -0.0348505639, -0.0318980254, 0.0044628042, -0.0412790552, -0.0256976523, 0.0076574767, 0.0703754649, 0.0396090746, -0.0104634594, -0.0145670967, -0.031721402, -0.0158915985, -0.0052340846, 0.0410812087, -0.0462972037, -0.0051403935, -0.0465300381, -0.0311083738, 0.0097335624, -0.0131695597, -0.0322171599, 0.0852508917, 0.0643647984, 0.0022796909, 0.0114746448, -0.0396281146, 0.0243470538, -0.0369592942, 0.023569515, 0.0461865775, 0.0399812162, -0.0338420495, 0.0139202103, 0.0046180906, -0.0484336093, 0.0138907991, 0.0347150639, 0.0127781769, 0.0140878968, 0.0185016766, -0.0026616757, 0.0496141538, -0.0306919459, -0.0403218083, -0.0151219787, -0.0074461978, -0.0571225956, 0.0032235549, 0.0312592275, -0.0512853004, -0.0123848664, -0.0200104769, -0.0698694959, -0.0252165869, -0.0152003011, 0.0203360096, -0.0120946597, -0.017382063, 0.0180445723, 0.0287416335, -0.0213481225, -0.0288134217, 0.0196065512, -0.0255845897, -0.0056579132, -0.0098894797, 0.0374398828, -0.0527269579, -0.0199222676, 0.0222752392, 0.0123441853, 0.0419884808, -0.0093818782, 0.0831690878, 0.0876243487, -0.070832558, 0.0028416237, -0.031109903, -0.0002156819, 0.0025898672, 0.0346741155, 0.0463179573, 0.0179917756, -0.0025346701, 0.0617854334, 0.0513083674, 0.0379094891, -0.0248320699, -0.0422019921, -0.0148126194, -0.0146978162, 0.0039821337, 0.0114215286, -0.0013528088, -0.0284103882, 0.0103510823, -0.0479998179, -0.0178466402, 0.0121708121, 2.607e-7, 0.0337843783, -0.0283074956, -0.0031498782, 0.0096423682, 0.0289601739, 0.0264155529, -0.1163787469, -0.0043699117, -0.0067065493, 0.0458892956, 0.0054909284, -0.0280687604, 0.1128008217, 0.0692627206, -0.1047140658, 0.0022554919, -0.0667740107, 0.0075955023, -0.0731684119, -0.0335383266, 0.0078932308, 0.0855149329, 0.012176414, -0.0265166387, -0.0301460698, -0.04919957, 0.0231514145, -0.0098860543, 0.001154836, -0.0666631237, -0.0017232571, 0.0259282645, -0.0003966416, -0.0005946953, -0.0799742937, 0.0828907937, 0.0418300368, 0.0420972109, 0.0204839464, 0.0172242485, 0.0556129701, 0.0333433822, -0.0471363179, -0.0207562149, 0.0400045551, 0.0199469049, -0.0369271711, -0.0209495295, 0.0207445547, 0.0436250418, 0.0360827148, -0.0377565324, -0.0260054506, -0.0291517954, 0.0375606157, -0.0421829559, 0.0161462948, -0.0136803901, -0.0098988144, 0.0382258072, -0.0390808098, -0.0007114505, 0.0093195457, -0.01991155, -0.0127893491, -0.0555505753, 0.0187241919, 1.852853915e-34, 0.0651146993, -0.0542197302, 0.0063606151, -0.0088272458, 0.0217203815, -0.0566271096, 0.0190623626, -0.0186511558, -0.0429719239, -0.0194950141, -0.0620152354 ]
http://arxiv.org/abs/2406.17600v1
http://arxiv.org/pdf/2406.17600v1
"Seeing the Big through the Small": Can LLMs Approximate Human Judgment Distributions on NLI from a Few Explanations?
2024-06-25T14:42:17
2024-06-25T14:42:17
Beiduo Chen, Xinpeng Wang, Siyao Peng, Robert Litschko, Anna Korhonen, Barbara Plank
cs.CL
cs.CL
Human label variation (HLV) is a valuable source of information that arises when multiple human annotators provide different labels for valid reasons. In Natural Language Inference (NLI) earlier approaches to capturing HLV involve either collecting annotations from many crowd workers to represent human judgment distribution (HJD) or use expert linguists to provide detailed explanations for their chosen labels. While the former method provides denser HJD information, obtaining it is resource-intensive. In contrast, the latter offers richer textual information but it is challenging to scale up to many human judges. Besides, large language models (LLMs) are increasingly used as evaluators (``LLM judges'') but with mixed results, and few works aim to study HJDs. This study proposes to exploit LLMs to approximate HJDs using a small number of expert labels and explanations. Our experiments show that a few explanations significantly improve LLMs' ability to approximate HJDs with and without explicit labels, thereby providing a solution to scale up annotations for HJD. However, fine-tuning smaller soft-label aware models with the LLM-generated model judgment distributions (MJDs) presents partially inconsistent results: while similar in distance, their resulting fine-tuned models and visualized distributions differ substantially. We show the importance of complementing instance-level distance measures with a global-level shape metric and visualization to more effectively evaluate MJDs against human judgment distributions.
"Seeing the Big through the Small": Can LLMs Approximate Human Judgment Distributions on NLI from a Few Explanations?: Human label variation (HLV) is a valuable source of information that arises when multiple human annotators provide different labels for valid reasons. In Natural Language Inference (NLI) earlier approaches to capturing HLV involve either collecting annotations from many crowd workers to represent human judgment distribution (HJD) or use expert linguists to provide detailed explanations for their chosen labels. While the former method provides denser HJD information, obtaining it is resource-intensive. In contrast, the latter offers richer textual information but it is challenging to scale up to many human judges. Besides, large language models (LLMs) are increasingly used as evaluators (``LLM judges'') but with mixed results, and few works aim to study HJDs. This study proposes to exploit LLMs to approximate HJDs using a small number of expert labels and explanations. Our experiments show that a few explanations significantly improve LLMs' ability to approximate HJDs with and without explicit labels, thereby providing a solution to scale up annotations for HJD. However, fine-tuning smaller soft-label aware models with the LLM-generated model judgment distributions (MJDs) presents partially inconsistent results: while similar in distance, their resulting fine-tuned models and visualized distributions differ substantially. We show the importance of complementing instance-level distance measures with a global-level shape metric and visualization to more effectively evaluate MJDs against human judgment distributions.
[ 0.0694849268, 0.0391089506, -0.0296425242, 0.034589231, -0.0173592456, 0.0082021551, 0.0054564644, -0.0035156384, -0.0220301002, -0.0604306757, -0.0594497174, -0.0219274126, 0.0036897468, -0.0183298644, 0.0408798866, -0.0561299324, -0.0036842988, -0.0395193696, 0.0414502062, -0.0060824747, 0.0301572643, 0.0045832405, -0.0121910935, 0.0714194849, -0.0115247816, 0.0218108427, -0.0149983745, 0.0056661554, 0.0384291597, -0.0120438077, 0.0055220528, 0.0670954213, 0.0444224626, -0.0033257238, 0.00000194, -0.0258554649, 0.0475187339, 0.000520558, -0.0175613705, 0.0411838666, 0.0096746581, 0.0035405078, 0.0312835649, -0.0199480075, -0.0129164374, -0.0153753133, 0.0524339527, 0.0297022983, 0.0541704334, 0.0192867741, 0.0169374254, -0.0528449193, -0.0454799645, -0.0277586505, 0.005718594, -0.0790459365, 0.0020258168, 0.0157314483, -0.0126705803, 0.0370596685, 0.0029739994, -0.00086818, -0.0101937726, 0.0001122102, 0.0219418686, 0.0092473598, -0.0102233188, -0.0327605531, 0.0165899992, 0.0682984069, 0.0497385263, -0.0344328322, -0.0385565311, -0.0208042171, 0.0330300517, 0.0407858938, -0.0023479429, 0.0323792323, -0.0184046533, 0.0179459713, 0.0129736299, 0.0412070602, 0.0311000179, 0.023483675, 0.0287759341, 0.0589499697, -0.0145858265, -0.0056484747, 0.0030626724, 0.0465276539, -0.0489852466, -0.031845361, 0.0046728468, -0.0008552946, -0.0401367098, -0.0093850736, -0.0024048341, -0.0213119145, 0.0284090396, -0.0320023783, 0.0264780689, -0.0008779748, -0.0180741046, 0.0481818989, 0.0242677424, 0.025640754, -0.0083886292, 0.0309845712, -0.0320079885, -0.0474227034, -0.05157502, 0.0139894774, -0.0039780843, -0.0078035342, -0.0392714888, 0.0310199615, -0.0523890778, -0.0126948943, -0.0672012269, -0.0340444073, 0.036952205, 0.0324719176, -0.0027100823, 0.0004842735, -0.0996373296, 0.0133974832, -0.0966742337, 0.0380537808, -0.0000325644, -0.0236430503, 0.0121684317, 0.0280201212, 0.0171783455, -0.0015886405, -0.0028240869, 0.0680477247, 0.029744219, 0.0063971188, -0.0425661914, -0.0307801217, -0.0327039398, -0.0425414145, -0.0343990885, -0.0597690716, 0.0211772174, 0.0094462037, -0.010415744, -0.0446235165, -0.0140903918, -0.0116321743, -0.0541289039, 0.0783176944, -0.0108771548, 0.0203808397, -0.003611733, -0.008970811, -0.0389460325, 0.0411937162, 0.0136400172, 0.0207400974, 0.027500771, -0.0230473392, 0.026690701, 0.0050904271, 0.0039971354, 0.0383118689, -0.0253035296, -0.0050738733, -0.0395210087, 0.0274947099, 0.0417633094, 0.0439670049, -0.0361642987, -0.0099518057, 0.0560907349, 0.0597299114, -0.0010548818, -0.0130971428, 0.0555978231, 0.0308114532, -0.0120819891, 0.0104561504, -0.0540695898, 0.0270513445, -0.028501844, -0.017944077, 0.0483136959, -0.0011059231, -0.0119569506, -0.0491859578, -0.0241849404, 0.0065103248, -0.0087735038, 0.003245634, 0.0416092761, -0.0020812252, 0.0374139175, 0.0690881461, 0.0926489308, -0.0357343629, 0.0004868336, -0.0589971244, 0.0373246297, 0.0482439585, 0.0054019662, 0.0022902351, 0.0233318619, 0.0237372238, -0.0425318964, 0.0206861403, 0.0701423064, -0.0076530525, -0.0128240474, -0.0205153581, -0.0423411019, -0.0134045547, 0.0492246039, 0.0215153657, 0.0262779333, -0.0292006135, 0.0216792449, -0.1264790595, -0.0261241645, 0.0223616064, 0.0362121649, 0.0117919696, -0.0059786313, 0.0233130939, 0.043623995, 0.0094684837, 0.0004155316, 0.0104475338, -0.0044017904, -0.0055946172, 0.0000426855, -0.0219818465, -0.0192484204, -0.0287841279, 0.0095625268, -0.0592748113, -0.0340757556, -0.029806966, -0.0517739207, -0.0118468739, -0.0125193717, -0.0190886147, -0.0297723226, -0.0445325375, -0.0134307072, -0.0130649665, -0.0079894029, -0.0549128465, 0.0816282853, 0.021965852, 0.0203912258, -0.0220038835, 0.0309183486, 0.0144007485, -0.002766266, -0.0261344165, -0.0299270805, 0.0114856493, 0.0126332864, 0.0318367258, 0.0079973117, -0.0311590862, -0.0316118002, -0.0736700073, -0.0260040164, -0.0050777653, -0.0163202472, 0.0099947546, -0.0210911408, 0.0183661114, -0.0083483132, -0.0301169679, -0.0159059558, -0.046844095, 0.0426735915, 0.0387858339, -0.0040110792, -0.0097864829, -0.0389753208, -0.0110676056, -0.0176071059, 0.0493169688, 0.0529489741, -0.0320041068, -0.0962400064, -0.0055411784, -0.0559595972, 0.0091202529, -0.0038006285, -0.0412630774, 0.0452934727, 0.0190438833, 0.0576687753, 0.0633712336, 0.0127537791, 0.028623715, 0.0553429313, 0.0126145231, -0.0587246828, -0.0023429089, -0.0111043323, -0.0337196328, -0.0427342616, -0.0562378913, 0.0010941153, 0.0602949075, 0.0139547205, -0.0173966605, -0.0717559755, 0.0275420956, -0.0164122451, -0.0290995259, -0.0326842442, -0.0216219518, -0.0106494827, 0.0779732391, 0.0034274289, -0.0588868037, 0.0931687728, -0.021430593, 0.0436990447, 0.0038344918, -0.0317806862, -0.0513183773, 0.0657740533, 0.0063431701, -0.0076656537, -0.0298253968, 0.0144445039, -0.0319553092, 0.0223602187, 0.0668089911, -0.0050643859, 0.0184845924, -0.067573972, 0.0024641594, -0.0172810238, -0.0271827914, -0.0057244077, -0.0574548654, -0.0729473233, -0.0167370345, 0.0738038868, -0.0022414245, -0.0296714399, 0.0376319662, 0.0270158108, -0.0136254951, -0.0018893864, 0.0286372118, -0.0210279673, -0.0110793822, 0.0143226516, 0.0309134927, 0.0204048734, -0.059811037, -0.035921894, 0.0189161599, -0.0200738125, 0.0060743592, -0.0181894079, -0.0036023303, 0.0264168046, 0.0252814386, -0.0229859501, 0.0495995432, 0.1088109985, -0.0126017826, 0.016268203, 0.0934909061, -0.0036455591, -0.0712240711, 0.0199646875, -0.0047536963, 0.0570686683, 0.0115645342, -0.0226197429, -0.048577521, -0.0642154962, 0.0088841049, 0.0130082294, -0.0024769162, -0.0241814405, 0.020201413, -0.0920035914, 0.0339971446, -0.0021726969, -0.0448960513, -0.0068385848, -0.0583824366, 0.0576746836, -0.0159790739, -0.0310867392, -0.0143321902, 0.0471128337, 0.025078563, 0.0361265317, -0.0034195704, -0.0342336372, 0.0652958155, -0.0695138648, 0.0582230352, 0.0120651992, 0.020734638, 0.0092684152, -0.0940504, 0.0285413265, -0.0023722248, 0.0444406532, 0.0173466112, 0.0747895911, -0.0018559137, -0.0119410316, -0.0609124787, 0.0268240161, -0.0094092879, -0.0095479162, -0.0115156472, 0.043197494, -0.0038704681, -0.0223354232, -0.0263554342, 0.014798644, -0.0203865971, 0.0783982724, 0.0530880168, -0.0674961135, -0.0323521271, -0.0079311254, -0.011280491, -0.0009579191, -0.0835195333, 0.0533242561, -0.0067211594, -0.0757828951, -0.0191482715, -0.018246185, 0.0014920271, 0.0213219579, 0.0104271211, -0.0166049879, 0.0194639601, 0.0174519755, 0.0390167758, 0.082097508, -0.0512715913, 0.0242357552, -0.0285082124, -0.0100841206, 0.0216736328, -0.0040958901, -0.0322975256, 0.0432294011, -0.0866602808, -0.0481518731, -0.0090947058, 0.0046591307, 0.01617782, 0.0364536121, 0.000938033, -0.0826204792, 0.0004584893, 0.0450690165, -0.0199257191, 0.0163900238, -0.001275464, 0.0167851094, -0.0039856704, 0.0415563397, -0.0325085409, 0.0373102725, 0.0462471992, 0.0145731056, -0.0357206687, -0.1001605019, 0.0179815292, 0.0143789919, 0.0546100587, -0.0258422494, -0.0318252034, -0.0288668294, -0.0343967564, 0.0146049308, -0.0511785187, -0.02879861, -0.0279653911, 0.0219428316, -0.0459106639, -0.0038212072, -0.0004596575, 0.0151756657, 0.0478687771, -0.0149725387, 0.0096626971, 0.0655436069, 0.0379714482, 0.003329593, -0.0036555559, -0.0309359152, -0.0254939329, -0.0380435437, 0.0492247716, 0.0164963361, 0.0367253795, 0.0247571431, 0.0021114668, -0.0103391716, -0.0062282565, -0.0488672741, -0.0328336768, 0.0013613165, 0.0192885362, 0.012669012, -0.0009168687, 0.0243524984, 0.0241328385, -0.0171111152, 0.0007839007, 0.0477847196, -0.0018090692, 0.0177834108, 0.0538267605, 0.0153684197, -0.0200837851, -0.0070882868, 0.0465504415, 0.0254486706, -0.0109005328, -0.0385360122, -0.0325016342, -0.0207141992, 0.0011732405, -0.0333836973, -0.0015943524, 0.0101505192, 0.024578901, -0.0818642601, -0.0563155338, -0.0414016843, 0.018720232, -0.0143000046, -0.0014628116, 0.0139756519, 0.0737018362, -0.0190819874, -0.0465829, -0.0404632352, -0.0048040501, 0.0331040546, 0.0356560126, -0.0611983761, -6.697910975e-33, -0.0204062611, -0.0628981814, -0.0337667018, -0.0592981242, -0.0009377168, -0.0073150112, 0.0099686896, 0.0107115358, -0.0300503206, 0.0632680655, 0.0157902632, -0.0012654272, 0.0157927871, -0.0109202247, 0.01737695, -0.0278419908, 0.0316354595, 0.0314366221, -0.0269902516, 0.0253594369, 0.0050013661, -0.0352488719, 0.0596816167, -0.1094119102, 0.0900680646, 0.0248856787, 0.0032498103, 0.035347227, -0.0249430723, 0.0314772725, 0.0143067567, 0.0510182977, 0.020691745, -0.0312245432, -0.0023302585, 0.010499266, -0.0681732818, -0.0143476827, 0.0608612783, -0.0056483857, -0.0465822965, -0.0187174957, 0.0455887429, 0.0157402139, -0.0102267014, -0.0070193489, 0.0775970966, -0.029291397, -0.003335601, -0.0070859389, 0.0158419367, 0.0426564664, -0.0257005934, 0.0004022178, -0.0235373303, 0.0182179753, -0.0167510007, 0.0631002858, -0.0689476281, 0.0270134509, 0.0336782858, -0.0097298436, 0.0327430815, 0.0099522853, -0.0078190053, 0.0667802542, 0.0056452267, 0.0287804361, -0.0606276318, -0.0931281894, -0.0054766014, 0.0639901906, 0.0498283096, 0.0648485869, 0.0235556848, 0.0156332124, -0.0409887359, 0.0115885632, -0.0297408234, 0.0153594902, 0.0413095169, 0.0185557399, -0.0659466311, -0.0285056774, -0.0516779087, -0.0472689644, -0.0502930097, 0.0328006037, 0.0134722935, -0.0299646296, -0.0172226485, -0.0155087141, 0.0097816, -0.0302675534, 0.015675405, -0.0164368413, 0.0067095063, 0.0027923675, -0.0072575207, -0.0028833975, -0.030587947, -0.0103282314, -0.0090723345, 0.0175580923, 0.0199721288, 0.0032687027, 0.0553823449, -0.0337260365, -0.0238141641, -0.0036012984, -0.037027467, -0.0484618805, -0.051082775, -0.0226081535, 0.0649963766, 0.0238913503, -0.0330904759, 0.0451215766, 0.0192942861, 0.0268272106, 0.0268176198, -0.0245694313, -0.0205457006, 0.0015110696, -0.0242799483, -0.0103097102, -0.0203233361, -0.0407260433, 0.0332423449, -0.0103927376, -0.0131735634, 0.0322224908, 2.656e-7, 0.0385097452, 0.0407970101, -0.0105798077, 0.0073963623, 0.0183157437, 0.0797776058, 0.0412600711, -0.0295092091, 0.0538302958, -0.0673114583, 0.0120143509, 0.0431395136, 0.0394073464, 0.0257420037, -0.0386047959, 0.011308765, -0.0653052926, -0.0154333543, -0.0019837618, 0.013900185, 0.0623279102, 0.061263077, 0.057866063, -0.0049221916, -0.007798837, -0.0376417525, 0.0556445271, -0.0304828361, -0.0106427157, 0.0325707495, 0.0543429218, -0.025275223, 0.0396640413, 0.0568121709, 0.0239007622, -0.0106358882, 0.0164430272, 0.0272143371, -0.0292903259, 0.0036048207, -0.0347517617, -0.0265179165, 0.0253717098, -0.0327380821, 0.0550297014, -0.04132393, -0.0622735545, -0.0082478784, -0.010976065, -0.0256191287, 0.0024393173, 0.0023095736, -0.0133085037, 0.025769081, -0.0089324554, -0.0100407256, -0.0375936516, -0.0512433983, -0.0455545299, 0.0191299282, -0.0510634892, -0.0518210866, -0.0016316627, 0.0037761235, 0.0315157063, -0.0431117564, 0.0321044773, 1.943926258e-34, 0.0450142398, 0.0136491274, 0.0136752101, -0.0135034891, 0.090540193, 0.0339445472, -0.040355783, -0.025666682, 0.044977773, -0.0444294214, -0.0073674652 ]
http://arxiv.org/abs/2406.17588v2
http://arxiv.org/pdf/2406.17588v2
LongIns: A Challenging Long-context Instruction-based Exam for LLMs
2024-06-25T14:31:26
2024-06-26T13:28:04
Shawn Gavin, Tuney Zheng, Jiaheng Liu, Quehry Que, Noah Wang, Jian Yang, Chenchen Zhang, Wenhao Huang, Wenhu Chen, Ge Zhang
cs.CL
cs.CL
The long-context capabilities of large language models (LLMs) have been a hot topic in recent years. To evaluate the performance of LLMs in different scenarios, various assessment benchmarks have emerged. However, as most of these benchmarks focus on identifying key information to answer questions, which mainly requires the retrieval ability of LLMs, these benchmarks can partially represent the reasoning performance of LLMs from large amounts of information. Meanwhile, although LLMs often claim to have context windows of 32k, 128k, 200k, or even longer, these benchmarks fail to reveal the actual supported length of these LLMs. To address these issues, we propose the LongIns benchmark dataset, a challenging long-context instruction-based exam for LLMs, which is built based on the existing instruction datasets. Specifically, in our LongIns, we introduce three evaluation settings: Global Instruction & Single Task (GIST), Local Instruction & Single Task (LIST), and Local Instruction & Multiple Tasks (LIMT). Based on LongIns, we perform comprehensive evaluations on existing LLMs and have the following important findings: (1). The top-performing GPT-4 with 128k context length performs poorly on the evaluation context window of 16k in our LongIns. (2). For the multi-hop reasoning ability of many existing LLMs, significant efforts are still needed under short context windows (less than 4k).
LongIns: A Challenging Long-context Instruction-based Exam for LLMs: The long-context capabilities of large language models (LLMs) have been a hot topic in recent years. To evaluate the performance of LLMs in different scenarios, various assessment benchmarks have emerged. However, as most of these benchmarks focus on identifying key information to answer questions, which mainly requires the retrieval ability of LLMs, these benchmarks can partially represent the reasoning performance of LLMs from large amounts of information. Meanwhile, although LLMs often claim to have context windows of 32k, 128k, 200k, or even longer, these benchmarks fail to reveal the actual supported length of these LLMs. To address these issues, we propose the LongIns benchmark dataset, a challenging long-context instruction-based exam for LLMs, which is built based on the existing instruction datasets. Specifically, in our LongIns, we introduce three evaluation settings: Global Instruction & Single Task (GIST), Local Instruction & Single Task (LIST), and Local Instruction & Multiple Tasks (LIMT). Based on LongIns, we perform comprehensive evaluations on existing LLMs and have the following important findings: (1). The top-performing GPT-4 with 128k context length performs poorly on the evaluation context window of 16k in our LongIns. (2). For the multi-hop reasoning ability of many existing LLMs, significant efforts are still needed under short context windows (less than 4k).
[ -0.008756944, 0.0393693522, -0.0178143103, 0.0712118447, -0.0639229789, 0.0205875449, 0.0233527031, 0.0000782907, 0.010602436, -0.0510865375, -0.0230490286, -0.0680123791, -0.0181226414, -0.0335265584, 0.0272367056, -0.0160022564, 0.0572472997, -0.0598413013, -0.0157929398, -0.016233705, -0.0273962282, -0.0035994323, -0.0085784923, 0.0005072865, 0.0215920657, -0.0402635112, -0.0179761332, -0.0050822813, 0.0088022863, -0.0406950973, 0.0073521989, 0.0354670323, 0.0210027117, 0.0204507988, 0.0000020535, -0.0406385623, -0.008557627, 0.0109704537, -0.048360832, 0.0430591255, 0.0055205002, 0.0254028197, -0.0082403347, 0.0291301366, -0.0297550075, 0.0121528637, -0.0021587117, 0.0096679619, 0.0562571809, 0.07900206, 0.0193193089, -0.0399192944, 0.0406697616, -0.0280306377, 0.042307537, -0.0569529422, 0.013653622, 0.0051157959, 0.046384804, 0.0136248991, -0.0100099426, 0.0156360343, 0.0630648434, 0.0354490727, -0.0683071166, 0.0240508318, -0.0406129621, 0.0421399623, 0.0102070495, 0.057660643, 0.0107815452, -0.000196664, -0.0168569405, 0.0124993008, -0.007061997, -0.0317923687, -0.0457911603, 0.0542175248, -0.0074128127, -0.0000924718, 0.0389230587, -0.0069053206, 0.0358605012, 0.0062818374, -0.0577939749, 0.0949274823, 0.0383404754, -0.0461428091, -0.0020651007, -0.0134083163, -0.0444475524, -0.032709416, -0.0245983806, 0.0525162816, -0.0262314826, 0.0269800927, -0.0149222519, 0.0261321831, 0.0346287824, -0.0086141899, 0.0167036392, 0.0389569774, -0.0190531109, 0.054635819, -0.035191074, 0.0745846853, -0.0127291577, 0.0277931057, -0.0425950028, -0.0237185191, -0.0404325016, 0.0128232446, -0.0191443637, 0.0302745681, -0.0472916663, 0.0155194206, -0.0428402387, 0.0286917817, -0.0297008194, 0.0258935224, 0.0110343182, -0.0289780963, -0.0148341376, 0.0025075711, -0.0407073312, 0.0085847154, -0.0414382704, -0.0027915109, 0.0112986024, -0.0556816384, -0.0270706695, 0.0166569706, -0.0088154739, -0.0612959862, 0.037236236, 0.0667760372, 0.0093640024, -0.0319103114, -0.0447971635, -0.0320075937, -0.020392945, -0.097627297, 0.0133394431, -0.0705432296, 0.0173723232, 0.0109887291, -0.011378428, -0.0055438858, 0.0061324062, 0.0390963852, 0.0235322844, 0.0535473749, 0.0363610536, -0.0068374304, 0.0020362462, 0.0401367284, 0.0099420259, 0.0734296963, 0.0052783331, 0.0242819078, 0.0431580059, -0.0124984924, -0.0236450117, 0.0332693718, 0.031687893, 0.0259941574, -0.0119410418, 0.0626710206, 0.0342070982, 0.0127018895, -0.0061618723, 0.065374434, -0.0180592369, -0.0375041217, 0.0224171225, 0.059181381, 0.030345073, -0.0208323449, 0.0535691306, 0.0100382473, -0.0015827247, 0.0155601446, -0.0629665107, 0.0137571264, -0.0375567712, 0.0383378081, 0.0334848873, 0.0108903032, -0.0307534747, -0.0290322322, -0.0228982419, 0.0049041635, -0.0220701043, 0.0453128405, 0.0462755002, -0.0347225592, -0.0377041548, 0.0677417591, 0.0228993762, -0.0481169149, -0.0253183693, -0.0553668588, 0.0226771049, 0.0414085351, 0.0386361554, -0.0290840212, -0.0662818104, 0.0116918469, 0.0305069201, 0.0242536739, 0.0450690761, -0.0085281795, 0.0179223772, -0.0309834108, -0.0172022898, -0.0402811244, 0.0035677981, 0.0097235413, 0.0527694933, 0.0194609594, 0.0333182923, -0.0627456233, 0.0244037788, 0.0169769619, -0.0456182659, -0.0425724201, -0.0322485305, 0.0144036869, 0.0361790992, 0.0139675597, -0.0075713908, 0.014796216, 0.0478153415, -0.0305875335, 0.0172254834, -0.0395713225, -0.0656304508, -0.0420026593, 0.0523796752, -0.0163913276, -0.0189839471, 0.0223635118, -0.0406511799, -0.064055033, 0.0080168182, -0.0127608404, 0.0225912724, -0.0288252067, 0.0124931, -0.0032043783, 0.0047017094, -0.0452444255, 0.0739120916, -0.0011413641, 0.0426856801, 0.0356635451, 0.0488521643, 0.0291294195, -0.0642824993, -0.0498282202, 0.0176232494, -0.0075952983, -0.0076137194, -0.0123526668, 0.0305464957, -0.0620917976, -0.0302004553, -0.1167945415, 0.0484376326, -0.039816808, 0.041439008, 0.0153960856, -0.0107996054, -0.0033427405, 0.0212249346, -0.0616352037, -0.0184136722, 0.0136330128, 0.0357365906, 0.0017513183, 0.0257424582, -0.0261620358, -0.0216033608, -0.0338916518, -0.0224403925, 0.027379889, -0.0422618128, 0.0237399265, -0.0328803174, 0.032088507, -0.0157792419, 0.0140581308, -0.02579269, -0.0098306397, 0.0276974626, -0.0425519422, 0.0266566146, 0.0545651391, -0.0112782568, 0.0098385429, 0.0334461443, -0.0276276823, -0.0502929017, 0.0497004688, 0.0224278048, -0.0369188488, -0.0389242396, -0.0143154934, -0.0274098571, 0.0706664324, -0.0195937566, -0.0240081977, -0.0179173201, -0.0099084647, -0.0224139113, 0.0021644386, -0.0227805153, 0.0315582342, 0.0046901489, 0.030540416, -0.0298007242, -0.0568563752, 0.0591589622, -0.0508684218, 0.0034233588, -0.0375570878, -0.0090691783, -0.0630840436, 0.0990723744, -0.0044968752, -0.0156450104, -0.0394463539, -0.0105000976, -0.0417423286, 0.0055145202, -0.0055799186, -0.0470112413, 0.0777600557, -0.0502384938, 0.0171257611, -0.034506619, -0.0498129986, -0.0384751149, 0.0251653977, -0.0052148062, 0.0315679051, 0.0038538596, 0.0009246236, -0.0086273653, -0.0035369536, 0.0128963487, 0.032975968, -0.0258818176, 0.027281614, -0.0181197897, 0.020458268, 0.0373051576, 0.0015497233, 0.0075834235, -0.052998852, 0.0442812033, 0.0570832901, 0.0672489107, -0.0101921335, -0.0065455656, 0.0091894977, 0.0259291008, -0.0513440296, -0.0235037133, 0.0457294025, 0.0749292523, -0.0899500847, -0.0339889042, 0.1282634288, 0.0231233519, -0.0457432792, -0.0056799855, -0.0066683851, 0.0134186363, 0.0332278237, 0.0391867757, -0.0944440588, -0.0400458463, 0.0151622733, -0.0039444873, -0.0097712195, -0.0207210351, -0.0051253033, -0.1264143437, 0.0223266874, 0.0050728396, -0.0517845713, -0.0144969653, -0.0328031182, -0.0084456336, 0.0231588427, 0.0457157902, 0.0374412946, -0.0103145763, -0.0136632035, 0.0031536673, 0.0034175718, -0.0258003939, 0.1032355055, -0.0099879857, 0.0470782593, 0.0125212185, 0.0020875293, -0.0049355146, -0.0538559891, 0.0551228374, -0.0614138469, 0.0089896973, -0.0048920726, 0.0206169244, -0.0098983347, -0.0024373871, -0.0041046343, 0.0173823517, -0.041457314, -0.0031214948, 0.0550138094, 0.1041707471, -0.0307792556, 0.003462371, -0.0220160484, -0.0203814749, -0.0127435196, 0.0719383135, 0.0145612061, -0.0097015221, -0.0535434671, -0.0337005518, -0.0420623459, -0.0497026667, 0.001600444, 0.0283350721, -0.0013364041, 0.0353105739, -0.0211802293, -0.0433222577, 0.047088895, 0.0554656945, -0.0138740307, 0.0149512626, -0.0109446235, 0.0099835601, -0.0765535906, 0.0430238135, -0.060710106, 0.0396494232, -0.084752813, 0.0036612041, 0.0200222284, -0.0041825268, -0.0158854909, 0.0318747014, -0.1109751016, -0.0474541113, -0.0306989178, 0.0265130699, 0.0219976958, 0.0418521613, -0.013705791, -0.0144870747, 0.0394381471, -0.0414318107, -0.0288321115, 0.0276569147, 0.0245095398, 0.0012769068, 0.0015094703, 0.0295505635, -0.0412166044, 0.0146455355, 0.0354482494, 0.0406186357, -0.0023901972, -0.011084076, 0.0138155799, 0.0363672525, 0.0085672801, 0.0309996866, -0.0142012173, 0.008280389, 0.0597014278, -0.028606372, -0.0080742072, -0.014784487, -0.0045448048, 0.0227550417, -0.0470863134, 0.0162228644, -0.0130961565, 0.0197781157, 0.0156544261, -0.0286990106, 0.0419958048, -0.003881742, 0.0033645721, 0.022000771, -0.011179762, -0.0307010226, -0.0221353266, -0.0095636714, 0.0607822947, -0.0206315294, 0.0838188529, 0.0044631884, 0.013087024, -0.004666178, 0.0177969243, -0.0788806304, -0.0215798523, 0.0099528246, 0.0205415972, 0.0482550673, 0.0016733295, 0.0131463716, -0.0053999866, -0.0002344466, -0.0305494741, 0.0069304504, 0.0058785737, -0.0401971526, 0.035498675, 0.0765486211, 0.0095561827, 0.0250199772, 0.0739615932, 0.020115871, 0.0420428105, 0.0414630473, 0.012268723, -0.0377647057, -0.0450057536, -0.0130435228, 0.0233245082, -0.0029126843, 0.007897499, -0.0444701761, -0.0857856274, 0.0075944327, 0.0214790236, -0.0018114115, -0.0126345782, 0.0184460394, 0.0726207718, 0.0031060041, -0.0521870442, -0.028061945, 0.0089737559, -0.0109351007, -0.011613097, -0.0042556846, -6.102021665e-33, -0.0273647122, -0.0228336249, 0.07232517, -0.0349714719, -0.0244775265, 0.0185380466, -0.0258416664, -0.0102242874, -0.0150057673, -0.0097745908, -0.0289050601, 0.0279488564, 0.00351978, 0.0435197949, -0.0071994197, -0.0297085345, 0.0038655177, 0.0145628406, 0.0010291593, 0.037665043, 0.0152172009, -0.0089056306, 0.0403932929, -0.0086296033, 0.0206867363, -0.0100840004, 0.0307196863, 0.0147227151, -0.0682726279, 0.0779491439, -0.0048784399, 0.0398653783, 0.0377700105, -0.0585708618, 0.0033454667, -0.0145801883, -0.0583185814, -0.0399568006, 0.0287233889, -0.0137059484, -0.0362796709, -0.0541314706, 0.0159348585, 0.0131646395, -0.0199718345, -0.0454413258, -0.0033970671, -0.0330736674, -0.0276342984, -0.0040127346, -0.0157264899, 0.0316472426, -0.0012437222, 0.0013789788, 0.0131282713, 0.0152086336, 0.0004297846, 0.0203957818, -0.030388331, 0.0066861496, 0.0027966714, 0.0645319894, 0.0751880109, -0.0078831008, 0.0136417728, 0.0304863602, 0.0078054434, 0.0246225391, -0.0025032551, -0.0201091245, 0.010122275, 0.066372402, 0.0362585858, 0.0178631973, 0.0366685502, 0.0078893173, -0.0879472643, 0.0088393772, 0.0337452479, 0.0451774076, 0.0258324724, -0.0068588112, -0.0514773615, -0.0143019883, -0.0483028181, 0.034833964, -0.0391149893, -0.0120328795, 0.0558625348, -0.0137017574, 0.0154420733, -0.010903039, 0.0388823785, -0.037284568, -0.003241041, 0.0238190144, 0.0000583303, 0.0105988597, -0.0040404699, 0.0626343861, -0.025247328, -0.0319432169, 0.0024387012, -0.0144909965, 0.0198962055, -0.0132811926, 0.0210191049, 0.0087855533, -0.0569310524, -0.0371780843, -0.07637804, 0.0170005895, -0.0241308212, 0.0179470815, 0.0439521968, 0.0249297898, 0.0221905895, 0.081963025, -0.023308685, -0.0277785286, -0.0230190642, -0.0215828829, -0.0028324185, 0.0029775789, -0.0116878003, 0.0085522179, -0.0314331278, 0.0245712791, -0.0034393533, -0.0422187112, 0.0111653293, 0.042641934, 2.709e-7, 0.0229737405, 0.028732311, 0.0237745661, 0.0422878265, 0.0139839463, -0.0033937343, -0.0344628841, -0.0113118282, 0.0015284496, -0.0271919034, -0.0080100289, 0.043321304, 0.0496073477, 0.0400661044, -0.064110741, 0.0216659438, -0.0141195478, -0.0148741966, -0.0545942448, 0.0254526064, 0.0642057508, 0.0960313156, -0.0122724306, -0.0051670116, 0.0042330124, -0.0497326627, 0.0258140229, -0.045075357, -0.0322349593, 0.0190734547, 0.0146187581, 0.0125944568, 0.016530415, 0.028379865, 0.0060107964, 0.0096459175, -0.0308993291, 0.0810730234, 0.0223817937, -0.0592429712, -0.0066535827, -0.0531218834, -0.0251156017, -0.0374221727, 0.1145654172, -0.0271150749, -0.0390402637, -0.0204871055, -0.0123860398, -0.0468808934, -0.0130816111, 0.0190590248, 0.01309318, 0.0052369693, -0.0028999508, -0.0513741374, 0.0075016231, -0.0901015326, -0.0132518243, 0.0248233993, -0.0415627621, -0.0637463927, -0.0555144846, -0.0058997958, 0.0297820382, -0.0696131736, -0.0325225778, 2.970708175e-34, 0.0397830829, -0.0292854067, 0.0302004181, -0.0005787369, 0.0398381092, -0.0181042608, -0.0568653829, 0.0031590473, 0.0015138277, -0.0928453356, -0.0190568026 ]
http://arxiv.org/abs/2406.17574v1
http://arxiv.org/pdf/2406.17574v1
Beyond Text-to-SQL for IoT Defense: A Comprehensive Framework for Querying and Classifying IoT Threats
2024-06-25T14:14:35
2024-06-25T14:14:35
Ryan Pavlich, Nima Ebadi, Richard Tarbell, Billy Linares, Adrian Tan, Rachael Humphreys, Jayanta Kumar Das, Rambod Ghandiparsi, Hannah Haley, Jerris George, Rocky Slavin, Kim-Kwang Raymond Choo, Glenn Dietrich, Anthony Rios
cs.CL
cs.CL
Recognizing the promise of natural language interfaces to databases, prior studies have emphasized the development of text-to-SQL systems. While substantial progress has been made in this field, existing research has concentrated on generating SQL statements from text queries. The broader challenge, however, lies in inferring new information about the returned data. Our research makes two major contributions to address this gap. First, we introduce a novel Internet-of-Things (IoT) text-to-SQL dataset comprising 10,985 text-SQL pairs and 239,398 rows of network traffic activity. The dataset contains additional query types limited in prior text-to-SQL datasets, notably temporal-related queries. Our dataset is sourced from a smart building's IoT ecosystem exploring sensor read and network traffic data. Second, our dataset allows two-stage processing, where the returned data (network traffic) from a generated SQL can be categorized as malicious or not. Our results show that joint training to query and infer information about the data can improve overall text-to-SQL performance, nearly matching substantially larger models. We also show that current large language models (e.g., GPT3.5) struggle to infer new information about returned data, thus our dataset provides a novel test bed for integrating complex domain-specific reasoning into LLMs.
Beyond Text-to-SQL for IoT Defense: A Comprehensive Framework for Querying and Classifying IoT Threats: Recognizing the promise of natural language interfaces to databases, prior studies have emphasized the development of text-to-SQL systems. While substantial progress has been made in this field, existing research has concentrated on generating SQL statements from text queries. The broader challenge, however, lies in inferring new information about the returned data. Our research makes two major contributions to address this gap. First, we introduce a novel Internet-of-Things (IoT) text-to-SQL dataset comprising 10,985 text-SQL pairs and 239,398 rows of network traffic activity. The dataset contains additional query types limited in prior text-to-SQL datasets, notably temporal-related queries. Our dataset is sourced from a smart building's IoT ecosystem exploring sensor read and network traffic data. Second, our dataset allows two-stage processing, where the returned data (network traffic) from a generated SQL can be categorized as malicious or not. Our results show that joint training to query and infer information about the data can improve overall text-to-SQL performance, nearly matching substantially larger models. We also show that current large language models (e.g., GPT3.5) struggle to infer new information about returned data, thus our dataset provides a novel test bed for integrating complex domain-specific reasoning into LLMs.
[ -0.0100571951, 0.0795632824, -0.0079133594, 0.0268057492, -0.0587800108, -0.0195292793, 0.0060460125, 0.0056741703, 0.0422976017, -0.0695841163, 0.0443086103, 0.0229053386, -0.0243227091, 0.0337019078, 0.0596829839, 0.0150076458, 0.0611094981, -0.0128849801, -0.010458136, 0.0603797883, 0.003914217, 0.0244474653, -0.002485848, 0.0072540208, -0.008166167, -0.0024600863, -0.0333219767, 0.0215474404, 0.0202801153, -0.0512565561, 0.002367422, 0.0156544074, 0.0466859452, 0.0171400607, 0.0000016931, -0.0475457273, -0.0657954589, 0.0052601793, -0.0337176695, -0.0428242311, 0.029058015, -0.0366328098, 0.0139750037, 0.0630579069, -0.0565735102, -0.0073319948, 0.0286606234, 0.0790305883, 0.0227332246, 0.1003995538, -0.0307520535, -0.0192445591, 0.0340566523, -0.0064706956, 0.0410431288, -0.0094843525, -0.0010651388, -0.0293947645, 0.0271165837, 0.011504082, 0.037411239, -0.0368211269, 0.0046169795, 0.0178671237, -0.0123883402, 0.0234071799, -0.0278978888, 0.0669077635, -0.0419161655, 0.026164202, -0.0052539078, -0.0304177478, 0.0408122838, -0.0065316409, -0.0234096367, -0.0316607617, 0.0025384186, -0.0297011621, 0.0159756932, 0.028366074, 0.005488622, 0.006004408, 0.0553846173, 0.0244958177, -0.0495188646, 0.0799290612, -0.0270498674, -0.0131988181, 0.0174960215, -0.0596915744, 0.0129984831, -0.0301665161, -0.0201351196, 0.0678005293, -0.0164882988, 0.0047438801, -0.0451444872, 0.0126180612, 0.0103826523, -0.0264859777, -0.0285412353, 0.0123467445, 0.0214470588, 0.0365638919, 0.0172076561, -0.0007504809, -0.0064911018, 0.0347232558, -0.0194419622, -0.0401917435, -0.0740602985, -0.0357091501, -0.0086759618, 0.0545872822, -0.002385119, -0.0062985192, -0.0177800134, 0.0172864981, 0.0034390122, 0.0282748546, -0.0314144865, 0.0117052626, 0.0046237851, -0.0109395673, -0.0127127813, -0.0347280987, -0.0021235358, 0.051363457, -0.0274291448, -0.0853062943, -0.0347207747, 0.0166802648, -0.0051311981, -0.0587451756, -0.0333723798, 0.0972475037, 0.0138819283, -0.0227332041, -0.0846805051, -0.0801815763, 0.0209090058, -0.0111022955, 0.0035846166, -0.0299562197, -0.0020689501, -0.008581643, -0.0056028487, -0.0599806271, -0.012607906, 0.0095936032, -0.0046358067, 0.0796269402, -0.0481031202, -0.0005769819, -0.0047094319, 0.0045084385, 0.0338040404, 0.0360222794, 0.056781996, 0.0194252357, 0.0009752513, -0.0031751022, 0.0189111177, 0.0042827083, 0.052926112, -0.0169847924, -0.0326450244, 0.0663026124, -0.0461758152, -0.0076637929, -0.0397974364, 0.0119449785, 0.0299002603, 0.0021551352, 0.0439159013, 0.0228914544, 0.0035011228, 0.0205321703, -0.015378844, 0.0842347518, -0.0141050452, -0.0330536179, -0.0767917559, -0.003283106, -0.0579739027, -0.0262410603, 0.0113115367, -0.0389820747, -0.0186853483, -0.0455603302, -0.0226521231, -0.0318535417, 0.0270961504, -0.0019628133, 0.0285815652, 0.0309406873, -0.00586322, 0.0162537694, -0.019494161, 0.0000467157, -0.0162629355, -0.0500896983, -0.0161798391, 0.0636864305, 0.032088507, -0.086316891, -0.0050763683, -0.0244932976, -0.0049821609, 0.0045875991, 0.0037511601, 0.0151641527, 0.0226185061, -0.0120579191, -0.0132770063, 0.0280597489, 0.0551772565, -0.0429316685, -0.0230109822, -0.0096034063, 0.0270340797, -0.0622362643, -0.045988854, 0.0273485184, 0.0319167301, -0.067556195, 0.0508699678, 0.0347840786, 0.0458576381, -0.0057946858, 0.013377727, 0.0284787454, 0.0497196838, 0.0007073157, -0.083690688, 0.0015713032, 0.0363774411, 0.0240682829, -0.0122500351, 0.0095018577, 0.022283379, 0.0072730291, -0.0666389391, -0.0157726277, 0.0585283488, -0.0427381657, -0.0031387075, -0.0140693402, -0.0543733463, 0.0255926438, -0.0020913281, -0.0159217771, 0.0218937974, 0.0254975036, -0.0457859077, -0.0095768496, -0.0367404744, 0.0386646017, -0.0451524444, 0.0332725421, -0.0563449673, -0.0361782163, 0.0275201481, 0.0081770141, 0.0403421633, -0.0109527921, 0.0188960694, -0.046194315, 0.0307385121, -0.0231366586, 0.0666038245, 0.0055496097, 0.0261749402, -0.0150641445, -0.0269100312, -0.0625338554, -0.0318696648, -0.0178541169, -0.005089913, 0.0646046773, 0.0133626619, -0.0012928639, -0.0098429043, -0.004760941, -0.0485372879, 0.0538172312, -0.0186217111, -0.0144268218, 0.0189935192, 0.0190158077, -0.0019055988, 0.0116596678, -0.0276028384, -0.0158119127, 0.0143603133, -0.0420268029, -0.0311776791, 0.02915271, 0.0253208075, -0.0176597573, 0.0332708918, -0.0245108232, -0.0417215191, 0.0275692586, -0.0362099819, 0.0150448363, -0.033123143, -0.0083144577, 0.0375332199, 0.0106184324, 0.0314929299, -0.0190653913, -0.0317259394, 0.035192892, 0.0598068833, -0.009855995, -0.0447336696, 0.0286869965, 0.0262778495, 0.0260068327, 0.0262698904, -0.0686683506, 0.0818478614, -0.0027740877, 0.0492419861, -0.0436101481, -0.0443302467, -0.0354634449, 0.0823945254, 0.0097712912, 0.0062527708, 0.0127125168, -0.0358657241, -0.0234490708, -0.0682533681, 0.005525582, -0.0025912728, 0.0300128423, -0.0564547107, 0.0164552107, -0.0117834602, -0.0206738971, -0.0444141068, -0.006959354, -0.0408039354, -0.0044577578, -0.0487693362, -0.0117318379, -0.0010041419, -0.0099250032, -0.0181117672, 0.0264489092, -0.0178451352, -0.026763238, 0.0066200676, -0.028231591, 0.0339804664, -0.0029366233, -0.0379835777, 0.0477743074, -0.0072710724, 0.0441151373, 0.0619928204, -0.0247488283, -0.0647976473, 0.0115926852, -0.0106119355, -0.0135110458, -0.007825518, 0.0487758182, 0.0047637532, -0.0588209219, -0.0034041794, 0.0884085372, 0.0082562473, 0.057801228, 0.0005039798, 0.020439364, -0.0037088285, -0.0091176713, 0.0169287454, -0.0999218822, -0.0601296686, -0.0049108099, 0.0550436676, -0.0071071805, -0.0528848693, -0.0538007468, -0.0913310871, 0.0243168455, 0.0159244891, 0.0080974689, -0.060964372, -0.0449895822, -0.0211094934, -0.0402789228, 0.0141957207, -0.0261493307, 0.028206652, -0.0066169263, 0.0023880443, -0.0476742275, -0.0409160741, 0.0007365241, -0.0048501613, 0.0140413893, 0.0398504697, 0.0522925779, 0.0148186134, 0.0215338524, 0.0259789005, -0.0255479887, -0.0080287252, -0.0208689701, 0.0569091216, 0.0221371166, 0.0374798961, 0.0184085891, 0.0756396577, -0.0431543738, -0.0328147635, 0.046184618, 0.0590120591, -0.1025707647, 0.0028904385, -0.0335131884, -0.0386462323, 0.0468575284, 0.0668923333, -0.0384168774, 0.0027587181, -0.0577099919, -0.036177814, -0.0226005428, -0.0234153308, -0.0328452848, 0.0187896676, 0.0002089736, -0.0531793907, -0.0258781575, -0.0220740847, 0.0087727476, 0.0033046897, 0.0301752165, 0.0475698858, 0.0300779976, -0.004249224, -0.0553471372, -0.0138485376, -0.0133975269, 0.0370321795, -0.0112522515, 0.0258625001, 0.0101356544, -0.051255852, -0.0341581404, 0.0332260877, -0.0947687402, -0.0478355512, -0.0118051516, 0.0034953363, 0.0256850813, 0.0347429253, -0.038414631, -0.0232328083, 0.0314414538, 0.0760279596, -0.0536186658, -0.0834002569, 0.0498580039, -0.009723871, 0.0181608386, 0.0424923636, -0.0616274811, 0.0030600491, 0.0007354545, -0.0341473669, -0.0482057594, 0.0127276452, 0.0534042045, 0.0397759825, 0.0772907212, 0.038572073, -0.0030866098, -0.0076798578, 0.0281732194, -0.0220113583, -0.0197023544, -0.0058229361, 0.034294337, -0.0024336462, 0.0184693001, -0.0102309324, -0.0106388079, 0.0077586486, -0.004956631, -0.0270998068, -0.008632767, 0.04132808, 0.0392964371, 0.0411472246, -0.042409949, 0.0230724066, 0.0469840877, 0.0272847489, 0.0383513011, -0.0422488078, -0.016394496, 0.0305754878, 0.0058016446, -0.0472090803, 0.031224167, -0.0463276692, -0.0306639206, 0.0531387553, -0.0020962541, 0.0186755266, 0.0054361289, 0.0152099868, -0.0554585233, 0.0263957437, -0.0248083305, 0.0493054613, -0.0031044453, 0.013650924, 0.0321259126, 0.0063252193, -0.0502865948, 0.0049177213, 0.0486564264, 0.0171073303, 0.0378072299, 0.0192466583, -0.0756073147, -0.0556061007, -0.0165629201, 0.0046421764, 0.0107226539, -0.0105129322, 0.0081341295, -0.035083048, -0.0187147316, 0.0081592062, -0.0041140984, 0.0015301746, -0.0009766473, 0.0430853702, 0.0687700734, -0.0005486364, -0.0203301013, -0.0505071655, -0.0124780405, -0.0281564873, -0.0550276898, -0.0004080033, -5.538227023e-33, 0.0215802267, -0.0237337276, 0.0272295866, 0.0176647734, 0.018200865, -0.019343609, 0.0044506369, -0.0487771891, -0.0317955464, -0.0197467152, -0.0299077649, 0.0457449183, 0.0158467088, 0.0376289226, -0.0132221477, -0.051089853, 0.0358213969, -0.0155434022, 0.0041459529, 0.0165648125, 0.0706337392, -0.00411072, 0.0498862229, -0.0744038522, 0.0272783134, -0.001106517, -0.0349949412, 0.0338901952, -0.051050093, 0.0329038836, 0.0022845464, 0.062947087, 0.0235721841, 0.0138838012, -0.0223116912, 0.0076272148, -0.0313741043, -0.0280390121, -0.0235369671, -0.0050315675, -0.0005752313, -0.0619244687, 0.0581642799, 0.0034516528, -0.0318803489, -0.0029906654, 0.0357195251, -0.0371706486, 0.0203033052, -0.0297907591, -0.0043681101, 0.0351830348, -0.0496269129, 0.0152242957, 0.0053885523, 0.0056986054, 0.0299666803, -0.0009540929, -0.0859054774, 0.03424979, 0.0207576044, 0.0472262576, 0.0235309955, -0.0169737618, 0.0010499819, 0.0099097891, -0.0013285441, -0.0534945652, 0.0118469298, 0.0534842834, -0.0174765401, 0.0637344867, 0.013957236, 0.041505307, 0.0439158492, -0.0427280255, -0.1307906955, 0.0473757088, -0.0163567998, 0.0216315072, 0.0270770937, -0.0252134837, -0.0453525707, -0.0220076703, -0.0020845735, -0.013515438, -0.0088482676, 0.008208557, 0.0084930565, 0.0000905138, 0.0188841242, -0.0158680771, -0.0134700844, -0.011869804, -0.0431006774, 0.0329205543, -0.0333370268, -0.0586166233, 0.0197353624, 0.0417003334, -0.0205089133, -0.0223985761, 0.0258545261, -0.0038041414, 0.034072753, -0.0130291209, 0.0399524271, 0.0794264823, -0.0610436127, -0.019179035, 0.0016733084, 0.0675752237, -0.0544784032, 0.0246493071, 0.0145158535, -0.0033376298, 0.0451476015, 0.0653215349, -0.0308207832, -0.0401905626, -0.0043135672, 0.0338770598, 0.0010064841, 0.0099713337, 0.0046702251, 0.0082743149, -0.0218329318, -0.0119365854, 0.0528991073, -0.0338885114, -0.0046169832, 0.0222983379, 2.67e-7, 0.0406157076, 0.0105257723, 0.0117743472, 0.0759116709, 0.0237420369, 0.0874878019, -0.0112671405, 0.0660476685, -0.0024577181, 0.0175352674, 0.0240270812, -0.0215927418, 0.0546423681, 0.0378421135, -0.0436945483, -0.0100076338, -0.0336586721, -0.0173646137, -0.076979503, 0.0034353368, 0.1087805927, 0.1130377576, 0.0098885763, -0.002885276, 0.0279642828, -0.0728647634, -0.012530827, -0.0115898466, 0.009276229, 0.008246284, 0.0528858788, 0.030324392, 0.0161842816, 0.0543175191, -0.0490781143, 0.0266903024, -0.0641676262, 0.033287067, 0.0165531971, -0.0058780247, -0.0260076635, -0.0670037344, 0.0233129412, -0.0166285206, 0.0914152712, -0.020114867, -0.0117439386, 0.0545173101, 0.0187441632, -0.0352568403, -0.0062337066, 0.0051134918, -0.0239746682, -0.0541895777, 0.0202487502, 0.0023829786, 0.0331735499, -0.0594314486, -0.0001825032, 0.0095132152, -0.0260596406, 0.0193822216, -0.0250611454, 0.031519793, 0.0469944514, -0.0023920075, 0.0023906645, 2.435380242e-34, -0.0120948935, -0.0021422787, -0.0174072701, -0.0220729541, 0.0331128724, 0.0162445791, 0.0013802639, -0.0144805107, 0.016945621, -0.0434067361, -0.0288989004 ]
http://arxiv.org/abs/2406.17566v1
http://arxiv.org/pdf/2406.17566v1
FrenchToxicityPrompts: a Large Benchmark for Evaluating and Mitigating Toxicity in French Texts
2024-06-25T14:02:11
2024-06-25T14:02:11
Caroline Brun, Vassilina Nikoulina
cs.CL
cs.CL
Large language models (LLMs) are increasingly popular but are also prone to generating bias, toxic or harmful language, which can have detrimental effects on individuals and communities. Although most efforts is put to assess and mitigate toxicity in generated content, it is primarily concentrated on English, while it's essential to consider other languages as well. For addressing this issue, we create and release FrenchToxicityPrompts, a dataset of 50K naturally occurring French prompts and their continuations, annotated with toxicity scores from a widely used toxicity classifier. We evaluate 14 different models from four prevalent open-sourced families of LLMs against our dataset to assess their potential toxicity across various dimensions. We hope that our contribution will foster future research on toxicity detection and mitigation beyond Englis
FrenchToxicityPrompts: a Large Benchmark for Evaluating and Mitigating Toxicity in French Texts: Large language models (LLMs) are increasingly popular but are also prone to generating bias, toxic or harmful language, which can have detrimental effects on individuals and communities. Although most efforts is put to assess and mitigate toxicity in generated content, it is primarily concentrated on English, while it's essential to consider other languages as well. For addressing this issue, we create and release FrenchToxicityPrompts, a dataset of 50K naturally occurring French prompts and their continuations, annotated with toxicity scores from a widely used toxicity classifier. We evaluate 14 different models from four prevalent open-sourced families of LLMs against our dataset to assess their potential toxicity across various dimensions. We hope that our contribution will foster future research on toxicity detection and mitigation beyond Englis
[ 0.0238831881, 0.0805670917, -0.0239410903, 0.0213706624, 0.0065387967, 0.0279656071, 0.0190036204, 0.016264027, 0.0576450713, -0.0327187292, -0.016068127, 0.0070245517, -0.0272108223, -0.0009474151, 0.0392999388, 0.005511885, 0.0530962348, 0.010543474, -0.0113281338, -0.0169098955, 0.0502894334, 0.0494557321, 0.0350458883, 0.0739874169, -0.021789046, -0.0228740703, 0.0623888634, 0.0209639762, 0.037294548, -0.0500525832, 0.0062277406, 0.034152627, -0.0442352518, 0.0455260016, 0.0000018137, -0.0221001767, 0.011324835, 0.0023293267, 0.0045083743, -0.0151853496, 0.0843210071, -0.0168582462, -0.0272402614, 0.0405861996, -0.0044521624, 0.0014972088, 0.0562102906, -0.0023618718, 0.0561106727, 0.0485026538, -0.027092414, 0.0172131862, -0.0122291381, 0.0035758161, 0.0150865037, 0.0474986359, 0.0677750334, -0.0569194965, -0.0092169885, -0.0167807043, -0.0455536917, 0.0290068332, 0.0166371241, -0.034078449, -0.0390972868, -0.0203490369, 0.0114408489, -0.0599271394, 0.0157315899, 0.0717734918, -0.017232256, -0.024445558, -0.0126349833, 0.0070991656, 0.0111431936, 0.0044110185, -0.0211794544, 0.0160091389, 0.0229295846, 0.0044513196, 0.0330238082, 0.005628984, 0.0342244357, 0.0249583274, 0.0326460339, 0.0451917164, 0.0195805617, -0.020052297, 0.0518893749, -0.0152668841, -0.0270847362, -0.0130119519, 0.0209859218, -0.0422271453, -0.0380127914, 0.0238703955, -0.0280046295, -0.0390676856, 0.0596265048, -0.0973384082, 0.0044688983, 0.0097161559, -0.0606461726, 0.0130932583, 0.0061678933, -0.0269234031, 0.0004848116, 0.0112613337, -0.0266095158, 0.0284880064, 0.021900991, -0.0004208546, 0.0015777305, -0.0429848246, 0.0100555494, 0.0386833511, -0.012811874, 0.0043257447, -0.0507071763, -0.0040868348, 0.0322037153, 0.0125163877, 0.011158146, 0.0230381154, -0.0006044963, 0.0215906724, -0.0359363519, 0.0258886684, -0.0160656944, -0.0772074535, 0.0178542286, 0.0132409381, -0.0484339669, -0.0449158028, 0.0078660641, 0.1001277044, 0.0222459659, 0.0061325147, -0.0161382928, -0.0338506997, -0.0646478832, -0.034926679, -0.0551491305, -0.0697267801, 0.0035911149, -0.0320219547, 0.0117674917, -0.0937856957, 0.0266160071, 0.0264373943, -0.0785806626, 0.030316554, -0.0484676287, -0.0123949144, 0.0405644774, 0.0386713445, -0.0384984612, 0.0682226643, 0.0050196983, 0.0006273434, 0.0232440699, -0.0202527791, -0.0301644802, 0.0777862817, 0.0162073709, 0.0448942408, -0.0341742486, -0.0098734144, 0.007520359, -0.0334808715, 0.0159760658, 0.0719193071, 0.0032126086, -0.014910589, 0.0828988701, -0.0045526982, -0.0448903516, 0.026526155, 0.0269647744, 0.0081635481, -0.0131054437, -0.0221567713, -0.0565603971, -0.0385049805, -0.0475912802, -0.056688711, 0.0287737083, 0.0249238908, 0.0157544557, -0.051372733, -0.0063851452, 0.0293073393, 0.0630070418, -0.0082924357, 0.0167878326, 0.0093393726, -0.0071974802, -0.0050210967, 0.0617730692, -0.0751546398, 0.0291813239, 0.0455746762, 0.0358123407, 0.0823052675, 0.0444686413, -0.0160234962, 0.0064796126, -0.0596256368, 0.003230491, 0.0116394442, 0.0007243276, -0.0180123616, 0.0026792993, -0.0102378447, -0.034705244, 0.0368075036, 0.0075353333, -0.0170313567, -0.0205412004, 0.0682438239, -0.0224788245, -0.0054530757, -0.0912386626, 0.0361821093, -0.0214930382, -0.0552461147, -0.0174217224, 0.0237244107, 0.0019565711, 0.0023581318, -0.0344571061, -0.0119185308, 0.0204135161, -0.0227113366, 0.0021410726, 0.0166408308, -0.0251059271, -0.015870491, 0.0398951136, -0.0390094109, -0.0401839763, 0.0369988941, -0.0113908192, 0.0148758162, 0.0114575215, -0.0445545577, -0.0050620562, -0.0063083, 0.0046121147, -0.0106359078, 0.0263426136, -0.0198269971, 0.0368642285, -0.021120904, 0.0378888994, 0.0191145912, -0.0366873816, -0.0128476769, -0.0171758346, -0.072937578, -0.0182115436, -0.024317896, 0.0152858077, 0.046165552, 0.0336286314, -0.0159371812, 0.0139582651, -0.0331808403, -0.0403585024, -0.0240928978, 0.0107739577, 0.020166371, -0.0115969945, -0.0278787073, 0.0056325309, -0.0435812213, -0.0504442006, -0.0092472862, 0.0461939871, -0.0261157267, -0.0121503612, -0.0179189015, -0.0071926676, 0.0044349325, -0.0372360051, 0.0583653152, -0.0128762601, -0.0281209927, 0.0043867393, 0.040141467, -0.0277393181, -0.0270729978, 0.0235457961, -0.0677865595, 0.0118768858, -0.0011035342, 0.0242252722, 0.038829349, 0.0236424319, 0.0427516699, 0.0473794043, 0.0040565911, -0.0297956243, -0.0169839729, -0.0278831441, -0.0271880124, -0.0300998967, -0.1002947167, -0.0222533494, 0.0272765309, 0.0135978498, -0.0043492597, -0.0245159958, 0.067482166, -0.0670204759, -0.0495900922, -0.0424081124, -0.0273641758, -0.017660277, 0.0074692559, -0.0118314857, -0.025295943, 0.041494403, 0.0166634321, 0.0203003231, -0.0005783807, -0.0091471151, 0.0422214568, 0.006900013, 0.0024316858, 0.0543945692, 0.0532905646, -0.0220676139, -0.0187539123, 0.0105951633, 0.0556914471, -0.0077646379, 0.0388093144, -0.0314413756, -0.0256773066, 0.0188578311, -0.067417644, -0.0394950546, -0.0366491415, -0.0152120683, -0.0244186632, 0.0337391347, 0.0142795527, -0.0022429032, -0.017872259, -0.0011887562, -0.0178364236, 0.0018647359, -0.0179738011, 0.0404950604, 0.0007304335, 0.0433731452, -0.0404014774, 0.0553279296, -0.0193354432, 0.0608471744, 0.0953207389, -0.0521931797, -0.006371839, 0.0140939932, 0.0246957298, 0.0248385612, 0.0094032893, -0.0191872679, 0.0070386496, 0.0175741334, -0.0383004509, -0.0278542433, 0.0623413101, -0.0016312988, -0.015314308, 0.0548748784, -0.0063727782, 0.039148137, 0.0191727523, -0.0337992907, -0.099351421, 0.0057638795, 0.0152195469, 0.0051668724, -0.0366683416, -0.0221824497, 0.0417607538, -0.0879073068, 0.0316162072, 0.0221822467, -0.0603267848, 0.0291364603, 0.0011076282, -0.0127390092, -0.0639999509, 0.0198926851, -0.0298196226, 0.0022699565, -0.0406621099, -0.0027098111, -0.0453584827, -0.0469133966, -0.0195931755, -0.0182366148, 0.0350910984, 0.0240499955, -0.0003383678, -0.0309032276, 0.0062408, 0.0535784997, -0.0022922261, -0.0336584039, -0.0097295698, 0.0441444777, 0.0299816169, -0.0244886186, -0.0251347311, 0.0406905487, -0.008028198, -0.0193125755, 0.0307870638, 0.0777771845, 0.0140112117, -0.0005272616, -0.0358498655, 0.0204513092, 0.0395555273, 0.0760076195, 0.041017279, -0.0669370964, -0.0007002715, -0.0752360821, -0.0422965623, -0.0347651988, -0.0812290534, 0.0100255562, -0.0286033507, -0.0788264945, -0.0656603277, -0.0137234945, 0.0188755337, -0.0242023524, 0.0347515717, -0.0054394151, -0.0312288199, 0.005722105, -0.0056652152, 0.0451133028, 0.0256075654, -0.0003111358, -0.0674322471, 0.0109114395, -0.0352571718, 0.0073870881, -0.0145680429, 0.0296620317, -0.0561565645, -0.0860622674, 0.0283703841, 0.0824091211, 0.0086115664, 0.0041443207, -0.003085797, -0.0626507699, 0.0759112835, 0.0482754372, -0.0173665248, -0.0239890702, 0.0352775231, -0.0223708227, 0.024961831, 0.0276256874, -0.0460785218, 0.0262998901, 0.0324932337, 0.0317411348, -0.0465084724, -0.0120331952, 0.0158833098, 0.0331875756, 0.034483999, -0.0614698678, 0.0067477757, -0.0377702713, -0.0193602759, -0.0192923583, -0.0158201158, 0.0234786011, 0.0501511469, 0.0323241241, 0.0031377203, -0.0124892602, -0.0097123561, -0.0002799159, 0.0206713006, -0.0506587029, 0.0314154178, 0.0358572789, 0.0179268606, 0.0666545033, 0.0419992507, 0.0021286823, 0.0267679654, -0.0003628532, 0.0065487972, -0.0253659301, -0.0001286577, 0.021264432, 0.0281632151, 0.0469832756, 0.0322723836, 0.0092347488, -0.0223900899, -0.0338659175, -0.0058388859, 0.0311865974, -0.0347782038, 0.0803817511, -0.0119767766, -0.0216785204, -0.0433790311, -0.016065143, -0.0091148205, -0.0689964071, -0.0075480728, -0.0062792241, -0.0376565121, -0.0282531679, 0.0650673509, -0.0125960791, 0.0474825837, 0.0509709045, -0.0213719551, -0.0163353011, -0.0034855944, 0.0304499492, 0.0077152732, -0.0503030606, -0.0088520758, -0.0671933368, -0.0770881996, -0.011014401, -0.0139940633, 0.0041401689, 0.0158475153, 0.0009894849, -0.0030050918, 0.0107085826, 0.0102063268, -0.0273480993, 0.0390371159, -0.0545031801, -0.0436446331, -0.0042628027, -5.446778705e-33, 0.0351630747, -0.0249142237, -0.0466896147, 0.0242961869, 0.0178127531, -0.0163077749, -0.0023049056, 0.0086515089, -0.0190451927, 0.0610277392, 0.0287274551, -0.0210654754, -0.0001172297, 0.0219481606, 0.0361887291, -0.0318684578, 0.0321714878, 0.0130514326, -0.0347165093, 0.0021778897, 0.0660915673, 0.040383283, 0.0378066339, -0.093477577, 0.0330859013, -0.0185958762, -0.0482119508, -0.0124090137, -0.0303096808, 0.0835867748, 0.0097203981, 0.0979419425, -0.0025154343, -0.0159857906, -0.003491092, 0.0553674661, -0.0300523993, -0.0350409523, 0.0167944767, 0.0106107462, -0.0167010445, -0.055293493, 0.0476222336, -0.0095298411, -0.0443475917, 0.0225361418, 0.0270282719, -0.0523795299, 0.0104706828, -0.0197729766, -0.0665155798, -0.0091660628, 0.0038467713, 0.0005918131, 0.0022545783, 0.1081764549, 0.017653672, -0.0048850807, -0.0453439876, 0.0166511927, -0.033273425, 0.0213319696, 0.0305376314, -0.0346070267, 0.0136189703, 0.0654602721, 0.0036319532, 0.0186531506, -0.0030692499, 0.0226037111, 0.0339301452, 0.0021805034, 0.0614123531, 0.0814832225, 0.0112377387, -0.0373149775, -0.0673648268, 0.0123881968, 0.0279476829, 0.0117206462, -0.0084143355, -0.0396641493, -0.0253804941, -0.0468296148, -0.0332765989, -0.0142817842, -0.0099882642, -0.0168114677, 0.0380733274, -0.0478798784, 0.0045812638, -0.0357314534, 0.0063061472, -0.0321125984, 0.0047982577, 0.0251452569, 0.0081001734, -0.0139228506, 0.0202680361, 0.0447941497, -0.050076846, 0.0402542129, -0.0090816012, -0.0317663141, 0.0320777781, 0.0135472165, 0.0769346282, 0.01854169, -0.0550196506, 0.0149327898, -0.0145674553, 0.0519796945, 0.0401455499, 0.0590081625, 0.0158219244, -0.0345178694, 0.025483381, -0.0051353555, 0.0812277868, 0.0340897813, -0.0110647548, -0.020004971, -0.0150288483, 0.019024035, -0.0318381749, -0.0122194355, -0.0256355293, 0.0317987651, 0.0590741113, -0.0825632438, -0.0206278935, -0.056354709, 2.444e-7, 0.0640242919, 0.0367868282, 0.0074312617, 0.0232415888, 0.0085292738, 0.0213988163, -0.0060974853, 0.0370546095, -0.0132701602, 0.0032511684, -0.0050282595, -0.0279652067, 0.0524063148, 0.0640058741, -0.0637790263, 0.0103954403, -0.0986264348, 0.0239528436, -0.04016684, 0.0017006106, 0.0204903539, 0.0926970765, -0.0057326667, -0.0229335502, -0.0080707297, -0.0277067963, 0.0188030712, -0.0680662915, 0.0056507746, -0.006171097, 0.0364919193, 0.0095848013, 0.0179561786, -0.0106571279, -0.0187898465, -0.0582531169, 0.0127636986, 0.0055525457, -0.0015419984, -0.010135252, 0.0237258058, -0.0657623708, -0.0022184951, -0.0597533323, 0.0354242958, 0.037125431, 0.0208690334, -0.0365787596, -0.0458863899, -0.0326416008, 0.0954639167, -0.0638937876, -0.0314412341, 0.0052791722, 0.0452658571, 0.0243109241, 0.0122228507, -0.0362980589, -0.0335297175, -0.0179325137, -0.0032313096, -0.0648385435, 0.0173906609, 0.0413777754, -0.0682061091, -0.0376515612, 0.0082753981, 1.961263652e-34, 0.0072582299, -0.0281736292, 0.0007212755, -0.028108811, 0.0364546701, -0.0079240222, 0.0145768961, 0.0054314327, -0.0281867757, -0.0242992006, -0.0117317243 ]
http://arxiv.org/abs/2406.17563v1
http://arxiv.org/pdf/2406.17563v1
Multi-property Steering of Large Language Models with Dynamic Activation Composition
2024-06-25T14:00:42
2024-06-25T14:00:42
Daniel Scalena, Gabriele Sarti, Malvina Nissim
cs.CL
cs.CL, cs.AI, cs.LG
Activation steering methods were shown to be effective in conditioning language model generation by additively intervening over models' intermediate representations. However, the evaluation of these techniques has so far been limited to single conditioning properties and synthetic settings. In this work, we conduct a comprehensive evaluation of various activation steering strategies, highlighting the property-dependent nature of optimal parameters to ensure a robust effect throughout generation. To address this issue, we propose Dynamic Activation Composition, an information-theoretic approach to modulate the steering intensity of one or more properties throughout generation. Our experiments on multi-property steering show that our method successfully maintains high conditioning while minimizing the impact of conditioning on generation fluency.
Multi-property Steering of Large Language Models with Dynamic Activation Composition: Activation steering methods were shown to be effective in conditioning language model generation by additively intervening over models' intermediate representations. However, the evaluation of these techniques has so far been limited to single conditioning properties and synthetic settings. In this work, we conduct a comprehensive evaluation of various activation steering strategies, highlighting the property-dependent nature of optimal parameters to ensure a robust effect throughout generation. To address this issue, we propose Dynamic Activation Composition, an information-theoretic approach to modulate the steering intensity of one or more properties throughout generation. Our experiments on multi-property steering show that our method successfully maintains high conditioning while minimizing the impact of conditioning on generation fluency.
[ -0.0118808877, 0.0510781184, -0.0105659608, -0.0085758669, -0.0475663766, -0.0008930812, -0.029272547, -0.0238879751, -0.0428166017, -0.0480308793, -0.0086747538, -0.0837073326, -0.013752535, 0.0119614331, 0.041306451, -0.0565765128, 0.0707159042, -0.0515489131, -0.0537863374, -0.0294297524, -0.0153108807, -0.0044275038, -0.0131000876, 0.0354350507, -0.0212831683, -0.0056855856, -0.0117568308, 0.0052096602, -0.0083844, -0.0351501629, -0.0074237785, 0.0573741496, 0.019163087, -0.0191582162, 0.0000020137, -0.0519807152, -0.0251589064, -0.0068437993, 0.0192664284, 0.0170147344, 0.0827391446, -0.005095101, -0.0315675102, 0.0225589387, -0.0268171802, -0.0019674543, 0.0604700372, 0.0344399959, 0.0206157733, 0.0553181395, -0.0147251757, -0.037972562, -0.0014846264, 0.0006152036, 0.001213375, 0.0018786736, 0.026574038, 0.0311028622, -0.0164022949, -0.018462237, -0.029118102, 0.0123287411, 0.0356709212, 0.0013724782, -0.0146177756, 0.0381614044, 0.009138572, -0.0126668941, -0.0276679415, 0.0681514144, -0.0211737417, -0.0219744742, -0.0075542624, 0.0008593245, -0.0087667974, -0.0257333722, -0.0145668508, -0.010486356, 0.0186431408, 0.0346318483, 0.0312547982, 0.0043172473, 0.0109462254, 0.0106049823, -0.0512047745, 0.0591639206, 0.0522791632, -0.0240908973, -0.0470318235, -0.0036765065, -0.0005350685, -0.0360189043, -0.0474594571, 0.0113522727, -0.0112182219, 0.0283488147, -0.0705499575, -0.0236606281, 0.0154358475, -0.0479101688, 0.0432375818, 0.0457392372, 0.0064070625, 0.0056796661, -0.0334070995, 0.0573165491, -0.0379275233, 0.0169420149, -0.0450750664, 0.0145823155, -0.0443545803, -0.0062635504, -0.0441160314, 0.0403505489, -0.006972834, -0.0201984458, -0.0703709573, 0.0389739163, 0.0111024063, -0.0471601114, -0.0333186202, -0.0056782141, -0.0538690425, 0.0103315394, 0.0211259443, 0.0096703721, -0.0406909063, 0.0012274577, 0.0043871994, -0.0440975875, -0.01842352, 0.037656229, 0.0032687066, 0.0074425456, -0.0189459808, 0.0759246126, -0.0331803747, -0.026352413, -0.064470388, -0.0539743826, 0.0043392023, -0.0460453629, 0.0199087784, -0.0506024137, 0.0127699198, 0.0065691597, -0.0233072061, -0.0741674602, -0.0030391254, -0.0002625515, -0.0189078879, 0.0273295473, 0.0201990828, 0.0135976728, -0.0174558498, -0.0100806961, -0.0117708892, 0.0988265723, 0.0421073847, 0.0053922483, -0.0124427341, 0.0160676315, -0.0140066231, 0.0501408055, 0.0359754339, 0.0431751423, -0.0763537064, -0.052109655, 0.0373316593, -0.0058470936, 0.0494552366, 0.0606762134, -0.0565885231, 0.0396755636, 0.0519433282, -0.0317128748, 0.0295560565, 0.0337552503, 0.1000781432, -0.0119614992, -0.0338671096, 0.0670948476, -0.0437414125, 0.000200507, -0.0488209948, 0.0505305901, -0.0036614726, 0.0772085413, -0.0208300967, -0.0361493751, 0.0012947208, -0.0133192427, 0.0484831221, -0.0545750037, 0.0291213561, 0.0520246327, -0.0669428632, 0.0228000451, 0.0514833443, -0.0413284637, 0.0240649153, -0.030559117, 0.0141073233, 0.0766869113, 0.0466700532, 0.0304299109, -0.016379578, 0.0370790958, 0.0280179959, 0.0452533104, 0.027886549, 0.0001547981, 0.0157811008, -0.0880555809, -0.0217373986, -0.0042725219, 0.0371700786, 0.0422389917, -0.0018195609, -0.0058858367, 0.0073975804, -0.0015875255, 0.0089616468, 0.0370113961, -0.0099018291, -0.0116495509, -0.0124107543, -0.0145850899, 0.0142798787, -0.0267630294, -0.0369632728, 0.0280551761, -0.0050293608, -0.0362795442, -0.0150604816, 0.0532074198, -0.0593285561, -0.0320721008, 0.0132723078, 0.046980232, 0.0351256914, -0.0691404566, 0.0046733469, -0.0471115187, 0.0607700273, -0.0240226686, 0.067900002, -0.0288303886, 0.0425293706, 0.0118197305, 0.024334792, -0.0247484203, 0.0666676909, -0.0203767233, -0.0261267573, 0.064357765, 0.0352622345, 0.0022509706, -0.0339458063, 0.0224300101, 0.0237759855, 0.0040602945, -0.0085722283, -0.0047253533, 0.0185869951, 0.0042216829, -0.0484641008, -0.0301306956, -0.02820554, -0.0359142758, 0.0239842907, 0.0293561015, 0.0153072234, 0.0109412856, 0.0391190127, -0.0531721674, -0.0164645184, -0.0092130024, 0.023368312, 0.0029424066, -0.0119748963, -0.0696305111, -0.029872587, 0.0003524603, -0.011662744, 0.0505869426, 0.0086315563, 0.0207065661, -0.0567503907, -0.0105835451, -0.0246789698, 0.0523771085, 0.0184054598, -0.0516182929, -0.0116756335, -0.0078636166, 0.088231951, 0.0078877611, 0.0063059768, -0.0102294078, 0.0677886829, -0.0209535565, -0.0534374677, 0.0061158054, -0.0576080829, 0.0283827297, -0.0336827412, -0.0187406968, 0.0048735277, 0.0847427845, 0.0215014331, 0.0025718836, -0.0901835412, -0.0008175813, -0.0299885701, -0.0568455681, 0.0387110785, -0.0350749083, -0.0220702775, 0.038000349, -0.0187070183, -0.0566289425, 0.0875502005, -0.0669336617, -0.0090054078, -0.0556634068, 0.0377418995, 0.0003416736, 0.007093004, 0.0124754095, 0.0353846438, 0.0516357645, 0.0145749599, -0.0627636239, -0.0302601457, 0.0362084135, -0.0079732426, 0.0201408137, -0.035135895, 0.0224526674, 0.0359657966, -0.0164516792, -0.0554257706, -0.0476392694, 0.0369159691, 0.0183452945, 0.0587583259, -0.0354954936, 0.0183162708, 0.0278465301, 0.0068042567, 0.031271819, -0.0051678238, -0.0076465486, -0.0254304092, 0.0222882275, 0.0610887557, 0.0008931627, 0.0197256319, -0.0380128436, 0.0194524713, 0.0115957046, 0.0461746603, 0.0240325332, -0.0311272945, 0.0348991714, 0.0236968379, 0.0058225463, -0.0157585945, 0.027958788, 0.0281762406, -0.0358634479, -0.031372834, 0.0861331746, 0.0318209939, -0.0346023142, 0.0396252908, -0.0047193272, 0.0980108902, 0.0130145065, 0.0159065668, -0.0817662179, -0.0100342212, 0.0080291815, 0.0153374178, -0.0250983369, 0.0170487165, -0.0236849692, -0.1076600403, 0.0376732163, 0.037066754, -0.013810344, -0.0276054572, -0.0175239965, -0.0112856273, 0.017404899, 0.0328007899, 0.0220782124, -0.0104104234, -0.0120561309, 0.0093502486, 0.0188123714, -0.0659562796, -0.0161089618, -0.0164028369, 0.0447651222, 0.1165239364, -0.0158958677, 0.0244404934, -0.0467365794, 0.0506537035, -0.0141327595, 0.0251566563, -0.0017093533, 0.0186183378, 0.0180437341, -0.0014102079, 0.00205099, 0.0408533104, 0.0042145494, -0.0315285325, -0.0578781292, 0.0379526131, -0.0322692432, -0.018853141, -0.0290112458, 0.0241899677, 0.0400135927, 0.0587361604, 0.0418683812, -0.0641885102, 0.0061828583, 0.001805159, -0.0090216137, -0.0763639882, -0.0365541466, -0.0126843555, -0.0190349612, -0.0321464762, -0.1128859073, -0.0195214245, 0.0119614536, 0.0394257568, 0.0055340752, -0.0171549022, -0.0008496013, -0.0469416156, -0.0012570252, 0.0105858687, -0.0233858097, 0.006871623, -0.0446216241, -0.0442891009, 0.0594560318, -0.0387173854, 0.0078053591, 0.026723437, -0.0258965101, -0.1113152504, 0.0179424994, 0.0269869138, 0.0181915145, 0.0756144449, -0.0282244571, -0.0254671127, 0.0277167615, 0.0067001772, 0.0046607223, 0.033195734, 0.0325850919, -0.0204828903, 0.0058492124, 0.0340297446, -0.0365659706, 0.0527007133, 0.0067152469, 0.0417122245, -0.0125047844, 0.0191362705, 0.0396006666, 0.0044221291, -0.0110474769, -0.0578678995, 0.0377138443, -0.031949345, 0.0066275401, -0.0272539798, -0.0181522835, -0.0548314005, 0.0252887607, 0.0092833992, 0.0071472465, -0.0013213266, -0.0250710379, 0.0100747682, 0.0279103667, 0.0047617918, -0.0050079697, -0.0464688651, 0.0023847711, 0.0376028828, 0.0111820735, 0.0124890078, 0.0574550517, -0.0543202981, 0.006058332, -0.0311434697, 0.0501761325, 0.0436770804, -0.0097019924, -0.070834823, -0.000590413, -0.0081463745, 0.0149298022, -0.0571803749, 0.0148056569, 0.0572745986, 0.016645696, 0.1053875536, -0.0432490669, -0.0055282726, 0.0135585014, 0.0315690376, 0.0291534401, -0.0422929786, 0.0239821989, 0.0213570558, -0.0277578421, -0.0115622105, 0.0810421705, -0.0147196976, 0.0702209398, -0.0589933135, 0.0482192449, -0.0086357761, -0.0248501915, -0.0120887635, 0.0090591572, -0.0142500121, 0.0190919377, -0.0802673027, -0.0393838808, -0.0038230231, -0.0026394462, 0.0127961971, -0.0168977287, 0.0721837729, 0.0315717384, -0.0291448981, -0.0036356917, -0.0029684594, -0.0736541077, 0.0243067816, -0.0035519921, -0.0339091457, -6.500311109e-33, -0.0512352921, 0.0202901494, 0.0190215837, -0.0160940923, -0.0249390434, -0.0208291691, -0.0357268229, 0.0042767571, -0.0059182206, 0.010696101, 0.0234411303, -0.0211316943, 0.0170762427, 0.0169721711, -0.0048907897, 0.0166466478, 0.0602633022, -0.0280045122, -0.0496627502, 0.0225185156, 0.0336621813, -0.0103498679, 0.0223212615, -0.0725048333, -0.0151588209, 0.0015499108, -0.047512047, 0.0161986016, -0.0424516574, 0.0620498769, -0.0080781383, 0.020354433, 0.015713498, -0.0575917475, -0.0534170642, -0.021385774, -0.0268356297, -0.0445967317, 0.0228062291, -0.0052202535, -0.0620357841, -0.0308636408, 0.0514908843, -0.0141277341, -0.0075747948, 0.0138769401, 0.0108916713, -0.0190395974, -0.0157670546, -0.0475150868, -0.1047558561, 0.0216781851, 0.001886308, -0.0410869122, 0.018141035, 0.0418839157, -0.0086761713, -0.006599023, -0.0232054796, 0.0300017707, 0.0182608608, 0.0424048267, 0.0548556596, 0.0436707363, 0.0290777292, 0.0346005857, 0.0279117953, 0.0115389349, -0.0289203841, -0.0034149806, 0.0579321086, 0.00192403, 0.0073755109, 0.0214056671, 0.003700933, -0.0212574992, -0.0221961495, -0.0161139369, 0.0328347646, 0.0101850731, 0.0300809648, 0.035567347, -0.0352536552, -0.0364506021, -0.0802070126, -0.0238343123, -0.0164698884, -0.0561417341, 0.0337034799, 0.0240811389, -0.0295268577, -0.0383961387, 0.0590934753, 0.0096690711, -0.019072352, 0.0432015322, -0.0589169078, 0.0382781588, -0.0406124704, 0.0008041319, -0.0126108099, -0.0440172926, 0.0441635847, -0.016913794, -0.0155543536, 0.0250101183, 0.0035543987, 0.0557210259, -0.0407562703, -0.0033745961, -0.0208242889, 0.0225609671, 0.0381735973, 0.0096706571, 0.0373070538, 0.0079006022, -0.0140227647, 0.0290718004, 0.0453117453, 0.0169208683, -0.0087383408, -0.014579081, 0.0135075096, 0.0795240998, -0.0461947992, -0.0019937588, -0.0434632562, -0.0158888008, 0.0275626462, -0.0266446248, -0.0447130799, 0.0294080637, 2.793e-7, 0.024250878, 0.0326474011, 0.0385598727, 0.0568104461, -0.0065633864, 0.0063556177, 0.0221515391, -0.0102489758, 0.0151771093, -0.0203987565, -0.0125681739, 0.0448026545, 0.0250760429, -0.004816893, -0.0762507319, 0.0214639753, -0.0318217091, 0.0181286931, -0.0259845089, -0.0390883349, 0.0511135124, 0.0829076618, -0.0179404318, -0.0053497609, 0.0297818463, -0.0053301812, 0.0594900884, 0.015002735, 0.0064760577, 0.055846747, -0.0061836606, -0.000296099, -0.0050119143, 0.0196583346, -0.0250980239, 0.0292640738, -0.0479949452, 0.0262927189, -0.0277249441, -0.0472538061, 0.0329728685, -0.0131590143, -0.0377372615, -0.0144915925, 0.0500735082, -0.0632297993, -0.0144410776, -0.0152677717, -0.0124187618, 0.0263649672, -0.0157331042, -0.0095103066, -0.0341348313, 0.0410040282, 0.0147279268, 0.038950745, -0.036653962, -0.0436152332, -0.0413718186, 0.0155149642, -0.0300730392, -0.0434869677, -0.0166527908, -0.0168245099, 0.0064707766, -0.0573908463, -0.0354995281, 3.063148179e-34, 0.0435798913, -0.0275068115, 0.0376943983, 0.0693662986, 0.0313715376, -0.0117760338, 0.0372329988, 0.0191744715, 0.0356843062, -0.073186025, -0.0151990382 ]
http://arxiv.org/abs/2406.17557v1
http://arxiv.org/pdf/2406.17557v1
The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
2024-06-25T13:50:56
2024-06-25T13:50:56
Guilherme Penedo, Hynek Kydlíček, Loubna Ben allal, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro Von Werra, Thomas Wolf
cs.CL
cs.CL
The performance of a large language model (LLM) depends heavily on the quality and size of its pretraining dataset. However, the pretraining datasets for state-of-the-art open LLMs like Llama 3 and Mixtral are not publicly available and very little is known about how they were created. In this work, we introduce FineWeb, a 15-trillion token dataset derived from 96 Common Crawl snapshots that produces better-performing LLMs than other open pretraining datasets. To advance the understanding of how best to curate high-quality pretraining datasets, we carefully document and ablate all of the design choices used in FineWeb, including in-depth investigations of deduplication and filtering strategies. In addition, we introduce FineWeb-Edu, a 1.3-trillion token collection of educational text filtered from FineWeb. LLMs pretrained on FineWeb-Edu exhibit dramatically better performance on knowledge- and reasoning-intensive benchmarks like MMLU and ARC. Along with our datasets, we publicly release our data curation codebase and all of the models trained during our ablation experiments.
The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale: The performance of a large language model (LLM) depends heavily on the quality and size of its pretraining dataset. However, the pretraining datasets for state-of-the-art open LLMs like Llama 3 and Mixtral are not publicly available and very little is known about how they were created. In this work, we introduce FineWeb, a 15-trillion token dataset derived from 96 Common Crawl snapshots that produces better-performing LLMs than other open pretraining datasets. To advance the understanding of how best to curate high-quality pretraining datasets, we carefully document and ablate all of the design choices used in FineWeb, including in-depth investigations of deduplication and filtering strategies. In addition, we introduce FineWeb-Edu, a 1.3-trillion token collection of educational text filtered from FineWeb. LLMs pretrained on FineWeb-Edu exhibit dramatically better performance on knowledge- and reasoning-intensive benchmarks like MMLU and ARC. Along with our datasets, we publicly release our data curation codebase and all of the models trained during our ablation experiments.
[ 0.0425819568, 0.0559946746, 0.0001985187, 0.0338833444, -0.0033952531, 0.0160738025, 0.0158506259, 0.0118710091, 0.0004883572, -0.0447154902, -0.0466959402, 0.0206590239, -0.0241232347, 0.0385839045, -0.0106556956, 0.0025936558, 0.023572769, -0.0191273596, 0.004293629, -0.0226322077, 0.0242743436, 0.0251837, 0.0143513242, 0.0064603537, 0.031615939, -0.0161858387, -0.0028835896, 0.0265131053, -0.0164601263, -0.0228681825, 0.0421079062, 0.0414914191, 0.0211625546, 0.0624252185, 0.0000022676, -0.0392809995, -0.0420425683, -0.0185254514, -0.0486188196, 0.0488613248, 0.0191843938, 0.0284037944, 0.009398913, 0.0366826989, -0.0409213044, -0.0160539523, 0.0173361078, 0.0337159894, 0.0623807795, 0.0666938499, -0.004320221, -0.0061543351, 0.0329884179, -0.0121743111, -0.0228374396, -0.0048593101, 0.0380424038, -0.0085980995, 0.0086520985, 0.0157785416, -0.014236833, 0.0259227976, 0.0151735153, -0.0087613668, -0.0168882553, 0.0121843498, -0.0714370757, -0.0290007051, 0.0204358399, 0.1012538075, -0.0142335873, -0.0621451214, 0.0284021012, 0.0075477236, -0.0404598638, -0.0085149137, -0.0450460985, 0.0326603539, 0.0122458227, 0.0590623431, -0.0194224883, -0.0204455871, 0.0500941649, 0.0367562771, -0.0161477961, 0.0382360481, 0.0173061192, -0.0251792502, 0.011499254, 0.0354464166, 0.0212249309, -0.0629286468, 0.0357643031, 0.0003399992, 0.0039405259, 0.0028329496, -0.0827368721, -0.041413866, 0.0566529259, 0.0235416517, 0.0143365711, 0.021758534, -0.0418483652, 0.0698577315, -0.0171592738, 0.0151820574, -0.0061985962, 0.0362696126, -0.0392317176, 0.0089893434, -0.0228356794, 0.0195386726, -0.0268587153, 0.0644639134, -0.0019548144, 0.0029315229, -0.0198898669, -0.0001700269, 0.0022009709, 0.0082649756, -0.014230134, -0.0235978812, -0.0312651582, -0.0093245571, 0.0056156269, -0.0064647798, -0.0300328601, 0.0321265757, 0.0040499354, -0.0722777024, -0.0197147205, 0.0064924872, 0.0440720096, -0.0304797124, 0.041337762, 0.0121521922, -0.0072922423, -0.0410742983, -0.0336502083, -0.0407520942, 0.0138318958, -0.0285145044, -0.0059675868, -0.0799419656, 0.0344010927, 0.0033135081, 0.0240713395, -0.0527119488, -0.0043688454, 0.0244269352, -0.0168111064, 0.0461349785, -0.0637978539, -0.0159766171, -0.0000080562, -0.0048125, 0.0116282692, 0.0556413122, 0.0086803511, 0.0611460395, 0.081687592, 0.015255738, -0.0027468202, 0.0469561778, 0.0530716889, 0.0273829252, 0.0090400372, -0.0023472954, -0.0037688729, 0.0061994987, 0.0433780439, 0.0340545177, -0.0629392415, -0.0222018491, 0.0429609269, 0.0405756421, 0.0127637964, 0.0161039121, -0.0011340466, 0.0348493792, 0.0306509528, -0.0107657742, -0.0397070572, -0.0193450786, -0.0263722837, -0.0139378691, 0.0593547635, 0.0461718217, -0.031801071, -0.0431631245, 0.0060114437, -0.0113939214, 0.0076411278, 0.0291126482, 0.0548358597, -0.0081018126, -0.0167606194, 0.0417535715, 0.0472764708, -0.0596591197, 0.0124645866, -0.0492117889, 0.0130055351, 0.0752735808, 0.0952104256, -0.0437467247, -0.0546802096, 0.0143083353, 0.0760905072, 0.0384267569, 0.0370237865, -0.0195717644, 0.0243558995, -0.0711531863, -0.086302191, -0.0147917941, 0.0022800982, -0.0128307948, 0.0524571426, 0.0138083035, 0.0225695893, -0.0618724376, -0.0285536163, 0.0453064628, -0.0463196449, 0.0004365282, -0.0422017239, -0.0128576681, 0.0052628266, -0.0090675633, -0.0066653695, -0.0154098235, 0.0627030879, -0.0136908954, -0.0000988658, -0.0529853441, -0.0058334633, -0.0171778556, 0.0964859873, -0.0567606427, 0.0123431357, 0.0644510835, -0.0663616061, -0.0292796623, -0.0101844268, 0.0114503009, 0.0474257842, -0.0222454444, 0.0339156762, -0.0271569155, 0.0136930645, -0.0073753111, 0.017964907, 0.0032604279, 0.0032027119, -0.0101425266, -0.0268881805, 0.0355856158, -0.0237665158, -0.0520611629, -0.0157419518, -0.036006555, 0.031367857, 0.0180839747, 0.0294543058, -0.0816438049, -0.0016348561, -0.0832552984, 0.0150583591, -0.0038103457, 0.0235860907, 0.0132528264, -0.002001517, -0.0008602898, -0.021822881, -0.0093522407, -0.0229425244, -0.0164080244, -0.0005111304, 0.0173169039, -0.0143130496, 0.0010400853, -0.0572260804, -0.0540553108, 0.0110203922, 0.0215205122, -0.0111603783, -0.0224255025, 0.0059430175, 0.0421814322, -0.0387791917, 0.0086704381, -0.0205573402, -0.0477058142, 0.007818535, -0.0279982239, -0.0023547148, 0.0402724855, -0.0262503438, -0.0004783364, 0.0590933487, 0.0074857064, -0.0686797574, -0.0042883097, -0.0061848606, 0.0075103785, -0.0438285172, -0.0071894191, 0.0017125971, 0.0262166187, -0.0434939191, -0.0174371544, -0.0212540124, 0.0182166994, -0.0112760738, -0.0149189821, -0.0481885597, 0.0006196303, -0.0083606476, 0.030246716, -0.0301670749, -0.0304783564, 0.0485644974, -0.029727051, 0.0252158418, -0.0423319638, 0.0308716577, -0.0236855429, 0.0187002718, -0.0223829355, 0.007921624, -0.0050933827, 0.0127299782, -0.0317880437, 0.0205615815, 0.0274192598, -0.0749194399, 0.0219777599, -0.0196805317, 0.0294911899, -0.0288963877, -0.0554471798, -0.0325403251, -0.0352577679, -0.034912765, -0.017018564, -0.0028219537, 0.0033048107, -0.0744729191, -0.030542912, 0.0073014153, 0.0609653555, -0.0162029099, -0.0014471562, -0.0105946716, 0.0381240621, 0.051739186, -0.0061946479, 0.0279451739, -0.026125934, -0.0022340622, 0.0966579542, 0.0692931563, -0.0108876815, 0.0075150044, 0.0019642694, 0.0722485259, -0.0392044708, -0.0044101165, -0.015287159, 0.0069518876, -0.0503879897, -0.0151740629, 0.1184980646, 0.0326804891, -0.0747940913, -0.0306168571, 0.0067648916, 0.0202449206, 0.0069513703, -0.0052877539, -0.1019394472, -0.0501156785, 0.0423969813, 0.0425725468, 0.0328984931, -0.0191249307, 0.0151128732, -0.0893964022, 0.0031634099, -0.005599115, 0.0158077348, -0.001479881, -0.0760980546, -0.0018293672, -0.0262322575, 0.0020017123, 0.0160744451, -0.0150749506, 0.0246450752, 0.01977624, 0.0148482136, -0.071278967, 0.0344027504, -0.0326109082, 0.026702242, 0.0105304951, 0.0557503849, -0.0002684678, -0.0284676235, 0.060147129, -0.0383820236, -0.0584932379, -0.0208230074, 0.0752868727, -0.0076269535, 0.0015172376, -0.0285122078, 0.0475810207, -0.0632671639, -0.0034049295, 0.0148799038, 0.0624157749, -0.0093277479, 0.0014133697, -0.0425806642, -0.0390964001, -0.0428430624, 0.0864907652, 0.0184906349, -0.0364462845, -0.0270515196, -0.0222367775, 0.0030435158, -0.0288124122, -0.0133562824, 0.0118347183, -0.044723507, 0.0127234459, -0.0551932901, -0.0270657986, 0.035569381, 0.017915966, 0.0147599764, 0.0059004351, 0.0095565766, -0.0073265131, -0.0677182153, 0.0633794963, -0.0815582871, -0.0041511096, -0.0963924155, -0.0116914222, -0.011686231, -0.0156657659, -0.037611898, 0.0239463132, -0.0814609751, -0.0405794866, 0.0273299739, 0.0042520296, 0.0233177915, 0.0412487201, -0.015262953, -0.07001587, 0.0555608384, 0.0001577651, -0.0454117842, -0.002816987, 0.0297197681, -0.0471278168, 0.0077604265, 0.0385659151, -0.0681350306, 0.0183098167, 0.0408800542, 0.0156601779, -0.0042464281, 0.0237293914, 0.0019694874, 0.0624073409, 0.0563061535, -0.0125617012, -0.0220047403, -0.0496267825, 0.039930176, 0.0025547515, -0.0170355216, -0.0321119241, 0.0346868001, 0.052464094, -0.0083787311, -0.0199844595, 0.0402643085, 0.0093656443, 0.0469816066, -0.0558314919, 0.018880818, -0.0006948857, 0.0269212928, 0.0403627381, -0.034417931, -0.0205420684, 0.0199685395, -0.0055690366, 0.0254089795, -0.0539822765, -0.0188258644, 0.0318575501, 0.0090409191, -0.022722749, 0.0610896423, -0.0468174368, -0.0244688876, -0.0342399999, -0.0129531091, 0.0465780236, -0.0023607397, 0.0742229, -0.0062798616, -0.0244325679, -0.0312239043, 0.0058344053, 0.00577125, 0.000561435, 0.0007071274, 0.0170895196, -0.026175376, 0.0465677232, 0.1005593464, 0.0006727574, 0.0404436588, 0.0474964604, 0.0055733733, -0.0242309682, -0.0109438663, 0.0180404708, 0.0505221561, -0.0083301412, -0.0434861556, -0.0115946047, -0.0439434908, 0.0070697614, 0.0066559566, -0.0321404375, 0.0041879765, 0.0531143919, 0.0707482919, 0.021743983, 0.0031385771, 0.0024530126, 0.0065438165, 0.0033525457, -0.0433928967, 0.0048936657, -6.934017832e-33, 0.0293860249, 0.0076631042, 0.0139163602, -0.0080120517, -0.0177666992, 0.0306202415, -0.021067854, -0.0449300632, -0.0437236652, 0.0067490833, -0.0374035612, 0.0486938655, 0.0194847137, 0.0509156547, -0.0084453775, -0.0026602901, 0.0828161836, -0.0120153427, -0.0589152947, 0.055396717, 0.0247297529, 0.0376136862, 0.0392140672, -0.0280179512, -0.0268796086, -0.021134343, -0.0108071221, -0.0029610021, 0.0157857444, 0.0678668693, -0.0584818088, 0.0456061736, 0.0144983372, -0.0094086211, -0.0158500243, -0.0172639433, -0.0951573998, -0.038458351, -0.0035840797, 0.0132720293, 0.0116347885, -0.0265693683, 0.045096878, -0.0119007491, -0.0084452527, -0.0046578799, 0.0441839546, -0.0129998289, -0.0431111865, -0.0350957997, -0.0153249325, 0.0318863764, -0.0203129705, -0.0112902969, 0.0630390644, 0.000521239, 0.0057005761, 0.0665393993, -0.0628941506, 0.0487824976, 0.0109415213, 0.04509129, 0.0424528457, -0.0370953344, 0.0180703774, 0.070217669, -0.0032951804, 0.0062627443, 0.0486768335, 0.0115899323, 0.066790916, 0.040135514, 0.0230487399, 0.0506884158, 0.0510229655, -0.0217516832, -0.0608133152, 0.0245590135, -0.0063169277, 0.0422008522, 0.0107158776, 0.0036476403, -0.0113476124, -0.0225046668, -0.0442990959, 0.0042613703, -0.0305608902, 0.0092864018, 0.0239005052, 0.0194637161, -0.0669819862, -0.0387919992, -0.0006128995, -0.0530981347, -0.0501324162, 0.0666991696, 0.0006151557, -0.0040271003, 0.0044503827, 0.0042517739, -0.0475049689, -0.0345432349, 0.0271202568, -0.0463192351, 0.009660936, -0.0199977607, 0.0283078961, 0.0284909885, -0.0544294454, 0.0210711733, -0.0310980026, -0.0132383266, 0.0286631603, 0.0184849799, 0.0395966582, 0.0104168374, 0.0219829474, -0.0293845721, 0.0816812962, 0.0206041969, -0.0438794941, -0.0076714279, 0.0031984823, -0.0001622432, -0.0149948932, -0.0125898467, -0.052431684, 0.0165739041, 0.0163928661, -0.1007341892, -0.0134853357, 0.0149202403, 2.858e-7, 0.0315155089, 0.0728706643, 0.0283607431, -0.009699069, 0.0106520914, 0.0198310614, -0.0256769527, 0.0039713387, 0.0092136711, 0.0167411547, 0.00430643, 0.017803641, 0.0499257445, 0.0240751058, -0.0718401074, 0.0358820148, -0.0353670828, -0.0240710359, -0.0531747639, -0.0505333841, -0.0020890192, 0.094120957, 0.0371721573, -0.0468319729, -0.0043221144, -0.0505383313, 0.0111129088, -0.0255491957, 0.0072317747, 0.0296127144, 0.0359011218, 0.0332253128, 0.0071631577, -0.0468742885, -0.0359402075, 0.0145871667, 0.0013374218, 0.0550978221, 0.0096890358, -0.0076218485, 0.0070969416, -0.0472607836, 0.0005033482, -0.071499452, 0.1067776978, 0.0257981922, 0.007694582, -0.0116896508, -0.0615797825, -0.0441614278, 0.0013741052, -0.0264108293, -0.019572936, 0.0240940861, 0.0169304889, -0.0062055001, -0.0307466444, -0.0866890997, -0.0251312181, -0.0157107003, -0.0416684672, -0.0657655969, -0.0474255197, -0.0103462879, 0.0158232804, -0.0444677845, 0.02624272, 3.059242874e-34, 0.0475488566, -0.0281007569, 0.0321423486, 0.0017611708, -0.0045093317, -0.0090106409, -0.0047869319, -0.0171342138, -0.0240312964, -0.0871180147, 0.0219110213 ]
http://arxiv.org/abs/2406.17553v1
http://arxiv.org/pdf/2406.17553v1
Retrieval-Augmented Code Generation for Situated Action Generation: A Case Study on Minecraft
2024-06-25T13:43:24
2024-06-25T13:43:24
Chalamalasetti Kranti, Sherzod Hakimov, David Schlangen
cs.CL
cs.CL
In the Minecraft Collaborative Building Task, two players collaborate: an Architect (A) provides instructions to a Builder (B) to assemble a specified structure using 3D blocks. In this work, we investigate the use of large language models (LLMs) to predict the sequence of actions taken by the Builder. Leveraging LLMs' in-context learning abilities, we use few-shot prompting techniques, that significantly improve performance over baseline methods. Additionally, we present a detailed analysis of the gaps in performance for future work
Retrieval-Augmented Code Generation for Situated Action Generation: A Case Study on Minecraft: In the Minecraft Collaborative Building Task, two players collaborate: an Architect (A) provides instructions to a Builder (B) to assemble a specified structure using 3D blocks. In this work, we investigate the use of large language models (LLMs) to predict the sequence of actions taken by the Builder. Leveraging LLMs' in-context learning abilities, we use few-shot prompting techniques, that significantly improve performance over baseline methods. Additionally, we present a detailed analysis of the gaps in performance for future work
[ -0.0065323124, 0.0412634984, -0.0189778935, 0.0150125176, -0.0291380137, -0.0141973309, -0.009763943, -0.008005403, 0.0298926663, -0.0588526763, -0.0289496593, -0.0230265055, -0.0277933478, -0.0063875858, 0.007668444, 0.011820971, 0.0164169725, -0.0424654447, -0.0086696586, 0.0103846537, -0.0286227763, -0.00539832, 0.0114338249, 0.0929207206, 0.0188883841, -0.0370766446, -0.0586677715, 0.0009104034, 0.004809272, -0.034288682, 0.0051762285, 0.0507070683, 0.0200929102, 0.0229237191, 0.0000020262, -0.0475606136, -0.005978757, -0.0306757744, 0.014355151, -0.0321849734, -0.0534839146, 0.0252651479, -0.015550497, 0.0339054838, -0.0564341024, -0.021693185, 0.0572961122, 0.0114295958, 0.0458513647, 0.0477173105, 0.000905346, -0.0287483633, -0.0243559107, -0.0074354243, 0.0564781427, 0.0038140386, 0.0243797936, 0.0592641421, 0.0674018189, 0.0694590434, -0.0060853041, 0.0241257511, 0.0091789784, -0.0052570505, -0.0009356284, 0.0572610982, -0.0034415494, 0.0295756906, -0.0062551089, 0.0640099049, 0.0195042137, -0.0035308325, 0.0561125316, -0.0046398388, 0.0049376371, 0.0054254895, -0.0591260418, 0.0446871072, -0.0052821413, -0.0010312657, -0.020095285, -0.0259426739, 0.0186579209, 0.0141884806, 0.0242857132, 0.0631908029, -0.0114792632, -0.0090136407, -0.0968593061, 0.0022785184, 0.027999416, -0.0231813472, -0.0337941796, 0.0491547659, -0.0613343529, -0.0000312666, 0.0016396893, 0.0246039871, -0.0177705884, 0.0034387512, 0.0765115023, 0.0457307734, 0.0053135836, 0.0604605041, -0.0338548757, 0.0624464191, 0.0091288295, -0.0063611683, -0.0508761294, 0.0528136417, -0.0868843645, -0.0024563544, -0.0497885682, 0.027210528, -0.0634969398, -0.019306412, 0.0018842077, 0.0413933508, -0.0246213302, 0.029189609, 0.0054487507, 0.0268092789, -0.0155746937, -0.0092622573, 0.0123242876, -0.0337747894, -0.0005625645, 0.0091178874, -0.0092233773, -0.0539158545, 0.0002380795, 0.0114423865, 0.0029940254, -0.0099594332, -0.0002092357, 0.0511680171, -0.0209567193, -0.0212619305, -0.0090664737, -0.0598453693, -0.0056946012, -0.0462406538, 0.0022212518, -0.0951136127, -0.0291018337, -0.0054338397, -0.0230374299, -0.0218810346, 0.0202772468, -0.0064732782, -0.0082402276, 0.0591479205, 0.0251159389, 0.0344919562, 0.0255464278, 0.0318893902, 0.0196530744, 0.0484890454, 0.0278146043, 0.0729643777, -0.0082319826, 0.025653502, 0.0368348025, 0.0442585684, -0.0433536768, 0.0057531493, 0.0191571489, -0.001446259, 0.0262419619, 0.0542918183, -0.0062373392, 0.0137272859, -0.0218991246, 0.0444838665, 0.05099722, 0.0592886209, 0.0211781934, 0.0198609438, 0.0723889619, 0.0281970613, -0.0507437289, -0.045632571, -0.0218516123, 0.0500913523, 0.0087729981, -0.0079991193, 0.0132188406, 0.0261186212, -0.0183523092, -0.045195207, 0.0065306611, -0.0070935879, 0.0040651858, -0.0080351038, 0.0254228227, -0.0127500435, -0.0704619288, 0.0025112883, -0.0301897805, -0.0611800589, 0.0103382273, -0.0624652281, 0.0527514815, 0.0899164453, 0.025977822, -0.0360056534, -0.0007926458, -0.008914032, 0.013148699, -0.0433966145, 0.0011627057, -0.032423079, -0.0018664572, -0.0213521179, -0.0517494343, -0.0321658067, 0.0132514937, 0.0331076682, 0.0169280805, -0.0166157242, 0.0290742666, 0.0206130259, -0.0088456264, 0.0185966641, -0.0236208253, -0.0752846822, 0.0025856718, 0.0134478146, 0.0145015353, -0.0051031336, 0.0007986416, 0.0077438713, 0.0492001921, -0.0392743461, -0.0320069194, 0.0235341322, -0.0507115088, -0.0623326451, 0.0177850276, 0.0156819355, 0.0228068922, 0.0012400018, -0.0280364063, -0.0207381491, 0.0484568737, -0.0164435953, 0.039334543, -0.012021671, 0.0116528822, -0.0054144999, -0.0567629747, -0.0594867133, 0.0542146936, -0.0081624156, 0.0126966918, 0.0335414708, 0.0247179884, 0.0041576987, -0.0204685144, 0.0048453808, 0.0240988117, -0.0155326938, 0.0473974198, 0.0377190821, 0.0395162106, -0.0438968241, -0.0262081381, -0.1200146154, -0.0217514709, -0.0449837074, 0.0080159465, 0.0075936331, 0.0145606045, -0.013966308, -0.0235086232, -0.0133248251, -0.0148430457, 0.0168266781, 0.0137811303, 0.0003833299, 0.0260173958, -0.0335935652, -0.0466943197, -0.0036655888, -0.0068852101, 0.0538616218, 0.0069857142, -0.010593473, -0.0607321598, 0.0796484128, 0.0218338612, 0.0146483285, -0.024088148, -0.0340162925, 0.0617957376, -0.083408691, 0.0745903701, 0.0589279197, -0.0374117605, 0.032238882, 0.0749292076, -0.0221803598, -0.0892455801, 0.0100039691, -0.0144992312, -0.0516966879, -0.0148376236, -0.003873809, 0.0428590402, 0.0663810968, -0.0030133871, 0.0105025098, -0.0458834916, 0.0056632501, 0.0121081406, -0.0350329466, -0.0010262199, -0.034430746, -0.021784788, 0.0272206701, -0.0092319781, -0.065560244, 0.0340421163, -0.0336648338, -0.0315776207, 0.00465792, 0.0066081183, -0.0147470729, 0.1008737385, -0.0014279082, 0.0083746566, -0.0030633383, -0.0445231684, -0.0431338064, -0.0294073932, 0.033407826, -0.0274289176, -0.0186841134, -0.0776280463, 0.0079333866, 0.0148907928, -0.0619899966, -0.0396454521, -0.0056221602, -0.0239542238, -0.0014302551, -0.0061161453, -0.002257932, -0.0092070345, 0.0133408522, 0.040029712, 0.0174007416, 0.0281943157, -0.0397246107, -0.0150141437, 0.0253888424, 0.0388152599, 0.0326993987, -0.0735789686, -0.0089011453, 0.0043400391, 0.0362129025, 0.0546741337, 0.0035231088, -0.0829394832, 0.0116357366, 0.0627542436, 0.0174373575, -0.0208314192, 0.0765715614, 0.0643830895, -0.0438204966, -0.0451540053, 0.0830882192, 0.0362503082, -0.0320063941, -0.0050052381, -0.0280660465, 0.0162345413, 0.0091892509, 0.0117104677, -0.0268837735, -0.0331538208, 0.0043534758, -0.0032367976, -0.0023943374, -0.0334838517, 0.0207143556, -0.13525033, 0.0418504365, 0.023120556, -0.0044967593, -0.0501865782, -0.0151211051, 0.0229031295, 0.0190658588, -0.0739410669, 0.0218408778, 0.0017107226, 0.029928444, 0.0264796317, -0.0396532603, -0.0262229294, 0.0433892831, -0.0469157137, 0.079757832, 0.1036083028, 0.0170394424, 0.0387230627, 0.0120858289, -0.0126682585, 0.0386241265, 0.044979766, -0.022486683, -0.0130020464, 0.0062584518, 0.0025646896, 0.0137897348, 0.0369165726, -0.0414595492, -0.0608124249, 0.0191361737, 0.0701979399, -0.0434594713, 0.011681444, -0.0156035246, 0.0083088959, 0.0734417364, 0.0260328613, -0.0043817037, -0.0224528685, -0.0306995977, -0.0194107108, -0.0046710996, -0.0510101579, -0.0385231115, -0.0382126421, 0.0145881223, 0.0785655901, -0.0863400921, -0.0260254703, 0.0361138396, 0.0397579446, -0.005345182, 0.004471363, -0.0011259385, 0.0382765718, -0.0656006038, 0.008127789, -0.0216614194, 0.0254095439, -0.0304608382, 0.0049673701, 0.0328330435, -0.0641575605, 0.0155613404, 0.0131526142, -0.0611438975, -0.0584670231, -0.0186629854, 0.0689921007, 0.0134861348, 0.0319997892, 0.0242921188, -0.0291486047, 0.004318824, 0.0270648282, -0.0308659468, 0.0191864148, 0.0240186118, 0.0279478319, 0.0037982948, -0.0028988845, -0.0501680821, 0.0026174036, 0.008115544, 0.0626023337, -0.0548540428, -0.0433434546, 0.0173812583, -0.0215541013, 0.0433154888, -0.0203849059, -0.0035879209, -0.0101939449, -0.0059218383, 0.0019278444, -0.0027025365, -0.045347698, 0.0432611555, 0.0149678411, -0.000387802, 0.0320153236, 0.0316075683, -0.0072701485, 0.0359599404, -0.0035763511, 0.0583415069, -0.0057775765, 0.0084605897, 0.074498713, 0.0170508139, -0.0255689174, 0.0289520193, 0.0128244935, 0.0367026441, 0.0037822418, 0.0338945277, 0.0257895626, 0.0104663614, -0.0338495597, 0.0363224894, -0.0286598336, -0.0668515563, -0.0017955529, 0.0135970945, 0.0751020759, 0.0033868039, 0.0071543395, -0.0397812612, 0.0357201844, -0.0420653857, 0.0408990681, 0.0271940511, -0.0734468028, 0.0001934, 0.0546096414, -0.009834446, -0.0058944696, 0.0641075745, 0.003320181, 0.0163113587, 0.0007921645, -0.0281614047, -0.0215616543, -0.0275987033, 0.0429278463, 0.0261530373, -0.0393853411, 0.0031589654, -0.0339466855, -0.0127188573, -0.0306132101, 0.043369364, -0.0056935535, -0.0234193895, 0.0300661791, 0.0835912004, 0.019410057, -0.0195643473, -0.0565575063, -0.0098862257, -0.0260381009, -0.0177323688, -0.0824011713, -5.159370337e-33, -0.0200102348, -0.0851069093, 0.0740981847, -0.0073472988, -0.0023903896, -0.0179068465, -0.0656480789, 0.0209367778, 0.0091769211, -0.042045936, 0.0011317227, 0.0006626719, 0.0034134747, 0.0126315532, 0.0624895915, 0.0157290027, 0.0402171724, 0.0122068627, -0.0427347571, 0.0249633286, 0.0021670698, 0.0141429892, -0.0038989868, -0.0389514752, -0.0088742031, -0.0160968676, -0.0044714329, -0.0054581463, 0.0092055453, 0.0175540317, -0.0110398727, -0.0016998459, 0.0061449111, -0.0091351159, 0.0031665484, 0.0346695371, -0.0410506688, -0.0049167094, 0.007993903, -0.0212593079, -0.0345842689, -0.0467508323, -0.0211598836, 0.000504796, -0.0538469367, -0.0069185272, 0.044678513, 0.0090663964, -0.0555346347, 0.0025529407, -0.0801878497, -0.0048682611, -0.0370374024, -0.0476446413, 0.0158358235, 0.0151863499, 0.0255045351, -0.0169819221, -0.0413210988, 0.0551414564, -0.0397536382, 0.0601009801, 0.0406038612, 0.0638067126, 0.0213734973, 0.056259945, 0.0202160273, 0.0234501772, -0.0419938937, 0.0007534608, 0.0236404315, 0.0716985539, -0.0001801921, 0.0467483625, 0.0058191339, 0.0573771484, -0.0351427756, 0.0248747561, -0.0207816157, 0.0682516992, 0.0462126918, 0.0427297503, -0.0201558527, -0.0151784448, -0.0258024335, 0.0232944228, 0.0016290458, -0.0272719022, 0.0422589295, 0.0328765623, 0.0035187798, -0.052360002, 0.0607882068, -0.0174414944, 0.0013607432, 0.0260938983, -0.0519556738, 0.0117822895, -0.0413004681, 0.0337530635, -0.0273267925, -0.0533046871, 0.0295955725, 0.0094626732, 0.0151012503, 0.0214807838, -0.0388532691, 0.0698157027, -0.0435848348, -0.0198345464, -0.0536708124, -0.0245409086, -0.0227178261, -0.0113510517, 0.0192462988, -0.0336991102, -0.0123622837, 0.0097240414, -0.0203443598, -0.0530564673, -0.0024447714, -0.0473960079, -0.0222303607, -0.0222341251, -0.0205438938, 0.0025135952, -0.0783699527, 0.0507829972, 0.0624154732, -0.0158711839, -0.025316529, -0.0022556945, 2.836e-7, -0.0212068409, 0.0382679775, -0.0243439116, -0.0012382384, -0.0285750888, 0.0533074513, -0.0263961889, 0.0109024076, 0.0267519858, 0.039944645, 0.0556124933, 0.0279353969, 0.0409901962, -0.0241773669, -0.1008368284, 0.0249184426, -0.0355007835, -0.0083795572, -0.0646656379, -0.0123884706, 0.0456774607, 0.016761763, 0.0002658553, 0.0172356945, 0.0279195402, -0.0947318822, 0.0365550369, -0.0373192802, 0.0076935352, 0.0111215394, -0.0226196814, -0.0132166101, -0.0161734745, -0.0627652481, -0.0125073735, -0.0366860852, -0.0064108046, 0.0496139899, -0.0004109497, 0.020403279, 0.0075498498, -0.0201979838, 0.003025529, -0.0012148392, 0.0165853631, -0.0735088885, 0.0223457702, 0.0244096983, 0.0068372963, -0.0117294742, -0.0394238718, -0.0013247867, 0.0022845557, -0.020903917, -0.010861095, -0.0238377545, 0.0059056506, -0.0802552998, -0.0165662058, -0.0042359806, -0.0541198142, -0.0489406958, -0.0341213755, -0.0081330528, -0.0103431623, -0.0268536862, 0.0014499555, 2.48530638e-34, -0.0259751193, 0.0064898506, 0.0203277897, 0.0549139343, 0.0346574225, 0.0068736775, -0.0143452343, 0.0516420826, 0.0481809974, -0.0612215251, -0.0152088748 ]
http://arxiv.org/abs/2406.17542v2
http://arxiv.org/pdf/2406.17542v2
CDQuant: Accurate Post-training Weight Quantization of Large Pre-trained Models using Greedy Coordinate Descent
2024-06-25T13:29:14
2024-06-26T07:44:42
Pranav Ajit Nair, Arun Sai Suggala
cs.LG
cs.LG, cs.AI, cs.CL
Large language models (LLMs) have recently demonstrated remarkable performance across diverse language tasks. But their deployment is often constrained by their substantial computational and storage requirements. Quantization has emerged as a key technique for addressing this challenge, enabling the compression of large models with minimal impact on performance. The recent GPTQ algorithm, a post-training quantization (PTQ) method, has proven highly effective for compressing LLMs, sparking a wave of research that leverages GPTQ as a core component. Recognizing the pivotal role of GPTQ in the PTQ landscape, we introduce CDQuant, a simple and scalable alternative to GPTQ with improved performance. CDQuant uses coordinate descent to minimize the layer-wise reconstruction loss to achieve high-quality quantized weights. Our algorithm is easy to implement and scales efficiently to models with hundreds of billions of parameters. Through extensive evaluation on the PaLM2 model family, we demonstrate that CDQuant consistently outperforms GPTQ across diverse model sizes and quantization levels. In particular, for INT2 quantization of PaLM2-Otter, CDQuant achieves a 10% reduction in perplexity compared to GPTQ.
CDQuant: Accurate Post-training Weight Quantization of Large Pre-trained Models using Greedy Coordinate Descent: Large language models (LLMs) have recently demonstrated remarkable performance across diverse language tasks. But their deployment is often constrained by their substantial computational and storage requirements. Quantization has emerged as a key technique for addressing this challenge, enabling the compression of large models with minimal impact on performance. The recent GPTQ algorithm, a post-training quantization (PTQ) method, has proven highly effective for compressing LLMs, sparking a wave of research that leverages GPTQ as a core component. Recognizing the pivotal role of GPTQ in the PTQ landscape, we introduce CDQuant, a simple and scalable alternative to GPTQ with improved performance. CDQuant uses coordinate descent to minimize the layer-wise reconstruction loss to achieve high-quality quantized weights. Our algorithm is easy to implement and scales efficiently to models with hundreds of billions of parameters. Through extensive evaluation on the PaLM2 model family, we demonstrate that CDQuant consistently outperforms GPTQ across diverse model sizes and quantization levels. In particular, for INT2 quantization of PaLM2-Otter, CDQuant achieves a 10% reduction in perplexity compared to GPTQ.
[ -0.0388786346, 0.1333447397, 0.0106212338, 0.0269220881, 0.0173252299, 0.0207526293, -0.0043918444, 0.0395180471, -0.0370271355, -0.0139473919, -0.0425520986, -0.037397068, -0.0010746699, 0.0076222187, 0.0017767652, -0.0220315084, 0.037456397, -0.0056540603, -0.03322291, -0.0251050405, -0.0049516605, 0.0341088772, 0.0248563029, 0.0282600243, -0.0055455067, -0.0229030587, -0.0236134343, 0.0144783808, -0.0167919789, -0.0227808356, 0.0028354381, 0.0810046718, 0.039821703, 0.0508287996, 0.0000023644, -0.0097672269, -0.0147610623, 0.0152448528, -0.0457427949, 0.0963558778, 0.0272392891, -0.0109787518, -0.0199905448, 0.001098371, -0.0237620063, -0.0489858948, 0.0641434863, 0.0449462198, 0.0432429165, 0.0424425229, -0.0033746297, -0.0401005521, 0.002730309, -0.0153808845, -0.0211521722, -0.0421193317, 0.0157776661, 0.019793218, -0.0070869722, 0.0161750726, -0.0052621569, 0.0291949362, 0.0261658374, 0.0328796022, -0.031852711, 0.025284104, -0.0449304879, -0.015280474, 0.0061372984, 0.0530190691, -0.0494467579, -0.0403501801, -0.0236252155, 0.0270190723, -0.0217306968, -0.0731722862, -0.0325048566, 0.0000262064, 0.0279840324, 0.0440110788, -0.0070693851, -0.0332792699, 0.0628592446, -0.016947018, -0.0626314133, 0.0195345059, 0.0526406616, -0.0087191202, -0.0268390384, -0.0141942715, -0.0748976395, -0.0261286013, 0.0142673785, -0.0282080062, 0.003770442, -0.0283995848, -0.0362455286, -0.0310592204, -0.009747264, 0.0060510752, 0.058597669, 0.0435203575, -0.0102337385, 0.0245017204, -0.0088307578, 0.0370166674, -0.025146421, 0.0100122942, -0.0258451737, 0.0066032205, -0.0341980308, -0.0038175883, -0.0620621368, 0.0344920233, -0.0273108762, -0.0159891304, -0.001370925, -0.0244154725, 0.0042781373, -0.063523367, 0.054869052, 0.0344227515, -0.0254782327, -0.0074702976, 0.0347938612, 0.0391864032, 0.0202398729, 0.0265696999, -0.0194747485, -0.0312237311, -0.0219356455, 0.017024301, 0.0213177297, -0.0330538973, 0.0289211441, -0.0003211404, -0.003329122, -0.0476263538, 0.036417108, 0.0235743243, -0.0118306903, -0.0471948646, -0.0458240211, -0.0622348338, 0.0097731687, -0.0449053384, 0.0085328864, -0.0105164293, 0.0174873993, 0.0004684036, -0.0157659911, 0.0548986346, -0.0284802765, 0.007293459, 0.0369425006, 0.0195777565, 0.019808732, 0.0606282838, 0.0087788729, 0.0150069101, -0.0002320934, -0.0258411691, -0.0163879376, 0.0630563423, 0.0310304873, -0.0040157307, -0.0254816841, -0.0209376346, 0.0039129192, 0.0352402665, 0.0753245056, 0.0699875578, -0.087007381, -0.0207836982, 0.0332580879, 0.0778674632, 0.0362071507, 0.0327482112, 0.0232154001, 0.0424291268, 0.0054113409, -0.0320470333, -0.0406978652, 0.000543704, 0.0015525491, -0.012229383, 0.0162075832, 0.049150221, -0.0375558026, -0.0135060716, 0.0036079891, -0.0216338206, 0.0224276073, 0.014275074, 0.0326459855, 0.0206123032, -0.023775924, 0.0918908566, 0.0202865563, -0.064295657, -0.0076302299, -0.0583437197, -0.0174305178, 0.006774053, 0.0232496802, -0.003752264, 0.0000817905, 0.0332895331, -0.0052118041, -0.0144566391, -0.0164902769, 0.0117262416, 0.0584593676, -0.0498692133, -0.0174326673, 0.0411600508, 0.0336144082, -0.0063450965, 0.0533284023, 0.0193093996, -0.0019176519, -0.0271166898, -0.0168404095, 0.0275672004, -0.012962956, -0.0578399822, 0.0341891199, -0.0035675776, 0.0219096784, 0.0037995684, -0.0181883052, -0.0036771926, 0.0652556345, -0.0444114469, 0.0264012478, -0.0428659283, -0.0210537072, -0.0630627275, 0.0695579797, 0.0102330483, 0.0081349602, 0.1005059183, -0.0211418457, -0.047111962, 0.019315334, -0.0000471758, 0.0495120399, -0.0786499158, 0.0068768193, -0.0289652701, 0.0233981945, -0.0191357955, -0.0008376483, 0.0018238587, -0.0092372317, 0.0232555401, -0.0047924668, -0.0149388667, -0.0563057326, -0.0013611824, 0.0090249144, -0.0028909398, 0.0487566181, 0.0039932132, 0.0292289406, 0.022897888, 0.0088466415, -0.1436016113, 0.0061353673, -0.0403260514, 0.0043023801, 0.0337039083, 0.0370898955, 0.0022394464, -0.007007062, -0.0219632126, -0.0318971686, -0.0058727115, -0.0072414316, -0.0208993144, -0.0230132714, -0.0154732438, -0.0374423824, -0.0505371429, 0.030015802, 0.0492137857, -0.0673102215, -0.0322814658, -0.0462271124, 0.0222959667, -0.0222490653, 0.0351813063, 0.0164413303, -0.0565037467, 0.0067013525, -0.0294839647, -0.01553463, 0.0468318462, 0.0138335442, 0.0176386368, 0.0571947619, 0.0105234226, -0.0341811366, 0.0059523433, -0.0401244424, -0.0485069305, -0.0238071214, -0.0311514921, -0.030242959, 0.0904968753, -0.0229435358, -0.0179847423, -0.0537554771, -0.0339735039, -0.0161448866, -0.0102404281, -0.0512976609, 0.0103022205, -0.0328296125, 0.0508151352, -0.0040847068, -0.0421521701, 0.0397279933, -0.0455635861, -0.0066093882, -0.0370005295, -0.0081062932, -0.0274997782, 0.0651559457, 0.0088274041, 0.0474636517, -0.0229048561, -0.0047206907, -0.0433618575, 0.0052352967, -0.0222318191, -0.0249758959, -0.0363688171, -0.0001944591, 0.0153301517, -0.0163929351, -0.0415021367, 0.0311670769, -0.0113496501, -0.0328637362, 0.0357869379, 0.0147797307, -0.0477732942, 0.0278676804, 0.0329199657, -0.0083166109, 0.042791184, -0.0136874449, 0.0301168114, -0.0047205617, 0.035975419, 0.0245511141, 0.0053322208, 0.0450261757, -0.0068989936, 0.0264188405, 0.0321470015, 0.0204425398, 0.0289610829, -0.0302440003, 0.0157449171, 0.0463439412, 0.0242411308, -0.0332473852, -0.0041191913, 0.059668228, -0.0421168618, -0.0020919747, 0.0926625133, 0.0140274493, -0.0769259855, 0.0035387296, 0.0170255657, 0.0343697518, -0.0031176677, 0.0025065818, -0.089540042, -0.043965809, 0.0139930909, -0.0059997546, 0.0623552166, 0.0158039518, 0.0103861848, -0.0626487583, 0.0389699191, 0.0557640418, -0.000213179, 0.0099701909, -0.0181091689, -0.0188347623, -0.0247943178, 0.0399520881, 0.0424263328, -0.0041392366, 0.0222389977, 0.0218272414, 0.0071588722, -0.0255187973, 0.0423149355, -0.021154616, 0.0069596395, 0.0633858964, 0.0236172266, 0.0077716005, -0.0156426262, 0.0381016172, 0.007669914, -0.0390768461, -0.0002874921, -0.0167712532, 0.0174277108, -0.0125958128, 0.0092493333, 0.0395419113, -0.0198868793, 0.0036109532, -0.0067560691, 0.081004858, 0.0095595662, -0.0253429059, 0.0126755787, -0.0366704501, -0.0597907566, 0.0638159961, 0.0508796759, -0.018497128, -0.0401284359, -0.0019419045, -0.0126371393, -0.0247109775, -0.0387568362, 0.0233334042, 0.0066399393, -0.0765173808, -0.0439545102, 0.0029315895, 0.0210890584, 0.0140982186, -0.0412344746, 0.0032851205, -0.010120051, -0.0002386693, -0.1012367308, 0.0376276039, -0.0231146906, 0.0131800221, -0.0699158609, 0.0260326769, -0.0054036244, -0.0228287596, -0.0139021855, -0.0280288961, -0.0966039076, -0.0511505455, 0.0058213985, 0.0308827609, 0.0112869758, 0.0355152674, -0.010087247, -0.0496291965, 0.0220784359, 0.0335984342, -0.0139781041, 0.0580517054, 0.0459832959, -0.0744468421, 0.0013396183, 0.056540288, -0.0170315579, 0.0419634394, -0.0389824472, 0.0285001751, -0.042816475, -0.0767684951, -0.0026181769, 0.0056644264, 0.0855455622, -0.0125938114, 0.0160178915, 0.0053237602, 0.0226630401, -0.0082190903, 0.0173917785, -0.0584731773, 0.0114323702, 0.0395246856, 0.0214676708, -0.0207762308, -0.0087734871, 0.0387175865, 0.0029247121, 0.0021340216, -0.0006133977, -0.0501868911, 0.0063973693, 0.0621168911, -0.0081935041, -0.0189960841, 0.0184236653, -0.0382594876, 0.0085847573, -0.0428446867, -0.0108969752, -0.0183424279, 0.0305031631, -0.0430596545, 0.0477146432, -0.0210993625, -0.1028636247, -0.0527545959, -0.0331329182, 0.0635225624, 0.0152427936, 0.0540274791, 0.0363177657, -0.049667988, -0.038066972, 0.0217292067, -0.0164842922, -0.0051740846, 0.0244780313, 0.0119373901, -0.0261835866, 0.0070413509, 0.0686196238, 0.0424586274, 0.0201870222, 0.0110433064, 0.0012191649, 0.007745815, -0.0108131133, -0.0108621037, 0.0358378775, -0.0314900577, -0.0350493826, 0.0092807664, -0.0723171458, 0.026694892, 0.0020873155, -0.0483178832, -0.0176107716, 0.044781357, 0.0019437443, -0.0063179694, -0.0071161594, -0.0149252694, -0.0331539325, 0.0044242674, -0.0393208675, 0.0171920918, -7.026570379e-33, 0.0016613071, 0.0417903736, -0.0001897365, -0.0209925734, -0.0389096849, 0.0306670573, -0.0136433551, -0.0114023518, -0.0022809056, 0.005073864, -0.0114103584, 0.0301664583, 0.0035419585, 0.0215298869, -0.0134807844, 0.0301581249, 0.0810539201, -0.0260810331, -0.0026214805, 0.0482275523, -0.0035704237, 0.0433485061, 0.060901951, 0.0142579805, -0.0639158785, 0.0205831807, 0.0043437593, 0.0151331155, 0.0045630173, 0.0420952663, -0.0616284348, -0.0143591706, -0.0206463505, -0.0004849815, -0.0292167496, -0.0168499947, -0.0685700849, 0.0129896943, -0.0016963714, 0.0020089771, -0.0028808443, -0.019706808, 0.0102052968, 0.0032668959, -0.0444874354, -0.0398047678, -0.0244500339, -0.0136773484, -0.0475793891, -0.0810580105, -0.0061510676, 0.0349608921, 0.0130104776, -0.0112615852, 0.0772849321, -0.0123539455, -0.0582133196, 0.0215054546, -0.0476157032, -0.0276034717, 0.0040004924, 0.0753552839, 0.034909226, 0.0316851623, 0.0389487706, 0.0394638926, 0.0437360704, -0.0511696339, -0.0123558436, 0.0383834653, 0.0604590662, 0.0311655719, 0.0566520281, 0.0313667767, 0.0982375517, -0.008857321, -0.076200828, 0.001275045, -0.0502209552, -0.0081089661, 0.0090941191, 0.0500712059, 0.0151679451, -0.0267356467, -0.0516468138, -0.0206870232, -0.0341937467, -0.0252693947, -0.0019780137, 0.0228900127, -0.0555500947, 0.0140397009, 0.0459447429, -0.0450551547, -0.0089964904, -0.0070127752, -0.0334215499, 0.0072757532, 0.0183501635, 0.0380752385, -0.0271665007, -0.0096192323, 0.0197502971, 0.0029694559, 0.0147119416, -0.0205301307, 0.0194853228, 0.0405147858, -0.0534814149, -0.0237424076, -0.0287082307, -0.0246754829, 0.0134347174, 0.0822302029, 0.0269777812, 0.0130447662, 0.0356178693, 0.0532883145, 0.0056792214, 0.0115043176, -0.0570073351, 0.0143694719, 0.0368505009, 0.0528633669, -0.0085915169, 0.0091091944, -0.071874775, 0.0397602096, 0.037193086, -0.0720091835, -0.0504510775, 0.0400106385, 2.922e-7, 0.0474847034, 0.0169449877, 0.0130765904, -0.0555499569, 0.002213808, 0.0116618518, -0.0035686123, 0.0178662501, -0.0113139059, -0.0081968307, 0.0304943733, 0.0567066111, 0.0436773077, 0.0173676126, -0.0801609382, 0.0088264728, -0.0641720369, -0.005833115, -0.0526561774, -0.0471505634, -0.0034845499, 0.1345913112, -0.0097855767, -0.0244481973, 0.0497996174, 0.0151050352, 0.049198918, -0.011244663, 0.0167298596, -0.0024673259, 0.0027086949, 0.0047210534, 0.002632285, -0.0786553025, 0.0020666528, 0.0115495864, -0.0118954303, 0.0835891291, -0.04227506, 0.0009502477, 0.0163313672, 0.0110316863, -0.0298153851, -0.0501452163, 0.0543610826, -0.0234336797, 0.017293321, -0.1060630381, -0.0113858953, -0.03471037, 0.0223437492, -0.0306702666, 0.0336489007, 0.0484728515, 0.0238581933, -0.0161002707, -0.0114379246, -0.0840304494, -0.0296393335, 0.0245586503, -0.0262591764, -0.0152734667, -0.0577025414, -0.0060048383, 0.0275497027, -0.0217178613, 0.0245037861, 2.999862502e-34, 0.0322520509, -0.0107609555, 0.028450748, 0.0487391204, 0.0023376886, -0.0510860905, 0.0227555726, 0.0191926379, -0.0361656882, -0.0657123998, -0.0200853441 ]
http://arxiv.org/abs/2406.17535v1
http://arxiv.org/pdf/2406.17535v1
Disce aut Deficere: Evaluating LLMs Proficiency on the INVALSI Italian Benchmark
2024-06-25T13:20:08
2024-06-25T13:20:08
Fabio Mercorio, Mario Mezzanzanica, Daniele Potertì, Antonio Serino, Andrea Seveso
cs.CL
cs.CL, cs.AI
Recent advancements in Large Language Models (LLMs) have significantly enhanced their ability to generate and manipulate human language, highlighting their potential across various applications. Evaluating LLMs in languages other than English is crucial for ensuring their linguistic versatility, cultural relevance, and applicability in diverse global contexts, thus broadening their usability and effectiveness. We tackle this challenge by introducing a structured benchmark using the INVALSI tests, a set of well-established assessments designed to measure educational competencies across Italy. Our study makes three primary contributions: Firstly, we adapt the INVALSI benchmark for automated LLM evaluation, which involves rigorous adaptation of the test format to suit automated processing while retaining the essence of the original tests. Secondly, we provide a detailed assessment of current LLMs, offering a crucial reference point for the academic community. Finally, we visually compare the performance of these models against human results. Additionally, researchers are invited to submit their models for ongoing evaluation, ensuring the benchmark remains a current and valuable resource.
Disce aut Deficere: Evaluating LLMs Proficiency on the INVALSI Italian Benchmark: Recent advancements in Large Language Models (LLMs) have significantly enhanced their ability to generate and manipulate human language, highlighting their potential across various applications. Evaluating LLMs in languages other than English is crucial for ensuring their linguistic versatility, cultural relevance, and applicability in diverse global contexts, thus broadening their usability and effectiveness. We tackle this challenge by introducing a structured benchmark using the INVALSI tests, a set of well-established assessments designed to measure educational competencies across Italy. Our study makes three primary contributions: Firstly, we adapt the INVALSI benchmark for automated LLM evaluation, which involves rigorous adaptation of the test format to suit automated processing while retaining the essence of the original tests. Secondly, we provide a detailed assessment of current LLMs, offering a crucial reference point for the academic community. Finally, we visually compare the performance of these models against human results. Additionally, researchers are invited to submit their models for ongoing evaluation, ensuring the benchmark remains a current and valuable resource.
[ -0.0231401492, 0.0242754333, -0.0489689559, 0.0369844697, -0.0370732099, 0.0074159214, 0.0325496793, 0.0218755547, 0.0260929503, -0.0096810767, -0.0339929387, -0.033729922, 0.0047557582, -0.0035595966, 0.0403592177, -0.0423124656, 0.0448765121, -0.0352432504, -0.0343345515, -0.0049927556, -0.0012660667, -0.0075117922, -0.0222726688, 0.0333262384, 0.0052868887, 0.0171387419, 0.0298735853, -0.0155240828, 0.0101466412, -0.0195532069, 0.0160218365, 0.0378563032, 0.0007935312, 0.0063258996, 0.0000021896, -0.0731602088, 0.0125661911, -0.0355190895, 0.021172896, 0.0462145992, 0.045438908, 0.0409393311, -0.0458347984, 0.0100775743, -0.0314325579, 0.044784449, 0.0191752426, -0.061909955, 0.015887009, 0.052670531, -0.0061017969, 0.0029208742, 0.0349282511, -0.0182070863, -0.0383426137, 0.0004567564, 0.004412205, 0.0193034504, 0.0207885485, -0.0056175729, -0.0027199234, 0.024990093, -0.0096990652, -0.0023622136, -0.0420036688, -0.0058534257, 0.0046129511, -0.023375392, 0.0044893599, 0.0272270311, 0.0197404139, -0.0184402876, 0.0040965187, 0.0025089909, 0.0101004262, 0.031473998, -0.0629304126, 0.0140569825, 0.0150660491, 0.0094182445, 0.050284896, 0.0160396639, 0.0017707775, 0.0059882682, -0.0006792123, 0.0781994313, 0.0668595284, -0.0517027341, 0.0435927361, 0.0060658087, -0.0102884816, -0.0452714898, -0.0053880014, 0.0554597043, -0.0288289879, 0.0063532135, -0.0177531708, -0.0181105379, 0.0728690699, -0.0509782732, 0.0617252775, 0.0318852589, -0.0152384182, 0.0262389611, -0.004951091, 0.1052015871, -0.0002942495, 0.0096149007, -0.0442294516, 0.0114699397, -0.0356636718, 0.0110557331, -0.0152837522, -0.0185331646, -0.0431114584, 0.0300843287, -0.0821313187, -0.0039394186, -0.0452509001, 0.0179353505, -0.0050555891, -0.0726507604, -0.0034596741, 0.0371504277, -0.0328783989, 0.0421082601, -0.0476083308, -0.0024539956, -0.0179098733, -0.063267, 0.042874746, 0.0255008023, -0.0315301567, -0.0210023057, 0.0279706381, 0.0448912457, -0.0018813085, -0.0137975439, -0.0192885548, -0.0214939024, -0.0342815295, -0.0861734375, 0.0206499137, -0.0767357424, 0.0275020748, -0.0078894803, 0.0066992962, -0.0531922989, -0.0041630589, 0.0343188085, 0.0354385115, -0.0141935926, -0.0281259418, 0.0041853539, 0.0264614578, 0.0122448066, -0.0015418245, 0.0510774292, 0.0578209758, -0.006699821, 0.0340448692, -0.036911577, 0.0006898086, 0.0998582318, 0.0108806863, 0.045262266, -0.0563146397, 0.0043594185, 0.0051186685, 0.0295848642, 0.0378423594, 0.0497963764, 0.0221400782, 0.0406230316, 0.0550751239, 0.0668292567, 0.0243199021, -0.0003256761, 0.0491772927, 0.010981177, 0.0472383983, -0.0094664525, 0.0068852203, -0.0151120927, -0.0092631923, 0.0156214023, -0.0303469002, 0.0572358109, -0.0096564898, -0.0340709463, -0.0104788048, -0.0246509593, 0.0247840453, 0.0039404789, 0.0624493398, -0.0263132267, -0.0241809022, 0.0242102481, 0.0445988849, -0.0705720931, 0.0417767763, -0.0446353629, 0.024835486, 0.0857625902, 0.0319540873, 0.0042321575, -0.015454377, 0.0197303463, 0.020094065, 0.0792076141, 0.0318368003, -0.0104276314, -0.0281230751, -0.0572576448, 0.0131316893, -0.0011440766, 0.0094898939, 0.0219204631, -0.0017848233, 0.0313638337, 0.0075187706, -0.0435543656, -0.0390820168, 0.0165933911, -0.0823300257, 0.0004896165, -0.0409474708, -0.0059191873, -0.0135100521, 0.0279316567, -0.0146114146, 0.0132684996, 0.0417940058, -0.0287167076, 0.0160927996, -0.0130017195, -0.0317518339, 0.0272787735, 0.0423728861, -0.0412529036, -0.0133896824, -0.0074019278, 0.0048016529, -0.0313185155, 0.0054129818, -0.024970239, 0.0108255297, -0.0163431093, 0.0414814502, -0.0534060933, 0.0213495307, -0.0412901826, 0.0461030751, -0.0182272084, 0.0629626513, 0.0172170009, 0.0209894814, 0.0254888646, -0.0461649448, -0.0614653118, 0.015958542, -0.0122499093, 0.0137512377, 0.0034129524, 0.0489354543, -0.0542941689, -0.068059735, -0.0942366794, 0.0055579888, -0.0047572609, 0.0307623241, 0.0171797667, -0.0225185975, -0.0336540118, 0.0412286632, -0.0381360836, -0.0311975181, 0.0327625908, -0.0007321917, 0.0308771469, -0.0291328859, -0.0581496023, -0.0562146381, -0.0114542563, -0.0049128686, -0.0053189923, 0.0279018469, 0.0196917206, -0.0114563005, 0.0374889672, -0.0422005542, -0.0100243734, -0.0012926679, -0.0445329547, 0.0241437145, 0.0133694382, 0.063060008, 0.0400877595, -0.0075554713, 0.010340862, 0.0496022925, -0.0478651226, -0.0540029183, -0.016336225, 0.0242964495, -0.0342350863, -0.0930967182, -0.104220286, -0.001980057, 0.0494772904, 0.0047941012, 0.0192675292, -0.0216902848, 0.0184111446, -0.0805712789, -0.026838759, -0.0168678686, 0.0171048138, -0.0366306305, 0.0704510361, -0.018422503, -0.0265981779, 0.0170457605, -0.0433462113, 0.0316986293, -0.0095229382, -0.0095067425, -0.0393472984, 0.0310188364, -0.0141643845, -0.0024820685, -0.051499106, 0.0110012116, -0.0449780039, -0.0169171877, 0.0158068985, -0.0785628334, 0.0608782172, -0.080543831, 0.0026407749, 0.0183252301, -0.017845802, -0.0041786944, 0.0457624905, 0.0073277028, 0.0366642028, 0.0177643411, -0.0219404455, -0.0035035056, -0.0201006699, 0.0112146186, 0.031420555, -0.0148991849, 0.0272342861, -0.0121082468, 0.0195219908, 0.0224999636, 0.0011679627, -0.0157373026, -0.0479677059, 0.0294724517, 0.0163212884, 0.0250098649, 0.0273862276, -0.0219710041, 0.0280396212, 0.0396447964, -0.0442583486, -0.0289549381, 0.0328225978, 0.0392691568, -0.0400759242, -0.03130785, 0.1107157022, 0.0153490901, -0.0172184482, 0.0169571601, -0.0313178301, 0.0374013931, 0.032455463, 0.0199436545, -0.0773376226, 0.0441707782, 0.0002105248, 0.0294617694, -0.0341098383, -0.0092586512, -0.0040269583, -0.094716683, 0.0621762201, 0.0026146551, -0.0086022699, 0.0053016734, -0.034335535, -0.0113331089, 0.0177236721, 0.0299590081, 0.0112374173, -0.0120538669, -0.0053453296, 0.0264734551, -0.0509115979, -0.0693471283, 0.0550882854, -0.0252617467, 0.0585060157, -0.00255644, 0.0377708413, -0.0537101738, -0.0465323515, 0.03923776, -0.0042451667, -0.036047928, -0.0120176263, 0.0059686354, 0.0421904773, -0.0039685951, -0.0501463525, -0.0246724393, 0.01101318, -0.0124825276, 0.0643395334, 0.0767262504, 0.0068559162, 0.0247931406, -0.0068160468, -0.0196662676, -0.0003816272, 0.0547851287, 0.0561178997, -0.0713691786, 0.0289581809, -0.0004833268, -0.0158870984, -0.0360900275, -0.058484152, 0.0386662818, -0.0471651405, -0.0023562438, -0.0591964498, -0.0318397135, 0.032611426, 0.0105080595, 0.0165892001, -0.0212231539, 0.0066802809, -0.0442568362, -0.0348320641, 0.0402236953, -0.0172937568, 0.0382728688, -0.1086896956, -0.0017494189, 0.0233639441, 0.0036148857, -0.0387856625, 0.0023587688, -0.0845037475, -0.0542618446, 0.0075356816, 0.0674656779, 0.0277206693, -0.0190183651, 0.0134575376, -0.0467327982, 0.0263125617, -0.0443679318, -0.0033595155, 0.0556289181, -0.0205507558, -0.0111990254, 0.0173283108, 0.0023903267, -0.0301373936, 0.0247909967, -0.0055574379, 0.0487662889, 0.0021810625, -0.0520659648, 0.0258222111, 0.0608432777, 0.00330355, -0.0610309914, -0.0134492302, -0.0238060486, -0.0096935565, -0.0372250564, -0.0019531139, -0.0182723515, -0.0121642165, 0.0655031875, 0.0157189239, 0.007844151, -0.0212048814, -0.0498277619, 0.0160473324, -0.0063615721, 0.0469638221, -0.0216610786, 0.0342417695, 0.0015771996, 0.0187849291, -0.0150217554, 0.026615737, -0.0454542674, 0.0785926729, -0.01443044, 0.0629701614, -0.0254042801, 0.0114648426, 0.0221694671, 0.030031886, 0.0303758774, -0.0322934575, -0.0130741429, 0.0333160944, 0.0674443021, 0.0250588898, 0.093245469, -0.0207273029, 0.0075843371, -0.0388639532, -0.0136489505, -0.0275641978, -0.0284483451, 0.0180533957, 0.0050128065, 0.010566053, 0.0010355251, 0.0187781285, 0.0120210592, -0.0006768675, -0.0111606158, 0.0429156162, -0.0185351465, -0.0414818116, -0.0006164422, -0.0158759207, -0.0324386097, -0.0366415195, -0.0855266824, -0.0736348704, 0.0068674688, 0.0633486584, -0.0119128935, -0.0237250421, -0.001960468, 0.0268342476, -0.0379121341, 0.0221822914, -0.0607706606, 0.001720174, 0.00922867, -0.0419147834, -0.0389846489, -6.302465124e-33, -0.0483882949, -0.0276766866, 0.0668037161, 0.0071572866, -0.0197526664, -0.0225475244, -0.0065328549, 0.0011173198, -0.0185112171, -0.0220539682, -0.0112543739, -0.0563936234, 0.0067537455, 0.00298868, 0.022060534, 0.0262833741, -0.0098576937, -0.0011392873, -0.006653979, -0.0293638557, -0.0162389614, -0.0301634762, 0.077845335, -0.0387060493, 0.0608890727, -0.0468846262, -0.0356214046, -0.0026258186, -0.0027018117, 0.0787522942, 0.0123714386, 0.0341538154, 0.0318369679, -0.0090260282, 0.0091394437, 0.0132029522, -0.0318917632, -0.0449986793, 0.0489899851, 0.0243953373, 0.023940444, -0.0330970995, 0.0400785916, -0.0053953598, -0.0849154517, 0.006319731, 0.0241851714, -0.0251488332, -0.0031205181, 0.0013225535, -0.0602029786, 0.0146835493, -0.0182477199, 0.0160915218, -0.0188316517, 0.0609380938, -0.0243266467, -0.0090000331, -0.0371474326, 0.0379009284, 0.0140072228, 0.0335428864, 0.0777595416, -0.0007145875, -0.0117226159, 0.0606554709, 0.0536114164, 0.007603786, 0.0557839498, -0.0208888352, 0.0853945464, 0.0270426571, 0.057610698, 0.0676229671, 0.0153557435, -0.0120564783, -0.0707769245, 0.0157530401, 0.0578989908, 0.0122850016, -0.0444850363, -0.0178783536, -0.0446804166, -0.0282260142, -0.0659117848, -0.0087798852, -0.0029468886, -0.0391045064, 0.0787746757, -0.0698990449, 0.0485804714, -0.0210296288, 0.0228859782, -0.0203357823, -0.0294018984, -0.0222680196, -0.0110769561, 0.0481115207, -0.0179511663, 0.0614376105, -0.0362691879, -0.041527953, -0.0248200502, 0.017943047, 0.0321864709, 0.0159448069, 0.0069011129, -0.0228935424, -0.0556139126, 0.0163807645, -0.0192712769, 0.0021885391, 0.0107759861, 0.002594857, 0.0254843011, -0.018844001, 0.006210858, 0.017224947, 0.0372562222, -0.0247625411, -0.0205339156, -0.0269136261, 0.014253878, 0.0401026197, -0.0151371062, 0.0189512447, -0.0301706437, 0.040122319, 0.0388884544, -0.0214582346, 0.0050839884, -0.0086156595, 2.873e-7, 0.0503000356, 0.0681036413, -0.0215466078, 0.0188505407, -0.0215497799, -0.0152297216, -0.0610766374, -0.0024118421, 0.0324090198, 0.0284533333, 0.0058384766, 0.0165038779, 0.0503052697, 0.0205796193, -0.046819903, 0.0246855523, -0.0415308438, 0.0083597768, -0.0752430409, 0.0265365466, 0.0180994701, 0.0054837433, 0.0303770304, -0.041604694, -0.0653585121, -0.0494708307, 0.0563594848, -0.0341734774, -0.060138315, 0.0514923632, 0.0311154332, 0.0273923483, 0.0287639033, 0.0304990951, -0.0123052774, 0.0267565809, 0.0032381306, 0.0866099373, 0.008556121, 0.0276834257, 0.0448916331, -0.0593469664, -0.0408496484, -0.0539176837, 0.0203703828, -0.0006852243, 0.0113929436, -0.0440085456, -0.0687893853, -0.0085285958, 0.0222120788, -0.0595662184, -0.028963577, 0.0190871917, 0.0361873545, -0.02230273, -0.0053288047, 0.0405570269, -0.0178679209, 0.0147264833, -0.0338290893, -0.1040173396, -0.0400948673, 0.0408352278, -0.0176560413, -0.0200436544, 0.0011044213, 2.766794114e-34, 0.0246123318, -0.0038527742, 0.0394318514, 0.0352499411, 0.0343832187, 0.0084142555, -0.040331807, 0.0406858586, -0.0405046269, -0.0295995474, -0.0103156492 ]