{ "best_metric": null, "best_model_checkpoint": null, "epoch": 46.0, "eval_steps": 500, "global_step": 3726, "is_hyper_param_search": false, "is_local_process_zero": true, "is_world_process_zero": true, "log_history": [ { "epoch": 1.0, "eval_B": { "f1-score": 0.8331822302810515, "precision": 0.7901977644024075, "recall": 0.8811121764141898, "support": 1043.0 }, "eval_I": { "f1-score": 0.9471294753175105, "precision": 0.9443616777446711, "recall": 0.9499135446685879, "support": 17350.0 }, "eval_O": { "f1-score": 0.8992868897421833, "precision": 0.9103731674811195, "recall": 0.8884673748103187, "support": 9226.0 }, "eval_accuracy": 0.9267895289474637, "eval_loss": 0.20478281378746033, "eval_macro avg": { "f1-score": 0.8931995317802484, "precision": 0.8816442032093993, "recall": 0.9064976986310321, "support": 27619.0 }, "eval_runtime": 4.6128, "eval_samples_per_second": 17.343, "eval_steps_per_second": 2.168, "eval_weighted avg": { "f1-score": 0.9268447919078652, "precision": 0.9271861479533134, "recall": 0.9267895289474637, "support": 27619.0 }, "step": 81 }, { "epoch": 2.0, "eval_B": { "f1-score": 0.8682547399124939, "precision": 0.8806706114398422, "recall": 0.8561840843720039, "support": 1043.0 }, "eval_I": { "f1-score": 0.9590983771942466, "precision": 0.9592919333448654, "recall": 0.9589048991354466, "support": 17350.0 }, "eval_O": { "f1-score": 0.9226525313717006, "precision": 0.9208594256100194, "recall": 0.9244526338608281, "support": 9226.0 }, "eval_accuracy": 0.9435171439950758, "eval_loss": 0.16574698686599731, "eval_macro avg": { "f1-score": 0.9166685494928136, "precision": 0.9202739901315756, "recall": 0.9131805391227594, "support": 27619.0 }, "eval_runtime": 4.6238, "eval_samples_per_second": 17.302, "eval_steps_per_second": 2.163, "eval_weighted avg": { "f1-score": 0.9434932036816766, "precision": 0.9434846863370583, "recall": 0.9435171439950758, "support": 27619.0 }, "step": 162 }, { "epoch": 3.0, "eval_B": { "f1-score": 0.8791725434884815, "precision": 0.8625461254612546, "recall": 0.8964525407478428, "support": 1043.0 }, "eval_I": { "f1-score": 0.9588372159825352, "precision": 0.9557349825345015, "recall": 0.9619596541786744, "support": 17350.0 }, "eval_O": { "f1-score": 0.9204284621270084, "precision": 0.9282407407407407, "recall": 0.9127465857359636, "support": 9226.0 }, "eval_accuracy": 0.9430464535283681, "eval_loss": 0.15394224226474762, "eval_macro avg": { "f1-score": 0.9194794071993417, "precision": 0.9155072829121655, "recall": 0.9237195935541602, "support": 27619.0 }, "eval_runtime": 4.6233, "eval_samples_per_second": 17.303, "eval_steps_per_second": 2.163, "eval_weighted avg": { "f1-score": 0.9429985029052194, "precision": 0.9430314866542512, "recall": 0.9430464535283681, "support": 27619.0 }, "step": 243 }, { "epoch": 4.0, "eval_B": { "f1-score": 0.8800746616892207, "precision": 0.8572727272727273, "recall": 0.9041227229146692, "support": 1043.0 }, "eval_I": { "f1-score": 0.9537844036697248, "precision": 0.9488876212207644, "recall": 0.9587319884726225, "support": 17350.0 }, "eval_O": { "f1-score": 0.9100192149327477, "precision": 0.9220157970853265, "recall": 0.8983308042488619, "support": 9226.0 }, "eval_accuracy": 0.9364929939534379, "eval_loss": 0.19801080226898193, "eval_macro avg": { "f1-score": 0.9146260934305644, "precision": 0.9093920485262728, "recall": 0.9203951718787179, "support": 27619.0 }, "eval_runtime": 4.6132, "eval_samples_per_second": 17.342, "eval_steps_per_second": 2.168, "eval_weighted avg": { "f1-score": 0.9363812792925563, "precision": 0.9364514800186444, "recall": 0.9364929939534379, "support": 27619.0 }, "step": 324 }, { "epoch": 5.0, "eval_B": { "f1-score": 0.883877159309021, "precision": 0.8847262247838616, "recall": 0.8830297219558965, "support": 1043.0 }, "eval_I": { "f1-score": 0.9591500482927107, "precision": 0.9456643513331839, "recall": 0.9730259365994236, "support": 17350.0 }, "eval_O": { "f1-score": 0.9191176470588236, "precision": 0.9454503781801513, "recall": 0.8942120095382614, "support": 9226.0 }, "eval_accuracy": 0.9432999022412107, "eval_loss": 0.20074255764484406, "eval_macro avg": { "f1-score": 0.9207149515535183, "precision": 0.9252803180990656, "recall": 0.9167558893645271, "support": 27619.0 }, "eval_runtime": 4.6092, "eval_samples_per_second": 17.356, "eval_steps_per_second": 2.17, "eval_weighted avg": { "f1-score": 0.9429348139614955, "precision": 0.9432916158141273, "recall": 0.9432999022412107, "support": 27619.0 }, "step": 405 }, { "epoch": 6.0, "eval_B": { "f1-score": 0.8897338403041826, "precision": 0.882186616399623, "recall": 0.8974113135186961, "support": 1043.0 }, "eval_I": { "f1-score": 0.9624160683661308, "precision": 0.9707969272268809, "recall": 0.954178674351585, "support": 17350.0 }, "eval_O": { "f1-score": 0.9297955261331482, "precision": 0.9161493950552341, "recall": 0.9438543247344461, "support": 9226.0 }, "eval_accuracy": 0.9485861182519281, "eval_loss": 0.16565027832984924, "eval_macro avg": { "f1-score": 0.9273151449344872, "precision": 0.9230443128939126, "recall": 0.9318147708682424, "support": 27619.0 }, "eval_runtime": 4.6219, "eval_samples_per_second": 17.309, "eval_steps_per_second": 2.164, "eval_weighted avg": { "f1-score": 0.9487745648174827, "precision": 0.9491959030765336, "recall": 0.9485861182519281, "support": 27619.0 }, "step": 486 }, { "epoch": 6.17, "grad_norm": 4.767655849456787, "learning_rate": 1.7530864197530865e-05, "loss": 0.1585, "step": 500 }, { "epoch": 7.0, "eval_B": { "f1-score": 0.8802588996763754, "precision": 0.85, "recall": 0.912751677852349, "support": 1043.0 }, "eval_I": { "f1-score": 0.958423523458803, "precision": 0.9354661691269275, "recall": 0.9825360230547551, "support": 17350.0 }, "eval_O": { "f1-score": 0.9139526911210146, "precision": 0.9664088931851136, "recall": 0.8668978972469109, "support": 9226.0 }, "eval_accuracy": 0.9412723125384699, "eval_loss": 0.2591131329536438, "eval_macro avg": { "f1-score": 0.9175450380853977, "precision": 0.917291687437347, "recall": 0.9207285327180049, "support": 27619.0 }, "eval_runtime": 4.617, "eval_samples_per_second": 17.327, "eval_steps_per_second": 2.166, "eval_weighted avg": { "f1-score": 0.9406164485555297, "precision": 0.9425749115781907, "recall": 0.9412723125384699, "support": 27619.0 }, "step": 567 }, { "epoch": 8.0, "eval_B": { "f1-score": 0.8687910028116214, "precision": 0.84967919340055, "recall": 0.8887823585810163, "support": 1043.0 }, "eval_I": { "f1-score": 0.9595010203788336, "precision": 0.956997878562009, "recall": 0.9620172910662824, "support": 17350.0 }, "eval_O": { "f1-score": 0.9212035166275324, "precision": 0.9282491471332673, "recall": 0.9142640364188164, "support": 9226.0 }, "eval_accuracy": 0.9432999022412107, "eval_loss": 0.2693409323692322, "eval_macro avg": { "f1-score": 0.9164985132726624, "precision": 0.9116420730319422, "recall": 0.9216878953553717, "support": 27619.0 }, "eval_runtime": 4.6153, "eval_samples_per_second": 17.333, "eval_steps_per_second": 2.167, "eval_weighted avg": { "f1-score": 0.9432823550422136, "precision": 0.9433417293609165, "recall": 0.9432999022412107, "support": 27619.0 }, "step": 648 }, { "epoch": 9.0, "eval_B": { "f1-score": 0.8645343367826904, "precision": 0.8485687903970452, "recall": 0.8811121764141898, "support": 1043.0 }, "eval_I": { "f1-score": 0.9567003348057282, "precision": 0.934535260814599, "recall": 0.9799423631123919, "support": 17350.0 }, "eval_O": { "f1-score": 0.9104673003585861, "precision": 0.9586479683567062, "recall": 0.8668978972469109, "support": 9226.0 }, "eval_accuracy": 0.9384481697382236, "eval_loss": 0.3087053894996643, "eval_macro avg": { "f1-score": 0.910567323982335, "precision": 0.9139173398561168, "recall": 0.9093174789244975, "support": 27619.0 }, "eval_runtime": 4.6156, "eval_samples_per_second": 17.333, "eval_steps_per_second": 2.167, "eval_weighted avg": { "f1-score": 0.9377758584761231, "precision": 0.9393435743356523, "recall": 0.9384481697382236, "support": 27619.0 }, "step": 729 }, { "epoch": 10.0, "eval_B": { "f1-score": 0.8803126526624329, "precision": 0.897410358565737, "recall": 0.8638542665388304, "support": 1043.0 }, "eval_I": { "f1-score": 0.9590983771942466, "precision": 0.9592919333448654, "recall": 0.9589048991354466, "support": 17350.0 }, "eval_O": { "f1-score": 0.9220456265542221, "precision": 0.9197584124245038, "recall": 0.9243442445263386, "support": 9226.0 }, "eval_accuracy": 0.9437705927079184, "eval_loss": 0.2796257436275482, "eval_macro avg": { "f1-score": 0.9204855521369671, "precision": 0.9254869014450354, "recall": 0.9157011367335386, "support": 27619.0 }, "eval_runtime": 4.641, "eval_samples_per_second": 17.238, "eval_steps_per_second": 2.155, "eval_weighted avg": { "f1-score": 0.9437458232244597, "precision": 0.9437490553802075, "recall": 0.9437705927079184, "support": 27619.0 }, "step": 810 }, { "epoch": 11.0, "eval_B": { "f1-score": 0.8830917874396135, "precision": 0.8899707887049659, "recall": 0.8763183125599233, "support": 1043.0 }, "eval_I": { "f1-score": 0.9574634882843123, "precision": 0.9647182727751448, "recall": 0.9503170028818444, "support": 17350.0 }, "eval_O": { "f1-score": 0.9200619426496501, "precision": 0.9067466582465004, "recall": 0.9337741166269239, "support": 9226.0 }, "eval_accuracy": 0.9419964517180202, "eval_loss": 0.3198166787624359, "eval_macro avg": { "f1-score": 0.9202057394578587, "precision": 0.9204785732422037, "recall": 0.9201364773562305, "support": 27619.0 }, "eval_runtime": 4.6345, "eval_samples_per_second": 17.262, "eval_steps_per_second": 2.158, "eval_weighted avg": { "f1-score": 0.942161111514465, "precision": 0.9425303680165918, "recall": 0.9419964517180202, "support": 27619.0 }, "step": 891 }, { "epoch": 12.0, "eval_B": { "f1-score": 0.871306005719733, "precision": 0.8663507109004739, "recall": 0.8763183125599233, "support": 1043.0 }, "eval_I": { "f1-score": 0.9576855173984852, "precision": 0.9464205312922107, "recall": 0.9692219020172911, "support": 17350.0 }, "eval_O": { "f1-score": 0.9162135168127844, "precision": 0.9386084583901774, "recall": 0.8948623455451984, "support": 9226.0 }, "eval_accuracy": 0.9408740359897172, "eval_loss": 0.35264310240745544, "eval_macro avg": { "f1-score": 0.9150683466436677, "precision": 0.917126566860954, "recall": 0.9134675200408043, "support": 27619.0 }, "eval_runtime": 4.6326, "eval_samples_per_second": 17.269, "eval_steps_per_second": 2.159, "eval_weighted avg": { "f1-score": 0.9405699625961891, "precision": 0.9407871989028143, "recall": 0.9408740359897172, "support": 27619.0 }, "step": 972 }, { "epoch": 12.35, "grad_norm": 0.3602512776851654, "learning_rate": 1.506172839506173e-05, "loss": 0.021, "step": 1000 }, { "epoch": 13.0, "eval_B": { "f1-score": 0.877092300334768, "precision": 0.875, "recall": 0.8791946308724832, "support": 1043.0 }, "eval_I": { "f1-score": 0.9597242793665263, "precision": 0.9486991778353419, "recall": 0.9710086455331413, "support": 17350.0 }, "eval_O": { "f1-score": 0.9203392649259936, "precision": 0.9419040054464994, "recall": 0.8997398655972252, "support": 9226.0 }, "eval_accuracy": 0.943734385748941, "eval_loss": 0.3594338297843933, "eval_macro avg": { "f1-score": 0.9190519482090961, "precision": 0.9218677277606138, "recall": 0.9166477140009498, "support": 27619.0 }, "eval_runtime": 4.6042, "eval_samples_per_second": 17.376, "eval_steps_per_second": 2.172, "eval_weighted avg": { "f1-score": 0.943447393984779, "precision": 0.9436461164304495, "recall": 0.943734385748941, "support": 27619.0 }, "step": 1053 }, { "epoch": 14.0, "eval_B": { "f1-score": 0.8754052802223252, "precision": 0.8467741935483871, "recall": 0.9060402684563759, "support": 1043.0 }, "eval_I": { "f1-score": 0.9574973153224439, "precision": 0.9392880904856953, "recall": 0.9764265129682997, "support": 17350.0 }, "eval_O": { "f1-score": 0.9130164471825015, "precision": 0.9539388213062477, "recall": 0.8754606546715803, "support": 9226.0 }, "eval_accuracy": 0.9400412759332344, "eval_loss": 0.3840242028236389, "eval_macro avg": { "f1-score": 0.9153063475757568, "precision": 0.9133337017801101, "recall": 0.9193091453654186, "support": 27619.0 }, "eval_runtime": 4.6182, "eval_samples_per_second": 17.323, "eval_steps_per_second": 2.165, "eval_weighted avg": { "f1-score": 0.9395385738014427, "precision": 0.9406884180878825, "recall": 0.9400412759332344, "support": 27619.0 }, "step": 1134 }, { "epoch": 15.0, "eval_B": { "f1-score": 0.8797709923664121, "precision": 0.8755935422602089, "recall": 0.8839884947267498, "support": 1043.0 }, "eval_I": { "f1-score": 0.9603688219460511, "precision": 0.954302299112224, "recall": 0.9665129682997118, "support": 17350.0 }, "eval_O": { "f1-score": 0.922722283205269, "precision": 0.9346230820547031, "recall": 0.9111207457186213, "support": 9226.0 }, "eval_accuracy": 0.9448930084362215, "eval_loss": 0.3762665390968323, "eval_macro avg": { "f1-score": 0.9209540325059108, "precision": 0.9215063078090453, "recall": 0.920540736248361, "support": 27619.0 }, "eval_runtime": 4.586, "eval_samples_per_second": 17.444, "eval_steps_per_second": 2.181, "eval_weighted avg": { "f1-score": 0.944749483712443, "precision": 0.9447562007752335, "recall": 0.9448930084362215, "support": 27619.0 }, "step": 1215 }, { "epoch": 16.0, "eval_B": { "f1-score": 0.8858669256920835, "precision": 0.8976377952755905, "recall": 0.8744007670182167, "support": 1043.0 }, "eval_I": { "f1-score": 0.9586212929752427, "precision": 0.9701351590627397, "recall": 0.9473775216138328, "support": 17350.0 }, "eval_O": { "f1-score": 0.9235412474849094, "precision": 0.9027950310559006, "recall": 0.9452633860828095, "support": 9226.0 }, "eval_accuracy": 0.9439154205438285, "eval_loss": 0.36898088455200195, "eval_macro avg": { "f1-score": 0.9226764887174118, "precision": 0.923522661798077, "recall": 0.922347224904953, "support": 27619.0 }, "eval_runtime": 4.6024, "eval_samples_per_second": 17.382, "eval_steps_per_second": 2.173, "eval_weighted avg": { "f1-score": 0.9441554794131966, "precision": 0.9449027186622512, "recall": 0.9439154205438285, "support": 27619.0 }, "step": 1296 }, { "epoch": 17.0, "eval_B": { "f1-score": 0.8863309352517986, "precision": 0.8867562380038387, "recall": 0.8859060402684564, "support": 1043.0 }, "eval_I": { "f1-score": 0.9630614486386496, "precision": 0.9659090909090909, "recall": 0.9602305475504322, "support": 17350.0 }, "eval_O": { "f1-score": 0.9296685529506872, "precision": 0.9245363918962375, "recall": 0.9348580099718188, "support": 9226.0 }, "eval_accuracy": 0.9489481878417032, "eval_loss": 0.3569599986076355, "eval_macro avg": { "f1-score": 0.9263536456137119, "precision": 0.9257339069363891, "recall": 0.9269981992635691, "support": 27619.0 }, "eval_runtime": 4.6143, "eval_samples_per_second": 17.337, "eval_steps_per_second": 2.167, "eval_weighted avg": { "f1-score": 0.9490090650954502, "precision": 0.9490996138580476, "recall": 0.9489481878417032, "support": 27619.0 }, "step": 1377 }, { "epoch": 18.0, "eval_B": { "f1-score": 0.8778409090909091, "precision": 0.8671655753040225, "recall": 0.8887823585810163, "support": 1043.0 }, "eval_I": { "f1-score": 0.9557028055829692, "precision": 0.9355230471984544, "recall": 0.9767723342939482, "support": 17350.0 }, "eval_O": { "f1-score": 0.9098012570069645, "precision": 0.9524599881446354, "recall": 0.8707999132885325, "support": 9226.0 }, "eval_accuracy": 0.938049893189471, "eval_loss": 0.4151785373687744, "eval_macro avg": { "f1-score": 0.9144483238936143, "precision": 0.9183828702157042, "recall": 0.9121182020544989, "support": 27619.0 }, "eval_runtime": 4.6304, "eval_samples_per_second": 17.277, "eval_steps_per_second": 2.16, "eval_weighted avg": { "f1-score": 0.9374292386470396, "precision": 0.9385993125948691, "recall": 0.938049893189471, "support": 27619.0 }, "step": 1458 }, { "epoch": 18.52, "grad_norm": 10.955327033996582, "learning_rate": 1.2592592592592593e-05, "loss": 0.0058, "step": 1500 }, { "epoch": 19.0, "eval_B": { "f1-score": 0.8850241545893719, "precision": 0.8919182083739046, "recall": 0.87823585810163, "support": 1043.0 }, "eval_I": { "f1-score": 0.9626489459211732, "precision": 0.956838628857761, "recall": 0.9685302593659942, "support": 17350.0 }, "eval_O": { "f1-score": 0.927475898334794, "precision": 0.9375415282392027, "recall": 0.9176241057879905, "support": 9226.0 }, "eval_accuracy": 0.9481154277852203, "eval_loss": 0.3719603419303894, "eval_macro avg": { "f1-score": 0.9250496662817796, "precision": 0.9287661218236227, "recall": 0.9214634077518715, "support": 27619.0 }, "eval_runtime": 4.6041, "eval_samples_per_second": 17.376, "eval_steps_per_second": 2.172, "eval_weighted avg": { "f1-score": 0.9479681394332119, "precision": 0.9479408755404257, "recall": 0.9481154277852203, "support": 27619.0 }, "step": 1539 }, { "epoch": 20.0, "eval_B": { "f1-score": 0.8742857142857142, "precision": 0.8684957426679281, "recall": 0.8801534036433365, "support": 1043.0 }, "eval_I": { "f1-score": 0.9602204921293401, "precision": 0.9469289396996189, "recall": 0.9738904899135447, "support": 17350.0 }, "eval_O": { "f1-score": 0.9205305394560855, "precision": 0.9473503097040605, "recall": 0.8951875135486668, "support": 9226.0 }, "eval_accuracy": 0.9440602483797386, "eval_loss": 0.36765894293785095, "eval_macro avg": { "f1-score": 0.9183455819570466, "precision": 0.9209249973572025, "recall": 0.9164104690351826, "support": 27619.0 }, "eval_runtime": 4.6404, "eval_samples_per_second": 17.24, "eval_steps_per_second": 2.155, "eval_weighted avg": { "f1-score": 0.9437170171065533, "precision": 0.9441077562808465, "recall": 0.9440602483797386, "support": 27619.0 }, "step": 1620 }, { "epoch": 21.0, "eval_B": { "f1-score": 0.8885767790262172, "precision": 0.868252516010979, "recall": 0.909875359539789, "support": 1043.0 }, "eval_I": { "f1-score": 0.9616143138880155, "precision": 0.9537903271531439, "recall": 0.9695677233429395, "support": 17350.0 }, "eval_O": { "f1-score": 0.9245376759591498, "precision": 0.9420632242096973, "recall": 0.9076522870149577, "support": 9226.0 }, "eval_accuracy": 0.9466309424671422, "eval_loss": 0.374899685382843, "eval_macro avg": { "f1-score": 0.9249095896244609, "precision": 0.9213686891246068, "recall": 0.9290317899658954, "support": 27619.0 }, "eval_runtime": 4.5979, "eval_samples_per_second": 17.399, "eval_steps_per_second": 2.175, "eval_weighted avg": { "f1-score": 0.9464708542988716, "precision": 0.9466427045463329, "recall": 0.9466309424671422, "support": 27619.0 }, "step": 1701 }, { "epoch": 22.0, "eval_B": { "f1-score": 0.875898418782942, "precision": 0.8754789272030651, "recall": 0.8763183125599233, "support": 1043.0 }, "eval_I": { "f1-score": 0.9606979453806588, "precision": 0.9475306911822412, "recall": 0.9742363112391931, "support": 17350.0 }, "eval_O": { "f1-score": 0.9217236387930074, "precision": 0.94757326007326, "recall": 0.897246910903967, "support": 9226.0 }, "eval_accuracy": 0.9448205945182664, "eval_loss": 0.3899074196815491, "eval_macro avg": { "f1-score": 0.919440000985536, "precision": 0.9235276261528554, "recall": 0.9159338449010278, "support": 27619.0 }, "eval_runtime": 4.5958, "eval_samples_per_second": 17.407, "eval_steps_per_second": 2.176, "eval_weighted avg": { "f1-score": 0.9444764001104068, "precision": 0.9448239585256736, "recall": 0.9448205945182664, "support": 27619.0 }, "step": 1782 }, { "epoch": 23.0, "eval_B": { "f1-score": 0.8766079085278704, "precision": 0.8712121212121212, "recall": 0.8820709491850431, "support": 1043.0 }, "eval_I": { "f1-score": 0.9571775605897973, "precision": 0.9385663638378019, "recall": 0.9765417867435159, "support": 17350.0 }, "eval_O": { "f1-score": 0.9132322264193494, "precision": 0.9515920573375631, "recall": 0.877845220030349, "support": 9226.0 }, "eval_accuracy": 0.9400050689742568, "eval_loss": 0.4506397247314453, "eval_macro avg": { "f1-score": 0.9156725651790056, "precision": 0.9204568474624955, "recall": 0.9121526519863027, "support": 27619.0 }, "eval_runtime": 4.6193, "eval_samples_per_second": 17.319, "eval_steps_per_second": 2.165, "eval_weighted avg": { "f1-score": 0.939455202786939, "precision": 0.9403739808105457, "recall": 0.9400050689742568, "support": 27619.0 }, "step": 1863 }, { "epoch": 24.0, "eval_B": { "f1-score": 0.8786853552440792, "precision": 0.8859649122807017, "recall": 0.8715244487056567, "support": 1043.0 }, "eval_I": { "f1-score": 0.9558590028580501, "precision": 0.9575452599919023, "recall": 0.954178674351585, "support": 17350.0 }, "eval_O": { "f1-score": 0.9160280626011873, "precision": 0.9121883061049011, "recall": 0.9199002818122697, "support": 9226.0 }, "eval_accuracy": 0.9396067924255042, "eval_loss": 0.42392534017562866, "eval_macro avg": { "f1-score": 0.9168574735677723, "precision": 0.9185661594591683, "recall": 0.9152011349565038, "support": 27619.0 }, "eval_runtime": 4.6242, "eval_samples_per_second": 17.3, "eval_steps_per_second": 2.163, "eval_weighted avg": { "f1-score": 0.9396392856607877, "precision": 0.9396908279261412, "recall": 0.9396067924255042, "support": 27619.0 }, "step": 1944 }, { "epoch": 24.69, "grad_norm": 0.030169092118740082, "learning_rate": 1.0123456790123458e-05, "loss": 0.0026, "step": 2000 }, { "epoch": 25.0, "eval_B": { "f1-score": 0.8887832699619772, "precision": 0.88124410933082, "recall": 0.8964525407478428, "support": 1043.0 }, "eval_I": { "f1-score": 0.9607269808508199, "precision": 0.9556860955857192, "recall": 0.9658213256484149, "support": 17350.0 }, "eval_O": { "f1-score": 0.9237260273972602, "precision": 0.9340647163120568, "recall": 0.9136137004118795, "support": 9226.0 }, "eval_accuracy": 0.9457619754516818, "eval_loss": 0.398232638835907, "eval_macro avg": { "f1-score": 0.9244120927366858, "precision": 0.9236649737428652, "recall": 0.9252958556027124, "support": 27619.0 }, "eval_runtime": 4.5953, "eval_samples_per_second": 17.409, "eval_steps_per_second": 2.176, "eval_weighted avg": { "f1-score": 0.9456501103261954, "precision": 0.9456523566073828, "recall": 0.9457619754516818, "support": 27619.0 }, "step": 2025 }, { "epoch": 26.0, "eval_B": { "f1-score": 0.8827717133364974, "precision": 0.8740601503759399, "recall": 0.8916586768935763, "support": 1043.0 }, "eval_I": { "f1-score": 0.9590283123770422, "precision": 0.9489364103142809, "recall": 0.9693371757925072, "support": 17350.0 }, "eval_O": { "f1-score": 0.9188171447557869, "precision": 0.9393115942028986, "recall": 0.8991979189247779, "support": 9226.0 }, "eval_accuracy": 0.9429740396104132, "eval_loss": 0.4066696763038635, "eval_macro avg": { "f1-score": 0.9202057234897755, "precision": 0.9207693849643731, "recall": 0.9200645905369539, "support": 27619.0 }, "eval_runtime": 4.6361, "eval_samples_per_second": 17.256, "eval_steps_per_second": 2.157, "eval_weighted avg": { "f1-score": 0.9427162132687112, "precision": 0.942893668268613, "recall": 0.9429740396104132, "support": 27619.0 }, "step": 2106 }, { "epoch": 27.0, "eval_B": { "f1-score": 0.8731307284129282, "precision": 0.8786407766990292, "recall": 0.8676893576222435, "support": 1043.0 }, "eval_I": { "f1-score": 0.9568967963451661, "precision": 0.9424291543234028, "recall": 0.9718155619596541, "support": 17350.0 }, "eval_O": { "f1-score": 0.9140816781968311, "precision": 0.9418257070590941, "recall": 0.8879254281378712, "support": 9226.0 }, "eval_accuracy": 0.9398602411383468, "eval_loss": 0.44146138429641724, "eval_macro avg": { "f1-score": 0.9147030676516418, "precision": 0.920965212693842, "recall": 0.9091434492399229, "support": 27619.0 }, "eval_runtime": 4.6584, "eval_samples_per_second": 17.173, "eval_steps_per_second": 2.147, "eval_weighted avg": { "f1-score": 0.9394312730137688, "precision": 0.9398186802902107, "recall": 0.9398602411383468, "support": 27619.0 }, "step": 2187 }, { "epoch": 28.0, "eval_B": { "f1-score": 0.8846153846153846, "precision": 0.8871745419479267, "recall": 0.8820709491850431, "support": 1043.0 }, "eval_I": { "f1-score": 0.9619680394066098, "precision": 0.9559995446265938, "recall": 0.9680115273775216, "support": 17350.0 }, "eval_O": { "f1-score": 0.9258771929824562, "precision": 0.9367650321721767, "recall": 0.9152395404292217, "support": 9226.0 }, "eval_accuracy": 0.9471378398928274, "eval_loss": 0.4067588746547699, "eval_macro avg": { "f1-score": 0.9241535390014834, "precision": 0.9266463729155657, "recall": 0.9217740056639289, "support": 27619.0 }, "eval_runtime": 4.6224, "eval_samples_per_second": 17.307, "eval_steps_per_second": 2.163, "eval_weighted avg": { "f1-score": 0.9469909233612611, "precision": 0.9469752465094172, "recall": 0.9471378398928274, "support": 27619.0 }, "step": 2268 }, { "epoch": 29.0, "eval_B": { "f1-score": 0.8919301557338367, "precision": 0.8782527881040892, "recall": 0.9060402684563759, "support": 1043.0 }, "eval_I": { "f1-score": 0.964135443798838, "precision": 0.9576391652925457, "recall": 0.9707204610951009, "support": 17350.0 }, "eval_O": { "f1-score": 0.9300406995930042, "precision": 0.944059848146494, "recall": 0.9164318231086062, "support": 9226.0 }, "eval_accuracy": 0.9501430174879612, "eval_loss": 0.4006032943725586, "eval_macro avg": { "f1-score": 0.9287020997085597, "precision": 0.9266506005143763, "recall": 0.9310641842200277, "support": 27619.0 }, "eval_runtime": 4.5962, "eval_samples_per_second": 17.406, "eval_steps_per_second": 2.176, "eval_weighted avg": { "f1-score": 0.9500195009517102, "precision": 0.9501051209246456, "recall": 0.9501430174879612, "support": 27619.0 }, "step": 2349 }, { "epoch": 30.0, "eval_B": { "f1-score": 0.8796956728483118, "precision": 0.8726415094339622, "recall": 0.8868648130393096, "support": 1043.0 }, "eval_I": { "f1-score": 0.9600954979536152, "precision": 0.947067399349557, "recall": 0.9734870317002882, "support": 17350.0 }, "eval_O": { "f1-score": 0.9212857222438862, "precision": 0.9477363896848138, "recall": 0.8962714068935617, "support": 9226.0 }, "eval_accuracy": 0.9444223179695137, "eval_loss": 0.46608424186706543, "eval_macro avg": { "f1-score": 0.9203589643486044, "precision": 0.922481766156111, "recall": 0.9188744172110531, "support": 27619.0 }, "eval_runtime": 4.5968, "eval_samples_per_second": 17.403, "eval_steps_per_second": 2.175, "eval_weighted avg": { "f1-score": 0.9440950631702127, "precision": 0.9444802637418637, "recall": 0.9444223179695137, "support": 27619.0 }, "step": 2430 }, { "epoch": 30.86, "grad_norm": 0.0009192481520585716, "learning_rate": 7.654320987654322e-06, "loss": 0.0015, "step": 2500 }, { "epoch": 31.0, "eval_B": { "f1-score": 0.8760724499523355, "precision": 0.8710900473933649, "recall": 0.8811121764141898, "support": 1043.0 }, "eval_I": { "f1-score": 0.9596285601041844, "precision": 0.9430224794124193, "recall": 0.9768299711815562, "support": 17350.0 }, "eval_O": { "f1-score": 0.9194073408912335, "precision": 0.953328677839851, "recall": 0.8878170388033817, "support": 9226.0 }, "eval_accuracy": 0.9434809370360984, "eval_loss": 0.48370352387428284, "eval_macro avg": { "f1-score": 0.9183694503159178, "precision": 0.9224804015485452, "recall": 0.9152530621330426, "support": 27619.0 }, "eval_runtime": 4.6119, "eval_samples_per_second": 17.347, "eval_steps_per_second": 2.168, "eval_weighted avg": { "f1-score": 0.943037445605214, "precision": 0.9437487714612122, "recall": 0.9434809370360984, "support": 27619.0 }, "step": 2511 }, { "epoch": 32.0, "eval_B": { "f1-score": 0.8770764119601329, "precision": 0.868421052631579, "recall": 0.8859060402684564, "support": 1043.0 }, "eval_I": { "f1-score": 0.9595270577535243, "precision": 0.9465066726477515, "recall": 0.9729106628242075, "support": 17350.0 }, "eval_O": { "f1-score": 0.919819468434836, "precision": 0.9464510950579063, "recall": 0.8946455668762194, "support": 9226.0 }, "eval_accuracy": 0.9434809370360984, "eval_loss": 0.4701555371284485, "eval_macro avg": { "f1-score": 0.9188076460494976, "precision": 0.9204596067790788, "recall": 0.9178207566562945, "support": 27619.0 }, "eval_runtime": 4.6158, "eval_samples_per_second": 17.332, "eval_steps_per_second": 2.166, "eval_weighted avg": { "f1-score": 0.943149265559139, "precision": 0.9435392929265168, "recall": 0.9434809370360984, "support": 27619.0 }, "step": 2592 }, { "epoch": 33.0, "eval_B": { "f1-score": 0.8849976336961666, "precision": 0.8738317757009346, "recall": 0.8964525407478428, "support": 1043.0 }, "eval_I": { "f1-score": 0.9615693472524951, "precision": 0.9569585569128896, "recall": 0.9662247838616714, "support": 17350.0 }, "eval_O": { "f1-score": 0.9260009859232076, "precision": 0.9359982283246595, "recall": 0.9162150444396271, "support": 9226.0 }, "eval_accuracy": 0.9468843911799848, "eval_loss": 0.4380284249782562, "eval_macro avg": { "f1-score": 0.9241893222906231, "precision": 0.9222628536461612, "recall": 0.9262974563497138, "support": 27619.0 }, "eval_runtime": 4.5881, "eval_samples_per_second": 17.436, "eval_steps_per_second": 2.18, "eval_weighted avg": { "f1-score": 0.9467962563055651, "precision": 0.946817667512148, "recall": 0.9468843911799848, "support": 27619.0 }, "step": 2673 }, { "epoch": 34.0, "eval_B": { "f1-score": 0.8833652007648184, "precision": 0.880838894184938, "recall": 0.8859060402684564, "support": 1043.0 }, "eval_I": { "f1-score": 0.9618386073235572, "precision": 0.9541201156921681, "recall": 0.9696829971181556, "support": 17350.0 }, "eval_O": { "f1-score": 0.9257281286131146, "precision": 0.9406959829920555, "recall": 0.9112291350531108, "support": 9226.0 }, "eval_accuracy": 0.9469930120569173, "eval_loss": 0.4419035315513611, "eval_macro avg": { "f1-score": 0.9236439789004968, "precision": 0.9252183309563873, "recall": 0.9222727241465742, "support": 27619.0 }, "eval_runtime": 4.6449, "eval_samples_per_second": 17.223, "eval_steps_per_second": 2.153, "eval_weighted avg": { "f1-score": 0.9468126092923719, "precision": 0.9468684642086502, "recall": 0.9469930120569173, "support": 27619.0 }, "step": 2754 }, { "epoch": 35.0, "eval_B": { "f1-score": 0.8785405664906385, "precision": 0.8798076923076923, "recall": 0.8772770853307766, "support": 1043.0 }, "eval_I": { "f1-score": 0.9603404255319149, "precision": 0.945586592178771, "recall": 0.9755619596541787, "support": 17350.0 }, "eval_O": { "f1-score": 0.9217537000837756, "precision": 0.9508007835004033, "recall": 0.8944287882072404, "support": 9226.0 }, "eval_accuracy": 0.9447481806003114, "eval_loss": 0.4607029855251312, "eval_macro avg": { "f1-score": 0.920211564035443, "precision": 0.9253983559956221, "recall": 0.9157559443973985, "support": 27619.0 }, "eval_runtime": 4.595, "eval_samples_per_second": 17.41, "eval_steps_per_second": 2.176, "eval_weighted avg": { "f1-score": 0.9443616289800995, "precision": 0.9448443037746955, "recall": 0.9447481806003114, "support": 27619.0 }, "step": 2835 }, { "epoch": 36.0, "eval_B": { "f1-score": 0.8802714493456132, "precision": 0.8901960784313725, "recall": 0.8705656759348035, "support": 1043.0 }, "eval_I": { "f1-score": 0.9600229918091681, "precision": 0.9574089997133849, "recall": 0.9626512968299712, "support": 17350.0 }, "eval_O": { "f1-score": 0.9233949945593034, "precision": 0.9270264365304784, "recall": 0.9197918924777801, "support": 9226.0 }, "eval_accuracy": 0.9448568014772439, "eval_loss": 0.44131579995155334, "eval_macro avg": { "f1-score": 0.9212298119046949, "precision": 0.924877171558412, "recall": 0.9176696217475183, "support": 27619.0 }, "eval_runtime": 4.6044, "eval_samples_per_second": 17.375, "eval_steps_per_second": 2.172, "eval_weighted avg": { "f1-score": 0.944775851745562, "precision": 0.9447216249053674, "recall": 0.9448568014772439, "support": 27619.0 }, "step": 2916 }, { "epoch": 37.0, "eval_B": { "f1-score": 0.8851513695338779, "precision": 0.8872832369942196, "recall": 0.8830297219558965, "support": 1043.0 }, "eval_I": { "f1-score": 0.9621385956860945, "precision": 0.9576338928856915, "recall": 0.966685878962536, "support": 17350.0 }, "eval_O": { "f1-score": 0.9272399278412508, "precision": 0.9353700231609132, "recall": 0.9192499458053327, "support": 9226.0 }, "eval_accuracy": 0.9476809442774902, "eval_loss": 0.4391210675239563, "eval_macro avg": { "f1-score": 0.9248432976870743, "precision": 0.9267623843469414, "recall": 0.9229885155745885, "support": 27619.0 }, "eval_runtime": 4.6195, "eval_samples_per_second": 17.318, "eval_steps_per_second": 2.165, "eval_weighted avg": { "f1-score": 0.9475735214106575, "precision": 0.9475400373450994, "recall": 0.9476809442774902, "support": 27619.0 }, "step": 2997 }, { "epoch": 37.04, "grad_norm": 0.36456331610679626, "learning_rate": 5.185185185185185e-06, "loss": 0.0012, "step": 3000 }, { "epoch": 38.0, "eval_B": { "f1-score": 0.8848715509039011, "precision": 0.8781869688385269, "recall": 0.8916586768935763, "support": 1043.0 }, "eval_I": { "f1-score": 0.9604875396984349, "precision": 0.9536389977842168, "recall": 0.9674351585014409, "support": 17350.0 }, "eval_O": { "f1-score": 0.9237283475391805, "precision": 0.9374930237749749, "recall": 0.9103620203771949, "support": 9226.0 }, "eval_accuracy": 0.9455085267388392, "eval_loss": 0.43638840317726135, "eval_macro avg": { "f1-score": 0.9230291460471722, "precision": 0.9231063301325729, "recall": 0.9231519519240706, "support": 27619.0 }, "eval_runtime": 4.6165, "eval_samples_per_second": 17.329, "eval_steps_per_second": 2.166, "eval_weighted avg": { "f1-score": 0.9453527490407726, "precision": 0.9453961496579408, "recall": 0.9455085267388392, "support": 27619.0 }, "step": 3078 }, { "epoch": 39.0, "eval_B": { "f1-score": 0.8896321070234113, "precision": 0.8866666666666667, "recall": 0.8926174496644296, "support": 1043.0 }, "eval_I": { "f1-score": 0.9625824964131994, "precision": 0.9584571428571429, "recall": 0.9667435158501441, "support": 17350.0 }, "eval_O": { "f1-score": 0.9277944793659471, "precision": 0.9358253390671518, "recall": 0.9199002818122697, "support": 9226.0 }, "eval_accuracy": 0.9482964625801079, "eval_loss": 0.44264644384384155, "eval_macro avg": { "f1-score": 0.9266696942675193, "precision": 0.9269830495303205, "recall": 0.9264204157756145, "support": 27619.0 }, "eval_runtime": 4.6185, "eval_samples_per_second": 17.322, "eval_steps_per_second": 2.165, "eval_weighted avg": { "f1-score": 0.9482068310592221, "precision": 0.9481860074636412, "recall": 0.9482964625801079, "support": 27619.0 }, "step": 3159 }, { "epoch": 40.0, "eval_B": { "f1-score": 0.8865485878410723, "precision": 0.8852772466539197, "recall": 0.887823585810163, "support": 1043.0 }, "eval_I": { "f1-score": 0.9624967964233846, "precision": 0.9512016660100185, "recall": 0.9740634005763689, "support": 17350.0 }, "eval_O": { "f1-score": 0.9265749778172139, "precision": 0.9486713604360664, "recall": 0.905484500325168, "support": 9226.0 }, "eval_accuracy": 0.9478981860313552, "eval_loss": 0.4583560824394226, "eval_macro avg": { "f1-score": 0.925206787360557, "precision": 0.9283834243666682, "recall": 0.9224571622372332, "support": 27619.0 }, "eval_runtime": 4.643, "eval_samples_per_second": 17.23, "eval_steps_per_second": 2.154, "eval_weighted avg": { "f1-score": 0.9476291806512029, "precision": 0.947866868638148, "recall": 0.9478981860313552, "support": 27619.0 }, "step": 3240 }, { "epoch": 41.0, "eval_B": { "f1-score": 0.886558627264061, "precision": 0.8815165876777251, "recall": 0.8916586768935763, "support": 1043.0 }, "eval_I": { "f1-score": 0.9625846436359462, "precision": 0.9584047537424294, "recall": 0.9668011527377521, "support": 17350.0 }, "eval_O": { "f1-score": 0.927930883639545, "precision": 0.9363275215184286, "recall": 0.9196835031432907, "support": 9226.0 }, "eval_accuracy": 0.9482240486621528, "eval_loss": 0.45645976066589355, "eval_macro avg": { "f1-score": 0.9256913848465174, "precision": 0.9254162876461943, "recall": 0.9260477775915397, "support": 27619.0 }, "eval_runtime": 4.6174, "eval_samples_per_second": 17.326, "eval_steps_per_second": 2.166, "eval_weighted avg": { "f1-score": 0.9481376786914271, "precision": 0.9481263619938463, "recall": 0.9482240486621528, "support": 27619.0 }, "step": 3321 }, { "epoch": 42.0, "eval_B": { "f1-score": 0.8838095238095238, "precision": 0.8779564806054873, "recall": 0.8897411313518696, "support": 1043.0 }, "eval_I": { "f1-score": 0.9627247894377418, "precision": 0.9507137237270991, "recall": 0.9750432276657061, "support": 17350.0 }, "eval_O": { "f1-score": 0.9265310659108592, "precision": 0.9507299270072993, "recall": 0.9035334923043572, "support": 9226.0 }, "eval_accuracy": 0.9479343929903328, "eval_loss": 0.4557858109474182, "eval_macro avg": { "f1-score": 0.9243551263860416, "precision": 0.9264667104466285, "recall": 0.9227726171073108, "support": 27619.0 }, "eval_runtime": 4.6048, "eval_samples_per_second": 17.373, "eval_steps_per_second": 2.172, "eval_weighted avg": { "f1-score": 0.947654297555007, "precision": 0.9479715421451187, "recall": 0.9479343929903328, "support": 27619.0 }, "step": 3402 }, { "epoch": 43.0, "eval_B": { "f1-score": 0.8847795163584636, "precision": 0.875234521575985, "recall": 0.8945349952061361, "support": 1043.0 }, "eval_I": { "f1-score": 0.9635252309879175, "precision": 0.9506872370266479, "recall": 0.9767146974063401, "support": 17350.0 }, "eval_O": { "f1-score": 0.9281497159407374, "precision": 0.9546287809349221, "recall": 0.9030999349663993, "support": 9226.0 }, "eval_accuracy": 0.9490206017596582, "eval_loss": 0.46096372604370117, "eval_macro avg": { "f1-score": 0.9254848210957062, "precision": 0.9268501798458516, "recall": 0.9247832091929585, "support": 27619.0 }, "eval_runtime": 4.6289, "eval_samples_per_second": 17.283, "eval_steps_per_second": 2.16, "eval_weighted avg": { "f1-score": 0.9487344607868312, "precision": 0.949154506003899, "recall": 0.9490206017596582, "support": 27619.0 }, "step": 3483 }, { "epoch": 43.21, "grad_norm": 0.008026141673326492, "learning_rate": 2.7160493827160496e-06, "loss": 0.0008, "step": 3500 }, { "epoch": 44.0, "eval_B": { "f1-score": 0.8856053384175404, "precision": 0.8805687203791469, "recall": 0.8906999041227229, "support": 1043.0 }, "eval_I": { "f1-score": 0.9627508804077075, "precision": 0.9565341070717415, "recall": 0.9690489913544669, "support": 17350.0 }, "eval_O": { "f1-score": 0.9280184483610608, "precision": 0.9403582953154557, "recall": 0.9159982657706481, "support": 9226.0 }, "eval_accuracy": 0.9483688764980629, "eval_loss": 0.4500958025455475, "eval_macro avg": { "f1-score": 0.9254582223954363, "precision": 0.9258203742554482, "recall": 0.9252490537492793, "support": 27619.0 }, "eval_runtime": 4.6418, "eval_samples_per_second": 17.235, "eval_steps_per_second": 2.154, "eval_weighted avg": { "f1-score": 0.9482353578197027, "precision": 0.9482619054140469, "recall": 0.9483688764980629, "support": 27619.0 }, "step": 3564 }, { "epoch": 45.0, "eval_B": { "f1-score": 0.8907963757749165, "precision": 0.8861480075901328, "recall": 0.8954937679769894, "support": 1043.0 }, "eval_I": { "f1-score": 0.9618886240271116, "precision": 0.9586123991069895, "recall": 0.9651873198847263, "support": 17350.0 }, "eval_O": { "f1-score": 0.9268638794891386, "precision": 0.9334872471416007, "recall": 0.9203338391502276, "support": 9226.0 }, "eval_accuracy": 0.9475723234005576, "eval_loss": 0.4541914463043213, "eval_macro avg": { "f1-score": 0.9265162930970557, "precision": 0.9260825512795744, "recall": 0.9270049756706479, "support": 27619.0 }, "eval_runtime": 4.6, "eval_samples_per_second": 17.391, "eval_steps_per_second": 2.174, "eval_weighted avg": { "f1-score": 0.9475040515214316, "precision": 0.9474829225732714, "recall": 0.9475723234005576, "support": 27619.0 }, "step": 3645 }, { "epoch": 46.0, "eval_B": { "f1-score": 0.8897338403041826, "precision": 0.882186616399623, "recall": 0.8974113135186961, "support": 1043.0 }, "eval_I": { "f1-score": 0.9623334864492421, "precision": 0.9587003775311749, "recall": 0.9659942363112392, "support": 17350.0 }, "eval_O": { "f1-score": 0.9276581794339416, "precision": 0.9353239312472454, "recall": 0.9201170604812486, "support": 9226.0 }, "eval_accuracy": 0.9480792208262429, "eval_loss": 0.45604071021080017, "eval_macro avg": { "f1-score": 0.926575168729122, "precision": 0.9254036417260144, "recall": 0.927840870103728, "support": 27619.0 }, "eval_runtime": 4.6335, "eval_samples_per_second": 17.266, "eval_steps_per_second": 2.158, "eval_weighted avg": { "f1-score": 0.9480087167815329, "precision": 0.9480021282724854, "recall": 0.9480792208262429, "support": 27619.0 }, "step": 3726 } ], "logging_steps": 500, "max_steps": 4050, "num_input_tokens_seen": 0, "num_train_epochs": 50, "save_steps": 500, "total_flos": 6614186025837600.0, "train_batch_size": 4, "trial_name": null, "trial_params": null }