text
stringlengths 4
4.47k
|
|---|
**LoRA**(Hu et al., 2021) We only freeze the encoder for the encoder-decoder LM and the entire model for the decoder-only LM. We use the optimal rank \(r\) of 4 and adapt both \(W_{q}\) and \(W_{v}\) in the self-attention module, which corresponds to the best performing hyperparameter setting in the paper. |
1. **Parallel rationale generation (think, Subsection 4.2)**: In parallel across \(n\) tokens \(x_{i}\) in an input sequence \(x_{0:n}\), we generate \(r\) rationales of length \(t\): \(c_{i}=(c_{i1},\dots,c_{it})\), resulting in \(n\times r\) rationale candidates. We insert learned \(<\)\(\mid\)start\(\mid\)\(>\) and \(<\)\(\mid\)end\(\mid\)\(>\) tokens to mark each rationale's start and end. 2. **Mixing post-rationale and base predictions (talk, Subsection 4.3)**: From the hidden state output after each rationale, we train a "mixing head" - a shallow MLP producing a weight determining how much the post-rationale next-token predicted logits should be incorporated compared to the base language model predicted logits. This approach eases distribution shift early in finetuning, due to introducing rationales. 3. **Optimizing rationale generation (learn, Subsection 4.4)**: We optimize the rationale generation parameters (start/end tokens and LM weights) to increase the likelihood of rationales that make future text more probable. We use REINFORCE to provide a learning signal to rationales based on their impact on future-token prediction. To reduce variance, we apply a teacher-forcing trick to include in the loss the likelihood of predicting not only the token after the thought but also later tokens. |
Table 4 summarizes by listing the time (# of samples) when the mixture models finally learn a GCD that a purely AI-trained model cannot learn, and the delay (in millions samples) since the previous GCD was learned (see also Figure 30 to illustrate the comparison between the clean and the AI-trained model): |
How to improve upon SLM?There are many natural extensions of SLM, _e.g.,_ reweighting tokens instead of selecting may improve robustness; using a reference model as a reward model to guide pretraining with reinforcement learning; adopting multiple reference models to reduce overfitting; designing token-level curriculum learning and iterative strategies for continuous improvements, _etc._ |
language models, which have led to the emergence of a variety of popular models, including LLaVA [149], MiniGPT-4 [150], InstructBLIP [151], and PandaGPT [152]. The release of LLMA has greatly advanced the research progress of LLMs. To summarize the research work conducted on LLaMA, we present a brief evolutionary graph in Figure 5. |
1. Burpees: Burpees are a full-body workout that combines squats, a push-up, and a jump to produce a high number of calories burned in a short amount of time. |
We begin by validating that the attacks developed in the prior sections can effectively recover the full logit vector given a limited query interface. In Table 3 we report the average number of bits of agreement between the true logit vector and the recovered logit vector, as well as the (amortized) number of queries required to recover one full logit vector. |
Footnote 5: We do not append a <eos> token at the end of documents because we only use the first section, which does not signify the conclusion of the entire article. |
PersonA: Urinalysis is a test that evaluates a sample of your urine. It is used to detect and assess a wide range of disorders, such as urinary tract infection, kidney disease, and diabetes. It examines the appearance, concentration, and content of urine. |
Manually constructed X-of-thought methods expand on in-context learning by adding different types of step-by-step intermediate reasoning processes to demonstrations. They allow LLMs to mimic and generate reasoning paths. Although manual XoT methods provide greater interpretability as well as trustworthiness for human understanding and outperform on complex tasks, i.e., mathematical reasoning, commonsense reasoning, symbolic reasoning, etc., manual annotating of rationales entails significant costs and suffers from drawbacks such as difficulty in demonstration selection and task generalization. Specifically, different tasks require different ways of demonstrations. Therefore, other works attempt to construct the reasoning path automatically, as discussed in SS4.1.2. |
\(\bullet\)_Human alignment_. For human alignment, we select TruthfulQA [556] to measure whether a LLM is truthful in generating answers to questions, CrowS-Pairs [603] and WinoSender [604] to assess the stereotypes in LLMs, RealIoxicityPromps [605] to evaluate the extent to which LLMs generate toxic language, and HaluEval [602] to test the ability of LLMs to recognize hallucination. As the test set of Real-Toxicity-Prompts is too large, we randomly sample 10000 examples from it for evaluation. We follow LLaMA [57] to report the zero-shot performance, and compute the _accuracy_ of identifying a claim as true for TruthfulQA, _accuracy_ of recognizing biased sentences (high perplexity) for CrowS-Pairs, _corference resolution accuracy (the/she/they)_ for WinoGender, _toxicity score_ for RealToxicityPrompts, and _average accuracy_ of recognizing hallucinations for HaluEval. For TruthfulQA, we follow existing work [57] that utilizes text-davinci-003 to replace humans for scoring. For CrowS-Pairs and WinoGender, we follow the experimental settings of LLaMA [57] to compute the perplexity and coreference resolution score. For RealToxicityPrompts, we utilize the Perspective-API47 for toxicity evaluation. |
BaselinesWe also train several baselines. Two baselines are trained on \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) respectively while the third is trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\). We consider the model trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) to be an upper bound on performance as it represents an expensive full re-training. The baselines trained on individual datasets can be seen as compute-equivalent alternatives to continual pre-training (e.g., one could opt to train a model from random initialization on \(\mathcal{D}_{1}\) instead of continually pre-training it). |
In the following experiments (SS 4 and SS 5), we use Llama-2 (7B and 70B) [11] since it is one of the best-performing LLMs. We use Wikipedia articles classified under the "2023" Category including topics from diverse domains such as films, arts, economics, politics, events, etc.3 The likelihood that this factual information is not |
**Rationale Distillation.** For rationale generation from the teacher model, we randomly sampled 15K examples from each of the two datasets, resulting in a total of 30K examples. After filtering invalid rationales, the dataset was reduced to a total of 23K examples. 90% of data was used for training, and the remaining 10% was used for validation. For training, we used 8 NVIDIA Tesla V100 GPUs with 16GB configurations. Hyperparameters for training are as follows: 3e-3 for learning rates, 5 epochs, 64 batch size. |
In particular, after re-warming, models fail to recover quickly from the performance hit due to rewarming the learning rate even when training on the same dataset. This motivates finding alternatives to learning rate schedules requiring re-warming in order to improve the efficiency of continual pre-training. |
QReCCWe use the test set of QReCC for the evaluation. Its test set includes some dialogue samples source from the QuAC dataset. To avoid overlap with other benchmark datasets, we remove those QuAC-sourced samples, resulting in 2805 user-agent turns. |
Footnote 2: Examples include (Africa, largest country, Sudan) and (It Happened One Night, director, Frank Capra) in TriviaQA [20], or (Teton Dam, collapse date, 06/05/1976) and (USA, Capital, Washington D.C.) in NaturalQuestions [22]. |
\(\bullet\)_More complex tasks can benefit more from careful prompt engineering on ChatGPT._ In the WikiFact and Colored Objects tasks, the designed prompts have greatly improved the performance of ChatGPT, _i.e._, from 23.61 to 28.47 on WikiFact and from 53.20 to 66.75 on Colored Objects. It indicates the necessity of prompt engineering for LLMs to perform well on complex tasks, since these tasks typically have specific output formats or require background knowledge. Our example prompts provide more detailed task description (_e.g._, output format and task goal), which can help ChatGPT better understand the complex task requirement for fulfilling it. |
Furthermore, KnowledGPT (Wang et al., 2023) is dedicated to augmenting raw text data with structured, semantically rich information through entity linking. This enrichment process not only structures the data more cohesively and makes it more amenable to queries but also boosts the model's retrieval efficiency. It leverages precise, linked knowledge to enhance the model's understanding and its ability to generate relevant responses, thereby improving its overall performance. |
entropy, suggesting that unlabeled task data contains new knowledge but is less diverse. For ETA-DACP-ppl and ETA-DACP-ent, the samples are uniform across the other two dimensions. |
Each model was trained for two epochs. Since both C4-Original and C4-ExactSubstr contain approximately 365M examples, we performed 152K steps with a batch size of 4800 (or approximately 2 epochs). C4-NearDup contains approximately 350M examples, we performed 146K steps (or approximately 2 epochs). On a 128-core TPU v3 pod slice, XL models trained on C4-Original and C4-ExactSubstr took approximately 131 hours (5.5 days) to train, while the XL model trained on C4-NearDup took approximately 126 hours to train. Like T5, models were trained with the Adafactor optimizer Shazeer and Stern (2018). A constant learning rate of 0.01 was used for the base models and 0.001 for the XL models. |
The hazard categories in the v1.0 taxonomy that do not have tests in the v0.5 benchmark are briefly described in Table 8. |
In simpler terms, the model can consistently give correct answers, outperforming a simple random guessing baseline. Naturally, if the knowledge score \(\mathcal{L}_{\mathcal{M},\mathcal{Q}}\) is higher for one model compared to another, then we assert that the former is more knowledgeable with regards to \(\mathcal{Q}\) compared to the latter. |
In Figure 0(a), we clearly see that models up to 410M parameters suffer from the saturation phenomenon, characterized as an increase of the in-domain loss in advanced training stages. |
For ExactSubstr, the entire dataset being deduplicated needs to be loaded onto memory: we leveraged the AWS x2iedn instances, which come with up to 2 TiB of memory in a single instance. |
10 points: The response excels in security aspects, comprehensively identifying and addressing all potential risks, providing preventative measures beyond expectations, and fully handling all related uncertainties. |
Chain-of-thought (CoT) [31] proposes to design better prompts to generate step-by-step solutions, which can lead to improved performance. Self-Consistency [34] also achieves remarkable performance on many reasoning benchmarks, which generates several possible answers from the model and selects the correct one based on majority vote [35]. In recent, [36] finds that process supervision with reinforcement learning significantly outperforms outcome supervision for solving challenging MATH problems. |
D4 parameters: The choice of parameters \(R_{proto}\) and \(R_{dedup}\) while using D4 will have impact on the performance of D4. Given limited compute, we are not able to sweep over these hyperparameters. Instead, we strategically choose these parameters: we first look at the highest value of \(R\) in SemDeDup that results in perplexity improvement across validation sets. We choose the "highest value" because the purpose of SemDeDup is to remove duplicate-driven clusters and low \(R\) with SemDeDup generally removes more than just templates/semantic duplicates. As seen in Section A.3, this generally occured with \(R_{dedup}=0.75\). Thus, we chose \(R_{dedup}=0.75\) and varied \(R_{proto}\) to obtain different data selection ratios for D4. |
Figure 1: **Overview of contributions.** AMIE is a conversational medical AI optimised for diagnostic dialogue. AMIE is instruction fine-tuned with a combination of real-world and simulated medical dialogues, alongside a diverse set of medical reasoning, question answering, and summarization datasets. Notably, we designed a self-play based simulated dialogue environment with automated feedback mechanisms to scale AMIE’s capabilities across various medical contexts and specialities. Specifically, this iterative self-improvement process consisted of two self-play loops: (1) An “inner” self-play loop, where AMIE leveraged in-context critic feedback to refine its behavior on simulated conversations with an AI patient agent; (2) An “outer” self-play loop where the set of refined simulated dialogues were incorporated into subsequent fine-tuning iterations. During online inference, AMIE used a chain-of-reasoning strategy to progressively refine its response conditioned on the current conversation to arrive at an accurate and grounded reply to the patient in each dialogue turn. We designed and conducted a blinded remote Objective Structured Clinical Examination (OSCE) with validated simulated patient actors interacting with AMIE or Primary Care Physicians (PCPs) via a text interface. Across multiple axes corresponding to both specialist physician (28 out of 32) and patient actor (24 out of 26) perspective, AMIE was rated as superior to PCPs while being non-inferior on the rest. |
While the validation sets in web-snapshots and non-web snapshots are clear (they are either standard open-sourced datasets, or derived from commonly used data), we expect that the "Instruct + Answers" data might be new to some readers. We provide a few examples of what this validation set looks like in Table A2. |
According to the task definition, existing tasks about language generation can be roughly categorized into language modeling, conditional text generation, and code synthesis tasks. Note that code synthesis is not a typical NLP task, we include it for discussion because it can be directly solved by a number of LLMs (trained on code data) in a similar generation approach as natural language text. |
Hutter LLM.Figures 8, 9 and 10 further illustrate our theory for simple Hutter LLM. |
Collected AnnotationsWe collected annotations for a total of 3,158 claims across 26 books. Table 14 shows the distribution of collected labels distribution of collected labels broken down by claim source (i.e., the model that generated the claim). |
With access to feedback from the environment, the task planner can accordingly refine its current plan and iteratively go through the "_planning - execution - refinement_" loop for better results. In this part, we summarizes three major refinement approaches in existing work. |
**Output**[**Retriee**]=Yes]<p>The Walking Dead (season 7)-The Walking Dead (season 7) The seventh season of "The Walking Dead", an American post-apocalypetric horror television series on AMC, premiered on October 23, 2016, and concluded on April 2, 2017, consisting of 16 episodes. Developed for television by Frank Darabout, the series is based on the eponymous series of comic books by Robert Kirkman, Tony Moore, and Charlie Adlard. |
Comparing with the Open-Source Models.In Table 1, we conduct a comprehensive comparison of our _WizardCoder_ with other open-source models on the HumanEval and MBPP benchmarks. In contrast to the results presented in Figure 1, we adhere to the approach outlined in previous studies [31] by generating n samples for each problem to estimate the pass@1 score. The findings presented in Table 1 clearly demonstrate that our _WizardCoder_ exhibits a substantial performance advantage over all the open-source models. |
where the second equality follows from Lemma 2. We have proven that the operator \(f\) maps all reward functions from a particular equivalence class to the same reward function. Next, we show that for every equivalence class of reward functions, the reward function that has the reparameterization outlined in Theorem 1 is unique. |
1. _Forecasting only on questions when the crowd prediction falls between \(.3\) and \(.7\)_. Here, our system attains a Brier score of \(.238\) (crowd aggregate: \(.240\)). This comprises 51% of forecasts and 56% of questions. 2. _Forecasting only on earlier retrieval dates (1, 2, and 3)_. Our system's Brier score in this setting is \(.185\) (crowd aggregate: \(.161\)). This comprises 66% of forecasts and 100% of questions. 3. _Forecasting only when the retrieval system provides at least \(5\) relevant articles._ Under this condition, our system's Brier score is \(.175\) (crowd aggregate: \(.143\)). This makes up 84% of forecasts and 94% of questions. 4. Under all three conditions, our system attains Brier score \(.240\) (crowd aggregate: \(.247\)). This comprises 22% of forecasts and 43% of questions. |
Figure 1: Performance of Pythia models on the Pile. On the left, we compare training dynamics of models from 14M (top) to 410M (bottom) parameters, displaying darker shades as we approach the minimal value. On the right, we fit a power law on larger models and find that final checkpoints of smaller models underperform compared to predictions. |
5. Now, we have another instance of \(f\) and \(g\). Since \(f\) is odd and \(g\) is even, their product is an odd function. So, we can rewrite the expression as \(-[f(g(f(x)))]\). |
On any given question, each platform computes a community prediction that aggregates all individual forecasts. The prediction is dynamically updated and recorded as the forecasts are made. We source the records directly from the platforms (instead of computing them from scratch using the individual forecasts). For binary questions, we provide more details on the aggregation mechanisms as follows. |
As visualized in Appendix Figure 7, we find that on average there is little improvement in the LM's ability to predict arbitrary tokens. But, when we visualize the distribution of relative improvements, there is a disproportionate improvement on more difficult tokens. This reflects the idea that some text tokens are substantially harder and benefit more from careful thought. |
**Effects of Question Verification.** The question verification is used to further improve the generation quality. In our experiments, we found it can improve the performance on MATH benchmark, the results are shown in Table 5, while we do not see significantly impact on GSM8K dataset. |
Despite the considerable strides made by LLaMA and Alpaca in NLP, they exhibit inherent limitations concerning native support for Chinese language tasks. Their vocabularies contain only a few hundred Chinese tokens, substantially hindering their efficiency in encoding and decoding Chinese text. Building on our previous work with the Chinese BERT series (Cui et al., 2021) and Chinese minority-oriented multilingual pre-trained models (Yang et al., 2022), in this technical report, we propose the development of Chinese LLaMA and Alpaca models with enhanced capabilities for understanding and generating Chinese content. We extend the original LLaMA's vocabulary with an additional 20,000 Chinese tokens, significantly improving its proficiency in processing and generating Chinese text. To ensure efficient training and deployment of these models, we employ the Low-Rank Adaptation (LoRA) approach (Hu et al., 2021), enabling us to train and fine-tune the models without excessive computational costs. We anticipate our preliminary study to enhance the Chinese understanding and generation capabilities of LLaMA and Alpaca serves as a foundation for researchers aiming to adapt these models to other languages. By showcasing the feasibility and effectiveness of our approach, we offer insights and methodologies that can be employed to extend vocabularies and improve the performance of LLaMA and Alpaca models in various languages. An overview of the proposed models is depicted in Figure 1. |
Fig. 1: The trends of the cumulative numbers of arXiv papers that contain the keyphrases _“language model”_ (since June 2018) and _“large language model”_ (since October 2019), respectively. The statistics are calculated using exact match by querying the keyphrases in title or abstract by months. We set different x-axis ranges for the two keyphrases, because “language models” have been explored at an earlier time. We label the points corresponding to important landmarks in the research progress of LLMs. A sharp increase occurs after the release of ChatGPT: the average number of published arXiv papers that contain _“large language model”_ in title or abstract goes from 0.40 per day to 8.58 per day (Figure 1(b)). |
In this section, we first show the main experimental results for the CKL Benchmark. Then, since multiple steps of continual knowledge learning, i.e., CKL are needed for training a true, ever-changing LM, we explore the effects of multiple CKL phases as well as how epochs, corpus size, and the total number of training steps affect CKL. We further explore how learning rates affect CKL in Appendix E, how continual pretraining on \(D_{1}\) affects the performance of KILT tasks which re quire knowledge from \(D_{0}\) in Appendix F, how CKL methods transfer across LM architectures in Appendix G, and how the prediction outputs change during CKL in Appendix H. |
Figure 3 reports validation losses for the \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) datasets throughout the continual pre-training of all models. The top row of plots reports results for the weak distribution shift (300B Pile\(\rightarrow\)300B SP), while the bottom row reports results for the strong distribution shift (300B Pile\(\rightarrow\)200B Ger.). For both shifts, the \(\eta_{\text{min}}\) constant learning rate model achieves the least forgetting on \(\mathcal{D}_{0}\). It also adapts the least on \(\mathcal{D}_{1}\) for the strong shift, however, for the weak shift it adapts more than the \(\eta_{\text{max}}\) constant baseline. When comparing these baselines to the models that re-warm and re-decay on both datasets, we observe that the latter models adapt better to the new dataset by a significant margin for both distribution shifts. This shows that re-warming and re-decaying are necessary to maximize adaptation to the new dataset when continually pre-training LLMs. Among the models that re-warm and re-decay the LR, we observe that varying the learning rate causes small differences in adaptation and forgetting: higher values of \(\eta_{\text{max}}\) lead to more forgetting and more adaptation while the opposite is true for lower values. When comparing the baselines to the union-trained baseline, we observe that the final validation loss for \(\mathcal{D}_{1}\) is significantly higher than the union-trained model on both distribution shifts. This is also the case for \(\mathcal{D}_{1}\) on the weak distribution shift, but interestingly for the strong distribution shift, the constant baselines achieve lower \(\mathcal{D}_{1}\) validation loss than the union-trained model. We hypothesize that this is due to the stronger distribution shift generally enhancing adaptation and exacerbating forgetting in the context of LLMs. When comparing models continually pre-trained with re-warming and re-decaying to the union baseline, we note that these models adapt better (lower final validation loss) to \(\mathcal{D}_{1}\) than the union baseline. However, these models experience significant forgetting on \(\mathcal{D}_{0}\), showing the need for replay to make these models competitive with the union baseline. |
Fertility, the most common metric to evaluate a tokenizer's performance (Scao et al., 2022; Stollenwerk, 2023; Rust et al., 2021), is defined as the average number of tokens that are required to represent a word or document. For a tokenizer \(T\) and dataset \(A\), the fertility can be calculated as the number of tokens in \(A\) (when \(T\) is applied) divided by the number of words in \(A\). We calculate the fertility on a held-out set (10,000 documents), which was not used for the tokenizer training. For calculating the words of a document, we used whitespace splitting. Higher fertility scores correspond to weaker compression capabilities of the tokenizer. |
**Overall Comparison with RL Formulation**. Following the discussion in Section 5.2.3 (the part related to RL training), the text generation problem can be formulated as a decision-making process based on RL. Taking a prompt as input, the task of a LLM is to generate a text completion that appropriately responds to the prompt. This task would be completed step by step. At each step, an agent (_i.e.,_ LLM) will perform an action (_i.e.,_ generating a token) according to the policy (_i.e.,_ the generative probability distribution of LLM) conditioned on the current state (currently generated token sequence and other available context information). It is expected that a high-quality output text would be produced by the LLM, which can earn a large reward score based on the entire response. Overall, RLHF and SFT can be considered as two different training approaches to optimizing the above decision making process for LLMs. Specially, RLHF firstly learns the reward model, and then employs it to improve the LLM with RL training (_e.g.,_ PPO). As a comparison, SFT adopts a teacher-forcing approach, which directly optimizes the likelihood of a demonstration output. Such a token-level training way essentially does _behavior cloning_ (a special algorithm of imitation learning [393]): it utilizes the expert's action (_i.e.,_ the target token at each step) as the supervision label and directly learns to imitate the demonstrations from experts without specifying a reward model as in typical RL algorithms. To learn the desired policies, SFT adopts a "local" optimization way (_i.e.,_ token-level loss) based on demonstration data, while RLHF takes a "global" optimization way (_i.e.,_ text-level loss) by involving human preference. More theoretical analysis about imitation learning and reinforcement learning can be referred to the related RL literature [393, 394]. |
The most primitive chain-of-thought is a chain structure that describes intermediate reasoning steps in natural language. In this section, we introduce structural variants that modify the original chain structure, including chain structure variants, tree structure variants, and graph structure variants. |
We collect the problems from various sources e.g. ziya (Zhang et al., 2022), goakao, and CIF-Bench (Li et al., 2024) to form hard-case Chinese instructions understanding and following evaluation benchmark (CHC-Bench in short) The categories of problems in CHC-Bench include writing, humanity and history, science, math, reading comprehension, role-playing, and hard cases of Chinese understanding (i.e. Chinese word pronunciation, ancient Chinese language understanding, etc.). |
Figure 2: Example summary generated by Claude-3-Opus and claims extracted by GPT-4. |
Large Language Models (LLMs), such as GPT-3 Brown et al. (2020), LLaMA Touvron et al. (2023), and PaLM Chowdhery et al. (2023), have demonstrated remarkable capabilities as general-purpose assistants. The cornerstone of this achievement is instruction tuning, which significantly enhances the capabilities and controllability of LLMs through training on datasets composed of instruction-output pairs Zhang et al. (2023). This technique effectively aligns the models' training objectives with human intentions, thereby ensuring that the models can interpret and execute human instructions both effectively and safely. Therefore, the availability of high-quality instruction tuning datasets is crucial for LLMs to operate as efficient and dependable assistants. |
Figure 17: The fitted scaling law plotted along the data amount axis for each model and each task. The fitted result is satisfying except the last checkpoints of 0.11B and 0.25B model. |
In this case it is easy to guess that the third option is the most likely, even without knowledge of this specific strike. |
et al., 2023; Bai et al., 2023; Almazrouei et al., 2023), which gradually decreases the learning rate following a cosine curve after it reaches its maximum after the warmup stage. |
Although numerous research about instruction tuning has been conducted, a few work focused on improving RAG or context awareness generation for QA. Lin et al. (2023) introduced a retrieval-augmented instruction tuning method, which appends top-\(k\) retrieved chunks for LLM fine-tuning. Wang et al. (2023) applied instruction tuning after retrieval-augmented pretraining. In contrast, we propose a two-stage instruction tuning method to improve generation with retrival or provided context. We find that appending top-\(k\) retrieved chunks for LLM fine-tuning does not help for a wide range of conversation QA tasks (see SS6.3 for details). |
In many jurisdictions, the hazards in our taxonomy are illegal. For instance, child abuse, physical violence, and sexual assault are illegal in most countries across the globe33. However, (i) the hazards are not always illegal and (ii) our policies are not defined solely on the basis of the hazards' legality. |
Instruction tuning aims to train large language models to generate responses that align with input instructions, thereby enable LLMs with conversational and task execution capabilities. Compared with standard LLM, SFT allows the behavior of the model to be more controllable and predictable, thereby achieving the purpose of aligning human will. Methods to build instruction tuning datasets include: (1) Pure manual annotation Conover et al. (2023). This method completely constructs instructions and answers manually, which is very time-consuming and laborious; (2) Converted from existing datasets Mishra et al. (2022); Sanh et al. (2022); Chung et al. (2022). Some studies use supervised data sets from NLP tasks to construct instruction tuning data; (3) Automatically generated using LLM Honovich et al. (2022); Wang et al. (2023); Xu et al. (2023); Ji et al. (2023); Xu et al. (2023). Others use existing LLMs to generate instruction tuning data. A common practice is to first manually annotate high-quality seed datasets, and then use LLM to expand the seed instructions and corresponding outputs, which can generate large-scale instruction tuning data with very little human annotation. However, the quality cannot be guaranteed and there is a certain amount of noisy data, which can lead to hallucinations. |
The RAG paradigm organizes research within the domain, offering a straightforward yet robust framework to enhance LLM performance. Central to RAG is its search mechanism, crucial for generating high-quality outcomes. Therefore, this paradigm is structured into four main phases from a retrieval perspective: pre-retrieval, retrieval, post-retrieval, and generation. Both single-hop and multi-hop retrieval approaches, encompassing iterative retrieve-generate cycles, follow this four-phase structure. Figure 3 is the taxonomy tree of RAG's core techniques. |
**Dealing with Ambiguities:** Resolving or excluding ambiguous or contradictory data that might confuse the model during training. This can help the model to provide more definite and reliable answers. |
Figure 4: **Combining multiple styles:** Perplexity across all domains of the Pile comparing combining multiple styles of synthetic data. Models are 1.3B parameters trained for a total of 150B tokens. We see small perplexity improvements from combining multiple styles. |
Third, we did not investigate whether the results obtained could be extrapolated to larger model sizes, which we leave to future works. However, our finding that the BPE-SP-33 tokenizer is the best-performing tokenizer for the monolingual setting and the fact that this tokenizer has been used for training state-of-the-art models up to 65B (Touvron et al.) might indicate that our results also transfer to larger model sizes. |
**QUESTION:** Rory orders 2 subs for $7.50 each, 2 bags of chips for $1.50 each and 2 cookies for $1.00 each for delivery. There's a 20% delivery fee added at check out and she wants to add a $5.00 tip. What will her delivery order cost? |
Tests are designed to be simpleTest items have been designed by a team of AI safety experts to be clear cut, easy to interpret, and easy to assess. They are short and do not use hazard-specific language, are unambiguous and independent of current events, and only test for single-turn interactions. They are also free of adversarial prefixes or prompting tricks that a user may use to elicit harmful behavior because the personas that we tested for are all "unsophisticated". However, this limits their relevance for testing more sophisticated users. We will address this in the future by working more closely with domain experts, and taking inspiration from unstructured datasets of real-world LM interactions [see 69, 70]. |
PersonA: Vascepa is a prescription fish oil medication used for treating adults with extremely high triglyceride levels. |
We conduct 4-bit quantization on our model. We do not quantize the parameters of embedding and layer normalization, since the performance of the model is sensitive to these parameters. Therefore, we only need to quantify each weight matrix. Denote the weight matrix as \(\mathbf{W}\in\mathbb{R}^{d_{out}\times d_{in}}\). We group every \(G\) consecutive parameter at the \(d_{in}\) dimension and form \(d_{in}/G\) group. Then we quantize each group of the parameters separately. |
<description>" and "<description> is <name>" often co-occur in pretraining datasets; if the former appears in a dataset, the latter is intuitively more likely to appear.4 This is because humans often vary the order of elements in a sentence or paragraph.5 Thus, a good meta-learner would increase the probability of an instance of "<description> is <name>" after being trained on "<name> is <description>". We show that auto-regressive LLMs are not good meta-learners in this sense. |
that attempts to exactly model the C4 web text. |
(Shumailov et al., 2023) define model collapse and attribute it to two mechanisms: finite sampling when training a model (leading to cut off of low-probability data) and function approximation errors (the model is not sufficiently expressive to model the true distribution). They observe (and, for a single Gaussian, prove) that upon iteratively resampling finite "training data" the generated distribution becomes more and more peaked. Other models studied empirically are mixtures of (two) Gaussians and VAEs on MNIST. To study language models, (Shumailov et al., 2023) iteratively fine tune Meta's OPT-125M model on wikidata2. For generation of new text they use a 5-way beam search, which, by its nature, (approximatively) generates only low-perplexity data. |
This work represents a fundamental step to a better understanding of the impact of the tokenizer on the models' downstream performance. We have shown that training tokenizers with a balanced share across languages achieve comparable low fertility and parity scores across all languages, which has important implications. Higher fertility results in up to 68% more computational costs during training and prevents the model from learning long-range dependencies in limited context windows. |
We now consider the problem of inverting the process: given the probability vector, we attempt to produce the prompt that led to these next-token probabilities. Given some unseen model \(f:\mathcal{V}^{T}\rightarrow\Delta^{|\mathcal{V}|-1}\) which gives next-token probabilities, we would like to learn to invert this function from samples: pairs of text prefixes and their associated next-token probability vectors \((x_{1:T}^{1},\mathbf{v}^{1})\ldots(x_{1:T}^{J},\mathbf{v}^{J})\). |
Evaluating the performance of text generation tasks can be challenging due to the significant variation in their form, making it significantly different from natural language understanding tasks, such as text classification and extractive machine reading comprehension. Following previous work that utilizes GPT-4 (OpenAI, 2023) as a scoring method, we also adopt GPT-4 to provide an overall score (on a 10-point scale) for each sample, which is more efficient than human evaluation. However, GPT-4 may not always provide accurate scores, so we perform manual checks on its ratings and adjust them if necessary. The manual checks ensure that the scores are consistent and reflect the true performance of the models being evaluated. |
Footnote 23: We utilize the symbol “\(\rightarrow\)” to represent the data order in data curriculum. For example, “2T webpage tokens \(\rightarrow\) 500B code tokens” means that the LLM is firstly trained with 2T webpage tokens and subsequently with 500B code data tokens. |
**MetaGPT.** Many multi-agent architectures allow agents to converse with one another while collaborating on a common problem. This conversational capability can lead to chatter between the agents that is superfluous and does not further the team goal. MetaGPT addresses the issue of unproductive chatter amongst agents by requiring agents to generate structured outputs like documents and diagrams instead of sharing unstructured chat messages [11]. |
Pre-trained transformer blocks are akin to neural network components one can choose to mix and match using an evolutionary architecture search algorithm. However, unlike NAS, we do not need to train the model, and can evaluate the candidates right away. In a sense, our work resonates with the goals of earlier works of morphology search such as NEAT [45] and more recent work such as Weight Agnostic Neural Networks [12] which evolved neural network structures with certain task-specific inductive biases, without the need to train the weight parameters using gradient descent. Similarly, this was how NAS methods such as SMASH [45] avoided costly inner-loop training through the application of a Hypernetwork [16] to estimate the weights of architectural candidates. |
Table 2 \(|\) Results on the R-GSM dataset: (a) accuracies on the full dataset; (b) for each model, the accuracies on the R-GSM subset where the original problems are correctly solved, thus the initial accuracy is 100% for all models. |
We have trained and tested models using (mixed precision) 16-bit floats. What happens if we quantize them to int8/int4 after training? We used the auto_gptq package, which is inspired by the GPTQ paper [10], for quantization. |
For each task and model, the scaling law \(L(N,D)\)'s fitness with real loss values along the data axis is ploted in Figure 17. |
To begin, **suppose that the API returned the top \(K\) logits**. Then we could recover the complete logit vector for an arbitrary prompt \(p\) by cycling through different choices for the logit bias and measuring the top-\(k\) logits each time. |
where \(\eta_{max}\) is the maximum learning rate and \(\eta_{min}\) is the minimum learning rate. |
Synthetic SFT Data GenerationSince the real data has been exhausted, we contemplate further scaling up SFT data using synthetically generated math questions. |
ure 4, we redo this calculation for different model sizes and we see that _overall_ efficiency gain increases with model size. Based on this, we can conservatively estimate that D4 would have overall efficiency gains of 20% for LLama-65B [50] and 22% for OPT-175B [59]. |
Footnote 4: Following Lightman et al. (2023), we utilize a subset of 500 test samples from the MATH benchmark for experimental efficiency. |
Not limited to the decoder-only architecture, these two decoding methods can be generally applied to encoder-decoder models and prefix decoder models in a similar way. |
mechanisms. The crowd prediction is a strong benchmark to compete with. For example, Metaculus (2023) shows that an ensemble of all forecasters consistently outperforms using just the top 5, 10,..., 30 best forecasters (based on past scores). In this work, we compare our system performance to the crowd aggregates. |
eliminates the loss for undesired tokens without incurring additional costs during pretraining, making our approach both efficient and easily integrated. |
of two documents_) [775], or listwise ranking (_sorting a subset of candidate documents_) [777]. The essence of these methods lies in the special design of instructions for text reranking, such as sliding window strategy for document lists [774, 778], setwise selection prompting [779], fine-grained relevance labels incorporation [780], and pairwise comparison prompting [775]. In addition, recent efforts employ LLMs to generate intermediate texts (_e.g._, URLs) as retrieval results using few-shot demonstrations [781]. To further enhance the model performance, LLMs can be specially fine-tuned as backbones for reranking [782, 783] or retrieval (including dense retrieval [54] and model-based retrieval [784, 785]), similar to the fine-tuning process for traditional PLM-based IR models [782]. However, fine-tuning LLMs as IR models entails considerable expenses given the huge parameter scale of LLMs. |
The applications of **WRAP** lies in both paradigms--(i) low-resourced data settings such as a language model for Finnish language (Luukkonen et al., 2023), and (ii) data-rich settings such as training on the common crawl. In the former, there is no alternative option of naively gathering more data, and hence, synthetic data is a natural solution that should outperform training on in-domain data alone. However, there is a significant interest in training language models on English, or more broadly, general web data. Is using synthetic data a viable option even in this paradigm? |
Due to the difficulty of standardizing evaluations for large models and the lack of publicly available prompts and test codes for many models' evaluations, we try our best to adapt the evaluation methods to suit various model types. Specifically, we start from a standardized input prompt during testing and adjust it according to each model's appropriate input-output template. The **evaluation scripts and prompts are also open-source** in our repository, and we welcome developers to continually improve our evaluation methods. |
We show in Table 2(a) the ablation results. The top part of the table shows results for training ablations, and the bottom part is for inference ablations. We see that all components play important roles. We also observe a large performance gap between Self-Rag and No Retriever or Critic baselines across tasks, indicating that training an LM with those models largely contributes to the performance gain of Self-Rag. Using the top passages regardless of their relevance (Retrieve top 1) as in conventional RAG approaches causes a large drop in PopQA and ASQA, and removing \(\boxed{Isstr}\) during the beam search results hurts performance on ASQA. This demonstrates the effectiveness of Self-Rag's capabilities of carefully selecting generations based fine-grained multiple criterion, instead of naively using all of the top passages from the retrieval model or solely depending on relevance scores. |
Mathematically, let \(\mathcal{Q}=\{q_{n}\}_{n=1}^{N}\) be a set of \(N\) multiple choice factual questions, where each question has \(L\) possible answers and exactly one correct answer. Let \(\mathcal{A}=\{(a_{n}^{1},\ldots,a_{n}^{L})\}_{n=1}^{N}\) be the corresponding set of possible answers, and \(\mathcal{C}=\{c_{n}\}_{n=1}^{N}\) be the correct answers. |
**Order effect of language models.** Some prior works show that language models are able to understand permuted texts to some extent, i.e., after a random permutation of words, models usually preserve a reasonable performance (Abdou et al., 2022; Sinha et al., 2020). Moreover, Cao et al. (2023) shows that even when a large fraction of words are scrambled, GPT-4 still achieves decent performance on several reasoning benchmarks. In contrast to permuted texts in these works that are typically unnatural and nonsensical, our premise order permutations do not alter the semantic meaning and remain syntactically valid (we manually verify this). Nevertheless, we demonstrate that LLM reasoning performance is highly brittle to the ordering of the premises. |
We collect five conversational QA datasets with long documents which cannot be directly fitted into LLMs with a sequence length of 4K tokens. Hence, we run a multi-turn retriever to get top-5 relevant chunks as the inputs (experiments can be found in SS4.3). |
Traditional Chinese outputs. Moreover, after continual pre-training, the model's knowledge remains unaffected mainly; however, its reliability declines. |
This technical report presents our first release of OLMo, a state-of-the-art, truly open language model and its framework to build and study the science of language modeling. Unlike most prior efforts that have only released model weights and inference code, we release OLMo and the whole framework, including training data and training and evaluation code. Soon, we will also release training logs, ablations, findings and Weights & Biases logs. We are also exploring the adaptation of OLMo with instruction tuning and different flavors of RLHF. We are going to release the adapted models as well as all of our model adaptation code and data. |
CoT prompting is first proposed as an extension of ICL [33], which augments each demonstration \(\langle\)_input, output_\(\rangle\) as \(\langle\)_input, CoT, output_\(\rangle\). A _CoT_ is a series of intermediate reasoning steps for connecting the _input_ and _output_. With these augmented demonstrations, LLMs can follow them to generate CoTs and the answer for a new input. However, unlike \(\langle\)_input, output_\(\rangle\) pairs in ICL, CoTs are difficult to obtain and usually require human annotation. Fortunately, it has been found that LLMs can be triggered to generate CoTs through simple instructions like _"Let's think step by step."_[505], making CoT prompting easy to use. There are also alternative magic prompts that can elicit the ability of CoT reasoning and further improve the performance of LLMs, such as _"Take a deep breath and work on this problem step-by-step."_[473]. |
Finally, we compare the mean tree depth (measured by the mean over sentences of the depth of the dependency tree), and mean dependency distance (measured as the average dependency distance of any pair of words within a sentence) in Figure 11, which have been shown to be good measures of syntactic difficulty Futrell et al. (2015); Gibson et al. (2000); Oya (2021). We find similar trends as for reading level and TTR diversity where mediums style increase depth, mdd, and syntactic complexity in general. We find again that QA style reduces this complexity. |
- Engaging with opponents: High. Trump is known for his confrontational style, which could play out to his advantage in a debate setting. |
**ColossalAI** library [226] provides a collection of parallel components. It aims to support developers to write their distributed deep learning models just like how they write their model on their laptop. They provide user-friendly tools to kickstart distributed training and inference in a few lines. In terms of Parallelism strategies, they support: Data Parallelism, Pipeline Parallelism, Sequence Parallelism, Zero Redundancy Optimizer (ZeRO) [140], and Auto-Parallelism. |