diff --git "a/checkpoint-3969/trainer_state.json" "b/checkpoint-3969/trainer_state.json" --- "a/checkpoint-3969/trainer_state.json" +++ "b/checkpoint-3969/trainer_state.json" @@ -11,1960 +11,1960 @@ { "epoch": 1.0, "eval_B": { - "f1-score": 0.7905866302864938, - "precision": 0.7768096514745308, - "recall": 0.8048611111111111, - "support": 1440.0 + "f1-score": 0.8331822302810515, + "precision": 0.7901977644024075, + "recall": 0.8811121764141898, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9005331227169513, - "precision": 0.9637593111099372, - "recall": 0.8450919534905267, - "support": 21587.0 + "f1-score": 0.9471294753175105, + "precision": 0.9443616777446711, + "recall": 0.9499135446685879, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.8288892663043479, - "precision": 0.7463108800367001, - "recall": 0.9320156593144275, - "support": 10473.0 + "f1-score": 0.8992868897421833, + "precision": 0.9103731674811195, + "recall": 0.8884673748103187, + "support": 9226.0 }, - "eval_accuracy": 0.8705373134328358, - "eval_loss": 0.32930415868759155, + "eval_accuracy": 0.9267895289474637, + "eval_loss": 0.20478281378746033, "eval_macro avg": { - "f1-score": 0.8400030064359308, - "precision": 0.8289599475403894, - "recall": 0.8606562413053552, - "support": 33500.0 - }, - "eval_runtime": 4.7349, - "eval_samples_per_second": 17.107, - "eval_steps_per_second": 2.323, + "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": 0.8734092702599646, - "precision": 0.8877430445874, - "recall": 0.8705373134328358, - "support": 33500.0 + "f1-score": 0.9268447919078652, + "precision": 0.9271861479533134, + "recall": 0.9267895289474637, + "support": 27619.0 }, "step": 81 }, { "epoch": 2.0, "eval_B": { - "f1-score": 0.8656218739579861, - "precision": 0.8325849903784477, - "recall": 0.9013888888888889, - "support": 1440.0 + "f1-score": 0.8682547399124939, + "precision": 0.8806706114398422, + "recall": 0.8561840843720039, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9547168948727912, - "precision": 0.9423736462093862, - "recall": 0.9673877796822161, - "support": 21587.0 + "f1-score": 0.9590983771942466, + "precision": 0.9592919333448654, + "recall": 0.9589048991354466, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9013528191962081, - "precision": 0.9332379102341274, - "recall": 0.8715745249689678, - "support": 10473.0 + "f1-score": 0.9226525313717006, + "precision": 0.9208594256100194, + "recall": 0.9244526338608281, + "support": 9226.0 }, - "eval_accuracy": 0.9345970149253732, - "eval_loss": 0.2162298709154129, + "eval_accuracy": 0.9435171439950758, + "eval_loss": 0.16574698686599731, "eval_macro avg": { - "f1-score": 0.9072305293423284, - "precision": 0.902732182273987, - "recall": 0.9134503978466909, - "support": 33500.0 - }, - "eval_runtime": 4.753, - "eval_samples_per_second": 17.042, - "eval_steps_per_second": 2.314, + "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": 0.9342040950316517, - "precision": 0.9347982961417612, - "recall": 0.9345970149253732, - "support": 33500.0 + "f1-score": 0.9434932036816766, + "precision": 0.9434846863370583, + "recall": 0.9435171439950758, + "support": 27619.0 }, "step": 162 }, { "epoch": 3.0, "eval_B": { - "f1-score": 0.8603587759409076, - "precision": 0.8717034925160371, - "recall": 0.8493055555555555, - "support": 1440.0 + "f1-score": 0.8791725434884815, + "precision": 0.8625461254612546, + "recall": 0.8964525407478428, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9434195218082959, - "precision": 0.9677057963955188, - "recall": 0.9203224162690509, - "support": 21587.0 + "f1-score": 0.9588372159825352, + "precision": 0.9557349825345015, + "recall": 0.9619596541786744, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.8879310344827586, - "precision": 0.8459410391631365, - "recall": 0.9343072663038289, - "support": 10473.0 + "f1-score": 0.9204284621270084, + "precision": 0.9282407407407407, + "recall": 0.9127465857359636, + "support": 9226.0 }, - "eval_accuracy": 0.9216417910447762, - "eval_loss": 0.24082203209400177, + "eval_accuracy": 0.9430464535283681, + "eval_loss": 0.15394224226474762, "eval_macro avg": { - "f1-score": 0.8972364440773206, - "precision": 0.8951167760248975, - "recall": 0.9013117460428117, - "support": 33500.0 - }, - "eval_runtime": 4.7554, - "eval_samples_per_second": 17.033, - "eval_steps_per_second": 2.313, + "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": 0.9225019575751797, - "precision": 0.9255121957960801, - "recall": 0.9216417910447762, - "support": 33500.0 + "f1-score": 0.9429985029052194, + "precision": 0.9430314866542512, + "recall": 0.9430464535283681, + "support": 27619.0 }, "step": 243 }, { "epoch": 4.0, "eval_B": { - "f1-score": 0.8804020100502512, - "precision": 0.8504854368932039, - "recall": 0.9125, - "support": 1440.0 + "f1-score": 0.8800746616892207, + "precision": 0.8572727272727273, + "recall": 0.9041227229146692, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9589184992350025, - "precision": 0.945595388218339, - "recall": 0.9726224116366332, - "support": 21587.0 + "f1-score": 0.9537844036697248, + "precision": 0.9488876212207644, + "recall": 0.9587319884726225, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9091178797468354, - "precision": 0.9427750999897446, - "recall": 0.877780960565263, - "support": 10473.0 + "f1-score": 0.9100192149327477, + "precision": 0.9220157970853265, + "recall": 0.8983308042488619, + "support": 9226.0 }, - "eval_accuracy": 0.9403880597014925, - "eval_loss": 0.22083505988121033, + "eval_accuracy": 0.9364929939534379, + "eval_loss": 0.19801080226898193, "eval_macro avg": { - "f1-score": 0.9161461296773631, - "precision": 0.9129519750337626, - "recall": 0.9209677907339654, - "support": 33500.0 - }, - "eval_runtime": 4.7449, - "eval_samples_per_second": 17.071, - "eval_steps_per_second": 2.318, + "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": 0.9399744505088647, - "precision": 0.9406253819936745, - "recall": 0.9403880597014925, - "support": 33500.0 + "f1-score": 0.9363812792925563, + "precision": 0.9364514800186444, + "recall": 0.9364929939534379, + "support": 27619.0 }, "step": 324 }, { "epoch": 5.0, "eval_B": { - "f1-score": 0.8879781420765027, - "precision": 0.8736559139784946, - "recall": 0.9027777777777778, - "support": 1440.0 + "f1-score": 0.883877159309021, + "precision": 0.8847262247838616, + "recall": 0.8830297219558965, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9613201259451634, - "precision": 0.9539317642765919, - "recall": 0.9688238291564367, - "support": 21587.0 + "f1-score": 0.9591500482927107, + "precision": 0.9456643513331839, + "recall": 0.9730259365994236, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.916006030835076, - "precision": 0.9334853291038858, - "recall": 0.899169292466342, - "support": 10473.0 + "f1-score": 0.9191176470588236, + "precision": 0.9454503781801513, + "recall": 0.8942120095382614, + "support": 9226.0 }, - "eval_accuracy": 0.9442089552238806, - "eval_loss": 0.2257838249206543, + "eval_accuracy": 0.9432999022412107, + "eval_loss": 0.20074255764484406, "eval_macro avg": { - "f1-score": 0.921768099618914, - "precision": 0.9203576691196576, - "recall": 0.9235902998001855, - "support": 33500.0 - }, - "eval_runtime": 4.7845, - "eval_samples_per_second": 16.93, - "eval_steps_per_second": 2.299, + "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": 0.9440011117702735, - "precision": 0.9440889959185916, - "recall": 0.9442089552238806, - "support": 33500.0 + "f1-score": 0.9429348139614955, + "precision": 0.9432916158141273, + "recall": 0.9432999022412107, + "support": 27619.0 }, "step": 405 }, { "epoch": 6.0, "eval_B": { - "f1-score": 0.8873579056148813, - "precision": 0.8803827751196173, - "recall": 0.8944444444444445, - "support": 1440.0 + "f1-score": 0.8897338403041826, + "precision": 0.882186616399623, + "recall": 0.8974113135186961, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9622060984570168, - "precision": 0.9539267015706806, - "recall": 0.9706304720433594, - "support": 21587.0 + "f1-score": 0.9624160683661308, + "precision": 0.9707969272268809, + "recall": 0.954178674351585, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9182769530299342, - "precision": 0.9365567911040509, - "recall": 0.9006970304592762, - "support": 10473.0 + "f1-score": 0.9297955261331482, + "precision": 0.9161493950552341, + "recall": 0.9438543247344461, + "support": 9226.0 }, - "eval_accuracy": 0.9454925373134329, - "eval_loss": 0.2242102175951004, + "eval_accuracy": 0.9485861182519281, + "eval_loss": 0.16565027832984924, "eval_macro avg": { - "f1-score": 0.9226136523672773, - "precision": 0.923622089264783, - "recall": 0.9219239823156933, - "support": 33500.0 - }, - "eval_runtime": 4.7715, - "eval_samples_per_second": 16.976, - "eval_steps_per_second": 2.305, + "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": 0.945255312255509, - "precision": 0.9453351097376493, - "recall": 0.9454925373134329, - "support": 33500.0 + "f1-score": 0.9487745648174827, + "precision": 0.9491959030765336, + "recall": 0.9485861182519281, + "support": 27619.0 }, "step": 486 }, { "epoch": 6.17, - "grad_norm": 1.8705081939697266, + "grad_norm": 4.767655849456787, "learning_rate": 1.7530864197530865e-05, - "loss": 0.1601, + "loss": 0.1585, "step": 500 }, { "epoch": 7.0, "eval_B": { - "f1-score": 0.8876749743939911, - "precision": 0.8730691739422431, - "recall": 0.9027777777777778, - "support": 1440.0 + "f1-score": 0.8802588996763754, + "precision": 0.85, + "recall": 0.912751677852349, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9611415113943341, - "precision": 0.9581108451482232, - "recall": 0.9641914114976606, - "support": 21587.0 + "f1-score": 0.958423523458803, + "precision": 0.9354661691269275, + "recall": 0.9825360230547551, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9164739884393064, - "precision": 0.9247594050743657, - "recall": 0.9083357204239473, - "support": 10473.0 + "f1-score": 0.9139526911210146, + "precision": 0.9664088931851136, + "recall": 0.8668978972469109, + "support": 9226.0 }, - "eval_accuracy": 0.944089552238806, - "eval_loss": 0.22517083585262299, + "eval_accuracy": 0.9412723125384699, + "eval_loss": 0.2591131329536438, "eval_macro avg": { - "f1-score": 0.9217634914092105, - "precision": 0.9186464747216107, - "recall": 0.9251016365664619, - "support": 33500.0 - }, - "eval_runtime": 4.7926, - "eval_samples_per_second": 16.901, - "eval_steps_per_second": 2.295, + "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": 0.9440192791200506, - "precision": 0.9440287663891153, - "recall": 0.944089552238806, - "support": 33500.0 + "f1-score": 0.9406164485555297, + "precision": 0.9425749115781907, + "recall": 0.9412723125384699, + "support": 27619.0 }, "step": 567 }, { "epoch": 8.0, "eval_B": { - "f1-score": 0.894540522210919, - "precision": 0.8740888005301524, - "recall": 0.9159722222222222, - "support": 1440.0 + "f1-score": 0.8687910028116214, + "precision": 0.84967919340055, + "recall": 0.8887823585810163, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9638018385639943, - "precision": 0.9540255968049378, - "recall": 0.9737805160513272, - "support": 21587.0 + "f1-score": 0.9595010203788336, + "precision": 0.956997878562009, + "recall": 0.9620172910662824, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9203132648066569, - "precision": 0.9441598875163202, - "recall": 0.8976415544734078, - "support": 10473.0 + "f1-score": 0.9212035166275324, + "precision": 0.9282491471332673, + "recall": 0.9142640364188164, + "support": 9226.0 }, - "eval_accuracy": 0.9474925373134329, - "eval_loss": 0.24869142472743988, + "eval_accuracy": 0.9432999022412107, + "eval_loss": 0.2693409323692322, "eval_macro avg": { - "f1-score": 0.9262185418605234, - "precision": 0.9240914282838034, - "recall": 0.9291314309156524, - "support": 33500.0 - }, - "eval_runtime": 4.7743, - "eval_samples_per_second": 16.966, - "eval_steps_per_second": 2.304, + "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": 0.947228939205516, - "precision": 0.9475052218791054, - "recall": 0.9474925373134329, - "support": 33500.0 + "f1-score": 0.9432823550422136, + "precision": 0.9433417293609165, + "recall": 0.9432999022412107, + "support": 27619.0 }, "step": 648 }, { "epoch": 9.0, "eval_B": { - "f1-score": 0.8851236868858015, - "precision": 0.8643282594308405, - "recall": 0.9069444444444444, - "support": 1440.0 + "f1-score": 0.8645343367826904, + "precision": 0.8485687903970452, + "recall": 0.8811121764141898, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9617001723147618, - "precision": 0.9540067462849849, - "recall": 0.9695186918052532, - "support": 21587.0 + "f1-score": 0.9567003348057282, + "precision": 0.934535260814599, + "recall": 0.9799423631123919, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9165854609237966, - "precision": 0.9358272808675754, - "recall": 0.8981189725961998, - "support": 10473.0 + "f1-score": 0.9104673003585861, + "precision": 0.9586479683567062, + "recall": 0.8668978972469109, + "support": 9226.0 }, - "eval_accuracy": 0.9445074626865672, - "eval_loss": 0.3045330345630646, + "eval_accuracy": 0.9384481697382236, + "eval_loss": 0.3087053894996643, "eval_macro avg": { - "f1-score": 0.9211364400414532, - "precision": 0.9180540955278002, - "recall": 0.9248607029486324, - "support": 33500.0 - }, - "eval_runtime": 4.7599, - "eval_samples_per_second": 17.017, - "eval_steps_per_second": 2.311, + "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": 0.9443044555560967, - "precision": 0.9444685205421044, - "recall": 0.9445074626865672, - "support": 33500.0 + "f1-score": 0.9377758584761231, + "precision": 0.9393435743356523, + "recall": 0.9384481697382236, + "support": 27619.0 }, "step": 729 }, { "epoch": 10.0, "eval_B": { - "f1-score": 0.8815068493150685, - "precision": 0.8695945945945946, - "recall": 0.89375, - "support": 1440.0 + "f1-score": 0.8803126526624329, + "precision": 0.897410358565737, + "recall": 0.8638542665388304, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.956297959947595, - "precision": 0.9660159757999717, - "recall": 0.9467735211006625, - "support": 21587.0 + "f1-score": 0.9590983771942466, + "precision": 0.9592919333448654, + "recall": 0.9589048991354466, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9101049868766404, - "precision": 0.8937678357728068, - "recall": 0.9270505108373914, - "support": 10473.0 + "f1-score": 0.9220456265542221, + "precision": 0.9197584124245038, + "recall": 0.9243442445263386, + "support": 9226.0 }, - "eval_accuracy": 0.9383283582089552, - "eval_loss": 0.32627996802330017, + "eval_accuracy": 0.9437705927079184, + "eval_loss": 0.2796257436275482, "eval_macro avg": { - "f1-score": 0.9159699320464346, - "precision": 0.909792802055791, - "recall": 0.9225246773126846, - "support": 33500.0 - }, - "eval_runtime": 4.7534, - "eval_samples_per_second": 17.04, - "eval_steps_per_second": 2.314, + "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": 0.9386418940884026, - "precision": 0.9392845859659347, - "recall": 0.9383283582089552, - "support": 33500.0 + "f1-score": 0.9437458232244597, + "precision": 0.9437490553802075, + "recall": 0.9437705927079184, + "support": 27619.0 }, "step": 810 }, { "epoch": 11.0, "eval_B": { - "f1-score": 0.8901136755080951, - "precision": 0.8831168831168831, - "recall": 0.8972222222222223, - "support": 1440.0 + "f1-score": 0.8830917874396135, + "precision": 0.8899707887049659, + "recall": 0.8763183125599233, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620073001078947, - "precision": 0.9535359970874671, - "recall": 0.9706304720433594, - "support": 21587.0 + "f1-score": 0.9574634882843123, + "precision": 0.9647182727751448, + "recall": 0.9503170028818444, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.917413322945072, - "precision": 0.9361025539103647, - "recall": 0.8994557433400172, - "support": 10473.0 + "f1-score": 0.9200619426496501, + "precision": 0.9067466582465004, + "recall": 0.9337741166269239, + "support": 9226.0 }, - "eval_accuracy": 0.9452238805970149, - "eval_loss": 0.3123624622821808, + "eval_accuracy": 0.9419964517180202, + "eval_loss": 0.3198166787624359, "eval_macro avg": { - "f1-score": 0.9231780995203539, - "precision": 0.9242518113715716, - "recall": 0.922436145868533, - "support": 33500.0 - }, - "eval_runtime": 4.7539, - "eval_samples_per_second": 17.039, - "eval_steps_per_second": 2.314, + "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": 0.9449756719810304, - "precision": 0.9450588635199616, - "recall": 0.9452238805970149, - "support": 33500.0 + "f1-score": 0.942161111514465, + "precision": 0.9425303680165918, + "recall": 0.9419964517180202, + "support": 27619.0 }, "step": 891 }, { "epoch": 12.0, "eval_B": { - "f1-score": 0.8800819952169457, - "precision": 0.8661735036987223, - "recall": 0.8944444444444445, - "support": 1440.0 + "f1-score": 0.871306005719733, + "precision": 0.8663507109004739, + "recall": 0.8763183125599233, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9595463311598861, - "precision": 0.9512039692293687, - "recall": 0.9680363181544448, - "support": 21587.0 + "f1-score": 0.9576855173984852, + "precision": 0.9464205312922107, + "recall": 0.9692219020172911, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9115367743822197, - "precision": 0.9310035842293907, - "recall": 0.8928673732454884, - "support": 10473.0 + "f1-score": 0.9162135168127844, + "precision": 0.9386084583901774, + "recall": 0.8948623455451984, + "support": 9226.0 }, - "eval_accuracy": 0.9413731343283582, - "eval_loss": 0.3427491784095764, + "eval_accuracy": 0.9408740359897172, + "eval_loss": 0.35264310240745544, "eval_macro avg": { - "f1-score": 0.9170550335863505, - "precision": 0.9161270190524938, - "recall": 0.9184493786147926, - "support": 33500.0 - }, - "eval_runtime": 4.7455, - "eval_samples_per_second": 17.069, - "eval_steps_per_second": 2.318, + "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": 0.9411214734915179, - "precision": 0.9412337452750731, - "recall": 0.9413731343283582, - "support": 33500.0 + "f1-score": 0.9405699625961891, + "precision": 0.9407871989028143, + "recall": 0.9408740359897172, + "support": 27619.0 }, "step": 972 }, { "epoch": 12.35, - "grad_norm": 0.47551417350769043, + "grad_norm": 0.3602512776851654, "learning_rate": 1.506172839506173e-05, - "loss": 0.0222, + "loss": 0.021, "step": 1000 }, { "epoch": 13.0, "eval_B": { - "f1-score": 0.8896434634974534, - "precision": 0.8704318936877077, - "recall": 0.9097222222222222, - "support": 1440.0 + "f1-score": 0.877092300334768, + "precision": 0.875, + "recall": 0.8791946308724832, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620136667970457, - "precision": 0.9556591698665204, - "recall": 0.9684532357437347, - "support": 21587.0 + "f1-score": 0.9597242793665263, + "precision": 0.9486991778353419, + "recall": 0.9710086455331413, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9183177933177933, - "precision": 0.9343808676746714, - "recall": 0.9027976701995608, - "support": 10473.0 + "f1-score": 0.9203392649259936, + "precision": 0.9419040054464994, + "recall": 0.8997398655972252, + "support": 9226.0 }, - "eval_accuracy": 0.9454029850746268, - "eval_loss": 0.3816670775413513, + "eval_accuracy": 0.943734385748941, + "eval_loss": 0.3594338297843933, "eval_macro avg": { - "f1-score": 0.9233249745374308, - "precision": 0.9201573104096331, - "recall": 0.9269910427218392, - "support": 33500.0 - }, - "eval_runtime": 4.7247, - "eval_samples_per_second": 17.144, - "eval_steps_per_second": 2.328, + "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": 0.945242324238848, - "precision": 0.9453435001186778, - "recall": 0.9454029850746268, - "support": 33500.0 + "f1-score": 0.943447393984779, + "precision": 0.9436461164304495, + "recall": 0.943734385748941, + "support": 27619.0 }, "step": 1053 }, { "epoch": 14.0, "eval_B": { - "f1-score": 0.8865272540281112, - "precision": 0.8754231550440081, - "recall": 0.8979166666666667, - "support": 1440.0 + "f1-score": 0.8754052802223252, + "precision": 0.8467741935483871, + "recall": 0.9060402684563759, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9599206715247671, - "precision": 0.9557331129172981, - "recall": 0.9641450873210728, - "support": 21587.0 + "f1-score": 0.9574973153224439, + "precision": 0.9392880904856953, + "recall": 0.9764265129682997, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9141367826632559, - "precision": 0.9242631270739801, - "recall": 0.9042299245679366, - "support": 10473.0 + "f1-score": 0.9130164471825015, + "precision": 0.9539388213062477, + "recall": 0.8754606546715803, + "support": 9226.0 }, - "eval_accuracy": 0.9425671641791045, - "eval_loss": 0.3827010989189148, + "eval_accuracy": 0.9400412759332344, + "eval_loss": 0.3840242028236389, "eval_macro avg": { - "f1-score": 0.9201949027387114, - "precision": 0.9184731316784287, - "recall": 0.9220972261852255, - "support": 33500.0 - }, - "eval_runtime": 4.7319, - "eval_samples_per_second": 17.118, - "eval_steps_per_second": 2.325, + "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": 0.9424525763235196, - "precision": 0.9424426203479068, - "recall": 0.9425671641791045, - "support": 33500.0 + "f1-score": 0.9395385738014427, + "precision": 0.9406884180878825, + "recall": 0.9400412759332344, + "support": 27619.0 }, "step": 1134 }, { "epoch": 15.0, "eval_B": { - "f1-score": 0.8871919472405415, - "precision": 0.8868841082581541, - "recall": 0.8875, - "support": 1440.0 + "f1-score": 0.8797709923664121, + "precision": 0.8755935422602089, + "recall": 0.8839884947267498, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9616856471022177, - "precision": 0.9571841585975862, - "recall": 0.9662296752675221, - "support": 21587.0 + "f1-score": 0.9603688219460511, + "precision": 0.954302299112224, + "recall": 0.9665129682997118, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9181813798756087, - "precision": 0.9273470977795092, - "recall": 0.9091950730449728, - "support": 10473.0 + "f1-score": 0.922722283205269, + "precision": 0.9346230820547031, + "recall": 0.9111207457186213, + "support": 9226.0 }, - "eval_accuracy": 0.9450149253731344, - "eval_loss": 0.39016807079315186, + "eval_accuracy": 0.9448930084362215, + "eval_loss": 0.3762665390968323, "eval_macro avg": { - "f1-score": 0.9223529914061226, - "precision": 0.9238051215450831, - "recall": 0.920974916104165, - "support": 33500.0 - }, - "eval_runtime": 4.7376, - "eval_samples_per_second": 17.097, - "eval_steps_per_second": 2.322, + "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": 0.944882927148036, - "precision": 0.9448344388830637, - "recall": 0.9450149253731344, - "support": 33500.0 + "f1-score": 0.944749483712443, + "precision": 0.9447562007752335, + "recall": 0.9448930084362215, + "support": 27619.0 }, "step": 1215 }, { "epoch": 16.0, "eval_B": { - "f1-score": 0.8897338403041825, - "precision": 0.8857536132140399, - "recall": 0.89375, - "support": 1440.0 + "f1-score": 0.8858669256920835, + "precision": 0.8976377952755905, + "recall": 0.8744007670182167, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9608010520729992, - "precision": 0.9570693142121713, - "recall": 0.9645620049103627, - "support": 21587.0 + "f1-score": 0.9586212929752427, + "precision": 0.9701351590627397, + "recall": 0.9473775216138328, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9159121556540166, - "precision": 0.9240112719852298, - "recall": 0.9079537859257137, - "support": 10473.0 + "f1-score": 0.9235412474849094, + "precision": 0.9027950310559006, + "recall": 0.9452633860828095, + "support": 9226.0 }, - "eval_accuracy": 0.943820895522388, - "eval_loss": 0.40087994933128357, + "eval_accuracy": 0.9439154205438285, + "eval_loss": 0.36898088455200195, "eval_macro avg": { - "f1-score": 0.9221490160103993, - "precision": 0.9222780664704805, - "recall": 0.9220885969453588, - "support": 33500.0 - }, - "eval_runtime": 4.8145, - "eval_samples_per_second": 16.824, - "eval_steps_per_second": 2.285, + "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": 0.9437127476806678, - "precision": 0.9436689713560499, - "recall": 0.943820895522388, - "support": 33500.0 + "f1-score": 0.9441554794131966, + "precision": 0.9449027186622512, + "recall": 0.9439154205438285, + "support": 27619.0 }, "step": 1296 }, { "epoch": 17.0, "eval_B": { - "f1-score": 0.8904347826086957, - "precision": 0.89198606271777, - "recall": 0.8888888888888888, - "support": 1440.0 + "f1-score": 0.8863309352517986, + "precision": 0.8867562380038387, + "recall": 0.8859060402684564, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9615171394252979, - "precision": 0.9589034784611841, - "recall": 0.9641450873210728, - "support": 21587.0 + "f1-score": 0.9630614486386496, + "precision": 0.9659090909090909, + "recall": 0.9602305475504322, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9182546920750733, - "precision": 0.9232625482625483, - "recall": 0.9133008689009835, - "support": 10473.0 + "f1-score": 0.9296685529506872, + "precision": 0.9245363918962375, + "recall": 0.9348580099718188, + "support": 9226.0 }, - "eval_accuracy": 0.9450149253731344, - "eval_loss": 0.42787015438079834, + "eval_accuracy": 0.9489481878417032, + "eval_loss": 0.3569599986076355, "eval_macro avg": { - "f1-score": 0.9234022047030224, - "precision": 0.9247173631471673, - "recall": 0.9221116150369818, - "support": 33500.0 - }, - "eval_runtime": 4.7453, - "eval_samples_per_second": 17.07, - "eval_steps_per_second": 2.318, + "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": 0.9449366556964975, - "precision": 0.9448847160539952, - "recall": 0.9450149253731344, - "support": 33500.0 + "f1-score": 0.9490090650954502, + "precision": 0.9490996138580476, + "recall": 0.9489481878417032, + "support": 27619.0 }, "step": 1377 }, { "epoch": 18.0, "eval_B": { - "f1-score": 0.8895027624309393, - "precision": 0.8846153846153846, - "recall": 0.8944444444444445, - "support": 1440.0 + "f1-score": 0.8778409090909091, + "precision": 0.8671655753040225, + "recall": 0.8887823585810163, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9607206210979795, - "precision": 0.9555912007332722, - "recall": 0.9659054060314078, - "support": 21587.0 + "f1-score": 0.9557028055829692, + "precision": 0.9355230471984544, + "recall": 0.9767723342939482, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9159781610861478, - "precision": 0.9271322378716745, - "recall": 0.905089277188962, - "support": 10473.0 + "f1-score": 0.9098012570069645, + "precision": 0.9524599881446354, + "recall": 0.8707999132885325, + "support": 9226.0 }, - "eval_accuracy": 0.943820895522388, - "eval_loss": 0.422620564699173, + "eval_accuracy": 0.938049893189471, + "eval_loss": 0.4151785373687744, "eval_macro avg": { - "f1-score": 0.9220671815383555, - "precision": 0.9224462744067772, - "recall": 0.9218130425549381, - "support": 33500.0 - }, - "eval_runtime": 4.7703, - "eval_samples_per_second": 16.98, - "eval_steps_per_second": 2.306, + "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": 0.9436716210924735, - "precision": 0.9436432636210552, - "recall": 0.943820895522388, - "support": 33500.0 + "f1-score": 0.9374292386470396, + "precision": 0.9385993125948691, + "recall": 0.938049893189471, + "support": 27619.0 }, "step": 1458 }, { "epoch": 18.52, - "grad_norm": 0.1027584820985794, + "grad_norm": 10.955327033996582, "learning_rate": 1.2592592592592593e-05, - "loss": 0.0054, + "loss": 0.0058, "step": 1500 }, { "epoch": 19.0, "eval_B": { - "f1-score": 0.8924247665167762, - "precision": 0.8890420399724328, - "recall": 0.8958333333333334, - "support": 1440.0 + "f1-score": 0.8850241545893719, + "precision": 0.8919182083739046, + "recall": 0.87823585810163, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9599648245857633, - "precision": 0.959121387283237, - "recall": 0.9608097466067541, - "support": 21587.0 + "f1-score": 0.9626489459211732, + "precision": 0.956838628857761, + "recall": 0.9685302593659942, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9158252380724505, - "precision": 0.9179777436684574, - "recall": 0.913682803399217, - "support": 10473.0 + "f1-score": 0.927475898334794, + "precision": 0.9375415282392027, + "recall": 0.9176241057879905, + "support": 9226.0 }, - "eval_accuracy": 0.9432835820895522, - "eval_loss": 0.4375981390476227, + "eval_accuracy": 0.9481154277852203, + "eval_loss": 0.3719603419303894, "eval_macro avg": { - "f1-score": 0.9227382763916633, - "precision": 0.9220470569747091, - "recall": 0.9234419611131015, - "support": 33500.0 - }, - "eval_runtime": 4.7487, - "eval_samples_per_second": 17.057, - "eval_steps_per_second": 2.316, + "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": 0.9432623895656658, - "precision": 0.9432464129636805, - "recall": 0.9432835820895522, - "support": 33500.0 + "f1-score": 0.9479681394332119, + "precision": 0.9479408755404257, + "recall": 0.9481154277852203, + "support": 27619.0 }, "step": 1539 }, { "epoch": 20.0, "eval_B": { - "f1-score": 0.8907273354015857, - "precision": 0.8843258042436687, - "recall": 0.8972222222222223, - "support": 1440.0 + "f1-score": 0.8742857142857142, + "precision": 0.8684957426679281, + "recall": 0.8801534036433365, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620682542229149, - "precision": 0.9520958083832335, - "recall": 0.972251818223931, - "support": 21587.0 + "f1-score": 0.9602204921293401, + "precision": 0.9469289396996189, + "recall": 0.9738904899135447, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9177252296267345, - "precision": 0.9396698349174587, - "recall": 0.8967822018523823, - "support": 10473.0 + "f1-score": 0.9205305394560855, + "precision": 0.9473503097040605, + "recall": 0.8951875135486668, + "support": 9226.0 }, - "eval_accuracy": 0.9454328358208955, - "eval_loss": 0.44444534182548523, + "eval_accuracy": 0.9440602483797386, + "eval_loss": 0.36765894293785095, "eval_macro avg": { - "f1-score": 0.9235069397504118, - "precision": 0.9253638158481202, - "recall": 0.922085414099512, - "support": 33500.0 - }, - "eval_runtime": 4.7631, - "eval_samples_per_second": 17.006, - "eval_steps_per_second": 2.309, + "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": 0.9451388387095264, - "precision": 0.9452980165603072, - "recall": 0.9454328358208955, - "support": 33500.0 + "f1-score": 0.9437170171065533, + "precision": 0.9441077562808465, + "recall": 0.9440602483797386, + "support": 27619.0 }, "step": 1620 }, { "epoch": 21.0, "eval_B": { - "f1-score": 0.890790825059911, - "precision": 0.8784604996623903, - "recall": 0.9034722222222222, - "support": 1440.0 + "f1-score": 0.8885767790262172, + "precision": 0.868252516010979, + "recall": 0.909875359539789, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9615375783600082, - "precision": 0.9533284764593388, - "recall": 0.9698892852179553, - "support": 21587.0 + "f1-score": 0.9616143138880155, + "precision": 0.9537903271531439, + "recall": 0.9695677233429395, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9162201656113005, - "precision": 0.9351695336581486, - "recall": 0.8980234889716414, - "support": 10473.0 + "f1-score": 0.9245376759591498, + "precision": 0.9420632242096973, + "recall": 0.9076522870149577, + "support": 9226.0 }, - "eval_accuracy": 0.9445671641791045, - "eval_loss": 0.4363422691822052, + "eval_accuracy": 0.9466309424671422, + "eval_loss": 0.374899685382843, "eval_macro avg": { - "f1-score": 0.9228495230104065, - "precision": 0.9223195032599593, - "recall": 0.9237949988039397, - "support": 33500.0 - }, - "eval_runtime": 4.7902, - "eval_samples_per_second": 16.91, - "eval_steps_per_second": 2.296, + "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": 0.9443290831818184, - "precision": 0.9444332975177129, - "recall": 0.9445671641791045, - "support": 33500.0 + "f1-score": 0.9464708542988716, + "precision": 0.9466427045463329, + "recall": 0.9466309424671422, + "support": 27619.0 }, "step": 1701 }, { "epoch": 22.0, "eval_B": { - "f1-score": 0.8907216494845361, - "precision": 0.8816326530612245, - "recall": 0.9, - "support": 1440.0 + "f1-score": 0.875898418782942, + "precision": 0.8754789272030651, + "recall": 0.8763183125599233, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.961758728985775, - "precision": 0.9588801399825022, - "recall": 0.9646546532635383, - "support": 21587.0 + "f1-score": 0.9606979453806588, + "precision": 0.9475306911822412, + "recall": 0.9742363112391931, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9186952756663139, - "precision": 0.9258217783380199, - "recall": 0.9116776472834909, - "support": 10473.0 + "f1-score": 0.9217236387930074, + "precision": 0.94757326007326, + "recall": 0.897246910903967, + "support": 9226.0 }, - "eval_accuracy": 0.9453134328358209, - "eval_loss": 0.3998025953769684, + "eval_accuracy": 0.9448205945182664, + "eval_loss": 0.3899074196815491, "eval_macro avg": { - "f1-score": 0.9237252180455416, - "precision": 0.9221115237939155, - "recall": 0.925444100182343, - "support": 33500.0 - }, - "eval_runtime": 4.7313, - "eval_samples_per_second": 17.12, - "eval_steps_per_second": 2.325, + "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": 0.9452424023858794, - "precision": 0.9452247190072991, - "recall": 0.9453134328358209, - "support": 33500.0 + "f1-score": 0.9444764001104068, + "precision": 0.9448239585256736, + "recall": 0.9448205945182664, + "support": 27619.0 }, "step": 1782 }, { "epoch": 23.0, "eval_B": { - "f1-score": 0.8938053097345132, - "precision": 0.8765020026702269, - "recall": 0.9118055555555555, - "support": 1440.0 + "f1-score": 0.8766079085278704, + "precision": 0.8712121212121212, + "recall": 0.8820709491850431, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620159689850926, - "precision": 0.9584770313146641, - "recall": 0.9655811367952934, - "support": 21587.0 + "f1-score": 0.9571775605897973, + "precision": 0.9385663638378019, + "recall": 0.9765417867435159, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.918660748745658, - "precision": 0.9284251584592882, - "recall": 0.9090995894204144, - "support": 10473.0 + "f1-score": 0.9132322264193494, + "precision": 0.9515920573375631, + "recall": 0.877845220030349, + "support": 9226.0 }, - "eval_accuracy": 0.9456119402985075, - "eval_loss": 0.4420088231563568, + "eval_accuracy": 0.9400050689742568, + "eval_loss": 0.4506397247314453, "eval_macro avg": { - "f1-score": 0.9248273424884212, - "precision": 0.9211347308147264, - "recall": 0.9288287605904211, - "support": 33500.0 - }, - "eval_runtime": 4.7375, - "eval_samples_per_second": 17.098, - "eval_steps_per_second": 2.322, + "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": 0.9455299220929005, - "precision": 0.9455583057725045, - "recall": 0.9456119402985075, - "support": 33500.0 + "f1-score": 0.939455202786939, + "precision": 0.9403739808105457, + "recall": 0.9400050689742568, + "support": 27619.0 }, "step": 1863 }, { "epoch": 24.0, "eval_B": { - "f1-score": 0.89482699554642, - "precision": 0.8830290736984449, - "recall": 0.9069444444444444, - "support": 1440.0 + "f1-score": 0.8786853552440792, + "precision": 0.8859649122807017, + "recall": 0.8715244487056567, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.962135833103575, - "precision": 0.9544605005242285, - "recall": 0.969935609394543, - "support": 21587.0 + "f1-score": 0.9558590028580501, + "precision": 0.9575452599919023, + "recall": 0.954178674351585, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9185192391885975, - "precision": 0.9362356207854026, - "recall": 0.9014608994557434, - "support": 10473.0 + "f1-score": 0.9160280626011873, + "precision": 0.9121883061049011, + "recall": 0.9199002818122697, + "support": 9226.0 }, - "eval_accuracy": 0.945820895522388, - "eval_loss": 0.4822453558444977, + "eval_accuracy": 0.9396067924255042, + "eval_loss": 0.42392534017562866, "eval_macro avg": { - "f1-score": 0.9251606892795309, - "precision": 0.924575065002692, - "recall": 0.9261136510982436, - "support": 33500.0 - }, - "eval_runtime": 4.7345, - "eval_samples_per_second": 17.108, - "eval_steps_per_second": 2.323, + "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": 0.9456068386512209, - "precision": 0.945692428281427, - "recall": 0.945820895522388, - "support": 33500.0 + "f1-score": 0.9396392856607877, + "precision": 0.9396908279261412, + "recall": 0.9396067924255042, + "support": 27619.0 }, "step": 1944 }, { "epoch": 24.69, - "grad_norm": 0.10262901335954666, + "grad_norm": 0.030169092118740082, "learning_rate": 1.0123456790123458e-05, - "loss": 0.0037, + "loss": 0.0026, "step": 2000 }, { "epoch": 25.0, "eval_B": { - "f1-score": 0.8864342423196411, - "precision": 0.8812628689087165, - "recall": 0.8916666666666667, - "support": 1440.0 + "f1-score": 0.8887832699619772, + "precision": 0.88124410933082, + "recall": 0.8964525407478428, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9595285234126801, - "precision": 0.9537320946409775, - "recall": 0.9653958400889424, - "support": 21587.0 + "f1-score": 0.9607269808508199, + "precision": 0.9556860955857192, + "recall": 0.9658213256484149, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9129445923058311, - "precision": 0.9255298273155416, - "recall": 0.9006970304592762, - "support": 10473.0 + "f1-score": 0.9237260273972602, + "precision": 0.9340647163120568, + "recall": 0.9136137004118795, + "support": 9226.0 }, - "eval_accuracy": 0.942, - "eval_loss": 0.46240004897117615, + "eval_accuracy": 0.9457619754516818, + "eval_loss": 0.398232638835907, "eval_macro avg": { - "f1-score": 0.9196357860127174, - "precision": 0.9201749302884119, - "recall": 0.9192531790716284, - "support": 33500.0 - }, - "eval_runtime": 4.7408, - "eval_samples_per_second": 17.086, - "eval_steps_per_second": 2.32, + "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": 0.941823171912501, - "precision": 0.9418002131259403, - "recall": 0.942, - "support": 33500.0 + "f1-score": 0.9456501103261954, + "precision": 0.9456523566073828, + "recall": 0.9457619754516818, + "support": 27619.0 }, "step": 2025 }, { "epoch": 26.0, "eval_B": { - "f1-score": 0.892392049348869, - "precision": 0.8809201623815968, - "recall": 0.9041666666666667, - "support": 1440.0 + "f1-score": 0.8827717133364974, + "precision": 0.8740601503759399, + "recall": 0.8916586768935763, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9613188432492171, - "precision": 0.9556420233463035, - "recall": 0.9670635104461018, - "support": 21587.0 + "f1-score": 0.9590283123770422, + "precision": 0.9489364103142809, + "recall": 0.9693371757925072, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9172881355932202, - "precision": 0.9306278864105335, - "recall": 0.904325408192495, - "support": 10473.0 + "f1-score": 0.9188171447557869, + "precision": 0.9393115942028986, + "recall": 0.8991979189247779, + "support": 9226.0 }, - "eval_accuracy": 0.9447462686567164, - "eval_loss": 0.4685353934764862, + "eval_accuracy": 0.9429740396104132, + "eval_loss": 0.4066696763038635, "eval_macro avg": { - "f1-score": 0.9236663427304355, - "precision": 0.9223966907128114, - "recall": 0.9251851951017546, - "support": 33500.0 - }, - "eval_runtime": 4.7522, - "eval_samples_per_second": 17.045, - "eval_steps_per_second": 2.315, + "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": 0.9445908377418215, - "precision": 0.9446100073487664, - "recall": 0.9447462686567164, - "support": 33500.0 + "f1-score": 0.9427162132687112, + "precision": 0.942893668268613, + "recall": 0.9429740396104132, + "support": 27619.0 }, "step": 2106 }, { "epoch": 27.0, "eval_B": { - "f1-score": 0.8878923766816142, - "precision": 0.882111034955449, - "recall": 0.89375, - "support": 1440.0 + "f1-score": 0.8731307284129282, + "precision": 0.8786407766990292, + "recall": 0.8676893576222435, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9603469923524712, - "precision": 0.9467098748762265, - "recall": 0.9743827303469681, - "support": 21587.0 + "f1-score": 0.9568967963451661, + "precision": 0.9424291543234028, + "recall": 0.9718155619596541, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9131848640126133, - "precision": 0.943398147205538, - "recall": 0.8848467487825837, - "support": 10473.0 + "f1-score": 0.9140816781968311, + "precision": 0.9418257070590941, + "recall": 0.8879254281378712, + "support": 9226.0 }, - "eval_accuracy": 0.9429253731343283, - "eval_loss": 0.5011060833930969, + "eval_accuracy": 0.9398602411383468, + "eval_loss": 0.44146138429641724, "eval_macro avg": { - "f1-score": 0.9204747443488995, - "precision": 0.9240730190124045, - "recall": 0.9176598263765173, - "support": 33500.0 - }, - "eval_runtime": 4.7877, - "eval_samples_per_second": 16.918, - "eval_steps_per_second": 2.298, + "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": 0.942488376929505, - "precision": 0.9428977538797776, - "recall": 0.9429253731343283, - "support": 33500.0 + "f1-score": 0.9394312730137688, + "precision": 0.9398186802902107, + "recall": 0.9398602411383468, + "support": 27619.0 }, "step": 2187 }, { "epoch": 28.0, "eval_B": { - "f1-score": 0.8914835164835164, - "precision": 0.8817934782608695, - "recall": 0.9013888888888889, - "support": 1440.0 + "f1-score": 0.8846153846153846, + "precision": 0.8871745419479267, + "recall": 0.8820709491850431, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9599201652393882, - "precision": 0.9618175053483397, - "recall": 0.9580302960114884, - "support": 21587.0 + "f1-score": 0.9619680394066098, + "precision": 0.9559995446265938, + "recall": 0.9680115273775216, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9164245916472212, - "precision": 0.9141174235227056, - "recall": 0.9187434355008116, - "support": 10473.0 + "f1-score": 0.9258771929824562, + "precision": 0.9367650321721767, + "recall": 0.9152395404292217, + "support": 9226.0 }, - "eval_accuracy": 0.9433134328358209, - "eval_loss": 0.4657354950904846, + "eval_accuracy": 0.9471378398928274, + "eval_loss": 0.4067588746547699, "eval_macro avg": { - "f1-score": 0.9226094244567086, - "precision": 0.919242802377305, - "recall": 0.9260542068003962, - "support": 33500.0 - }, - "eval_runtime": 4.7715, - "eval_samples_per_second": 16.976, - "eval_steps_per_second": 2.305, + "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": 0.9433805259426951, - "precision": 0.9434653394986137, - "recall": 0.9433134328358209, - "support": 33500.0 + "f1-score": 0.9469909233612611, + "precision": 0.9469752465094172, + "recall": 0.9471378398928274, + "support": 27619.0 }, "step": 2268 }, { "epoch": 29.0, "eval_B": { - "f1-score": 0.8932038834951457, - "precision": 0.8623141564318035, - "recall": 0.9263888888888889, - "support": 1440.0 + "f1-score": 0.8919301557338367, + "precision": 0.8782527881040892, + "recall": 0.9060402684563759, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9610192139335576, - "precision": 0.9559099867088318, - "recall": 0.9661833510909343, - "support": 21587.0 + "f1-score": 0.964135443798838, + "precision": 0.9576391652925457, + "recall": 0.9707204610951009, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9164846896685592, - "precision": 0.9318136964673377, - "recall": 0.9016518667048601, - "support": 10473.0 + "f1-score": 0.9300406995930042, + "precision": 0.944059848146494, + "recall": 0.9164318231086062, + "support": 9226.0 }, - "eval_accuracy": 0.9442985074626866, - "eval_loss": 0.45431333780288696, + "eval_accuracy": 0.9501430174879612, + "eval_loss": 0.4006032943725586, "eval_macro avg": { - "f1-score": 0.9235692623657542, - "precision": 0.9166792798693243, - "recall": 0.931408035561561, - "support": 33500.0 - }, - "eval_runtime": 4.7759, - "eval_samples_per_second": 16.96, - "eval_steps_per_second": 2.303, + "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": 0.9441814781586727, - "precision": 0.9443536152670977, - "recall": 0.9442985074626866, - "support": 33500.0 + "f1-score": 0.9500195009517102, + "precision": 0.9501051209246456, + "recall": 0.9501430174879612, + "support": 27619.0 }, "step": 2349 }, { "epoch": 30.0, "eval_B": { - "f1-score": 0.8953290146607569, - "precision": 0.8794373744139317, - "recall": 0.9118055555555555, - "support": 1440.0 + "f1-score": 0.8796956728483118, + "precision": 0.8726415094339622, + "recall": 0.8868648130393096, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620589989448088, - "precision": 0.9528786295269687, - "recall": 0.9714179830453513, - "support": 21587.0 + "f1-score": 0.9600954979536152, + "precision": 0.947067399349557, + "recall": 0.9734870317002882, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9176964782884776, - "precision": 0.9394, - "recall": 0.896973169101499, - "support": 10473.0 + "f1-score": 0.9212857222438862, + "precision": 0.9477363896848138, + "recall": 0.8962714068935617, + "support": 9226.0 }, - "eval_accuracy": 0.9455820895522388, - "eval_loss": 0.4872027635574341, + "eval_accuracy": 0.9444223179695137, + "eval_loss": 0.46608424186706543, "eval_macro avg": { - "f1-score": 0.9250281639646811, - "precision": 0.9239053346469669, - "recall": 0.926732235900802, - "support": 33500.0 - }, - "eval_runtime": 4.7685, - "eval_samples_per_second": 16.987, - "eval_steps_per_second": 2.307, + "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": 0.945321689804427, - "precision": 0.9455079699926786, - "recall": 0.9455820895522388, - "support": 33500.0 + "f1-score": 0.9440950631702127, + "precision": 0.9444802637418637, + "recall": 0.9444223179695137, + "support": 27619.0 }, "step": 2430 }, { "epoch": 30.86, - "grad_norm": 8.135757446289062, + "grad_norm": 0.0009192481520585716, "learning_rate": 7.654320987654322e-06, - "loss": 0.0025, + "loss": 0.0015, "step": 2500 }, { "epoch": 31.0, "eval_B": { - "f1-score": 0.8891177480260899, - "precision": 0.8791581805838425, - "recall": 0.8993055555555556, - "support": 1440.0 + "f1-score": 0.8760724499523355, + "precision": 0.8710900473933649, + "recall": 0.8811121764141898, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9598950783461034, - "precision": 0.9535978787601719, - "recall": 0.9662759994441099, - "support": 21587.0 + "f1-score": 0.9596285601041844, + "precision": 0.9430224794124193, + "recall": 0.9768299711815562, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9137011538834482, - "precision": 0.9281000689451394, - "recall": 0.8997421942136924, - "support": 10473.0 + "f1-score": 0.9194073408912335, + "precision": 0.953328677839851, + "recall": 0.8878170388033817, + "support": 9226.0 }, - "eval_accuracy": 0.9425970149253732, - "eval_loss": 0.4678420424461365, + "eval_accuracy": 0.9434809370360984, + "eval_loss": 0.48370352387428284, "eval_macro avg": { - "f1-score": 0.9209046600852139, - "precision": 0.9202853760963846, - "recall": 0.9217745830711194, - "support": 33500.0 - }, - "eval_runtime": 4.7618, - "eval_samples_per_second": 17.01, - "eval_steps_per_second": 2.31, + "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": 0.9424112477025748, - "precision": 0.9424267824148957, - "recall": 0.9425970149253732, - "support": 33500.0 + "f1-score": 0.943037445605214, + "precision": 0.9437487714612122, + "recall": 0.9434809370360984, + "support": 27619.0 }, "step": 2511 }, { "epoch": 32.0, "eval_B": { - "f1-score": 0.8871355060034306, - "precision": 0.8766101694915254, - "recall": 0.8979166666666667, - "support": 1440.0 + "f1-score": 0.8770764119601329, + "precision": 0.868421052631579, + "recall": 0.8859060402684564, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9603147215128719, - "precision": 0.9538848263254114, - "recall": 0.966831889563163, - "support": 21587.0 + "f1-score": 0.9595270577535243, + "precision": 0.9465066726477515, + "recall": 0.9729106628242075, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9144436899796294, - "precision": 0.9292262198127156, - "recall": 0.9001241287119259, - "support": 10473.0 + "f1-score": 0.919819468434836, + "precision": 0.9464510950579063, + "recall": 0.8946455668762194, + "support": 9226.0 }, - "eval_accuracy": 0.9430149253731344, - "eval_loss": 0.4877029359340668, + "eval_accuracy": 0.9434809370360984, + "eval_loss": 0.4701555371284485, "eval_macro avg": { - "f1-score": 0.9206313058319773, - "precision": 0.9199070718765507, - "recall": 0.9216242283139185, - "support": 33500.0 - }, - "eval_runtime": 4.759, - "eval_samples_per_second": 17.02, - "eval_steps_per_second": 2.311, + "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": 0.9428285906597003, - "precision": 0.942854226568747, - "recall": 0.9430149253731344, - "support": 33500.0 + "f1-score": 0.943149265559139, + "precision": 0.9435392929265168, + "recall": 0.9434809370360984, + "support": 27619.0 }, "step": 2592 }, { "epoch": 33.0, "eval_B": { - "f1-score": 0.8910756701730572, - "precision": 0.8712674187126742, - "recall": 0.9118055555555555, - "support": 1440.0 + "f1-score": 0.8849976336961666, + "precision": 0.8738317757009346, + "recall": 0.8964525407478428, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9623726096415894, - "precision": 0.9516713470423046, - "recall": 0.9733172742854496, - "support": 21587.0 + "f1-score": 0.9615693472524951, + "precision": 0.9569585569128896, + "recall": 0.9662247838616714, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9176966843241121, - "precision": 0.9435199193141705, - "recall": 0.8932493077437219, - "support": 10473.0 + "f1-score": 0.9260009859232076, + "precision": 0.9359982283246595, + "recall": 0.9162150444396271, + "support": 9226.0 }, - "eval_accuracy": 0.9456417910447761, - "eval_loss": 0.49512049555778503, + "eval_accuracy": 0.9468843911799848, + "eval_loss": 0.4380284249782562, "eval_macro avg": { - "f1-score": 0.9237149880462528, - "precision": 0.9221528950230496, - "recall": 0.9261240458615757, - "support": 33500.0 - }, - "eval_runtime": 4.765, - "eval_samples_per_second": 16.999, - "eval_steps_per_second": 2.308, + "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": 0.9453410108748842, - "precision": 0.9456668228813668, - "recall": 0.9456417910447761, - "support": 33500.0 + "f1-score": 0.9467962563055651, + "precision": 0.946817667512148, + "recall": 0.9468843911799848, + "support": 27619.0 }, "step": 2673 }, { "epoch": 34.0, "eval_B": { - "f1-score": 0.8933423913043478, - "precision": 0.8743351063829787, - "recall": 0.9131944444444444, - "support": 1440.0 + "f1-score": 0.8833652007648184, + "precision": 0.880838894184938, + "recall": 0.8859060402684564, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620096067661051, - "precision": 0.9546159459952563, - "recall": 0.9695186918052532, - "support": 21587.0 + "f1-score": 0.9618386073235572, + "precision": 0.9541201156921681, + "recall": 0.9696829971181556, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9179849111706012, - "precision": 0.9362589356632248, - "recall": 0.9004105795856011, - "support": 10473.0 + "f1-score": 0.9257281286131146, + "precision": 0.9406959829920555, + "recall": 0.9112291350531108, + "support": 9226.0 }, - "eval_accuracy": 0.9454925373134329, - "eval_loss": 0.47959965467453003, + "eval_accuracy": 0.9469930120569173, + "eval_loss": 0.4419035315513611, "eval_macro avg": { - "f1-score": 0.9244456364136847, - "precision": 0.9217366626804866, - "recall": 0.9277079052784329, - "support": 33500.0 - }, - "eval_runtime": 4.7437, - "eval_samples_per_second": 17.075, - "eval_steps_per_second": 2.319, + "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": 0.9452946387888919, - "precision": 0.9454261735102102, - "recall": 0.9454925373134329, - "support": 33500.0 + "f1-score": 0.9468126092923719, + "precision": 0.9468684642086502, + "recall": 0.9469930120569173, + "support": 27619.0 }, "step": 2754 }, { "epoch": 35.0, "eval_B": { - "f1-score": 0.8896504455106238, - "precision": 0.878213802435724, - "recall": 0.9013888888888889, - "support": 1440.0 + "f1-score": 0.8785405664906385, + "precision": 0.8798076923076923, + "recall": 0.8772770853307766, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9619017054832203, - "precision": 0.9520805917320869, - "recall": 0.9719275489878167, - "support": 21587.0 + "f1-score": 0.9603404255319149, + "precision": 0.945586592178771, + "recall": 0.9755619596541787, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.917587251930785, - "precision": 0.9400100150225338, - "recall": 0.896209300105032, - "support": 10473.0 + "f1-score": 0.9217537000837756, + "precision": 0.9508007835004033, + "recall": 0.8944287882072404, + "support": 9226.0 }, - "eval_accuracy": 0.9452238805970149, - "eval_loss": 0.4806292951107025, + "eval_accuracy": 0.9447481806003114, + "eval_loss": 0.4607029855251312, "eval_macro avg": { - "f1-score": 0.923046467641543, - "precision": 0.9234348030634482, - "recall": 0.9231752459939125, - "support": 33500.0 - }, - "eval_runtime": 4.73, - "eval_samples_per_second": 17.125, - "eval_steps_per_second": 2.326, + "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": 0.9449420909633638, - "precision": 0.9451318357181789, - "recall": 0.9452238805970149, - "support": 33500.0 + "f1-score": 0.9443616289800995, + "precision": 0.9448443037746955, + "recall": 0.9447481806003114, + "support": 27619.0 }, "step": 2835 }, { "epoch": 36.0, "eval_B": { - "f1-score": 0.8926979773740144, - "precision": 0.8815165876777251, - "recall": 0.9041666666666667, - "support": 1440.0 + "f1-score": 0.8802714493456132, + "precision": 0.8901960784313725, + "recall": 0.8705656759348035, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9623375129414473, - "precision": 0.9559374714324893, - "recall": 0.9688238291564367, - "support": 21587.0 + "f1-score": 0.9600229918091681, + "precision": 0.9574089997133849, + "recall": 0.9626512968299712, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9190998156950239, - "precision": 0.9339576145884673, - "recall": 0.9047073426907285, - "support": 10473.0 + "f1-score": 0.9233949945593034, + "precision": 0.9270264365304784, + "recall": 0.9197918924777801, + "support": 9226.0 }, - "eval_accuracy": 0.946, - "eval_loss": 0.4832863509654999, + "eval_accuracy": 0.9448568014772439, + "eval_loss": 0.44131579995155334, "eval_macro avg": { - "f1-score": 0.9247117686701619, - "precision": 0.923803891232894, - "recall": 0.9258992795046107, - "support": 33500.0 - }, - "eval_runtime": 4.7435, - "eval_samples_per_second": 17.076, - "eval_steps_per_second": 2.319, + "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": 0.9458267865390921, - "precision": 0.9458669904374355, - "recall": 0.946, - "support": 33500.0 + "f1-score": 0.944775851745562, + "precision": 0.9447216249053674, + "recall": 0.9448568014772439, + "support": 27619.0 }, "step": 2916 }, { "epoch": 37.0, "eval_B": { - "f1-score": 0.8904958677685951, - "precision": 0.8831967213114754, - "recall": 0.8979166666666667, - "support": 1440.0 + "f1-score": 0.8851513695338779, + "precision": 0.8872832369942196, + "recall": 0.8830297219558965, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9614206416841329, - "precision": 0.9560696289509849, - "recall": 0.966831889563163, - "support": 21587.0 + "f1-score": 0.9621385956860945, + "precision": 0.9576338928856915, + "recall": 0.966685878962536, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9176459209826394, - "precision": 0.9296492259455222, - "recall": 0.9059486298099876, - "support": 10473.0 + "f1-score": 0.9272399278412508, + "precision": 0.9353700231609132, + "recall": 0.9192499458053327, + "support": 9226.0 }, - "eval_accuracy": 0.9448358208955224, - "eval_loss": 0.4824429750442505, + "eval_accuracy": 0.9476809442774902, + "eval_loss": 0.4391210675239563, "eval_macro avg": { - "f1-score": 0.9231874768117891, - "precision": 0.9229718587359942, - "recall": 0.9235657286799391, - "support": 33500.0 - }, - "eval_runtime": 4.7583, - "eval_samples_per_second": 17.023, - "eval_steps_per_second": 2.312, + "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": 0.9446867812559204, - "precision": 0.9446774537964444, - "recall": 0.9448358208955224, - "support": 33500.0 + "f1-score": 0.9475735214106575, + "precision": 0.9475400373450994, + "recall": 0.9476809442774902, + "support": 27619.0 }, "step": 2997 }, { "epoch": 37.04, - "grad_norm": 0.2996862828731537, + "grad_norm": 0.36456331610679626, "learning_rate": 5.185185185185185e-06, - "loss": 0.0016, + "loss": 0.0012, "step": 3000 }, { "epoch": 38.0, "eval_B": { - "f1-score": 0.8921302578018996, - "precision": 0.8720159151193634, - "recall": 0.9131944444444444, - "support": 1440.0 + "f1-score": 0.8848715509039011, + "precision": 0.8781869688385269, + "recall": 0.8916586768935763, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9625111778598124, - "precision": 0.9529192772178334, - "recall": 0.9722981424005188, - "support": 21587.0 + "f1-score": 0.9604875396984349, + "precision": 0.9536389977842168, + "recall": 0.9674351585014409, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9186359410930084, - "precision": 0.9420028095524784, - "recall": 0.8964002673541488, - "support": 10473.0 + "f1-score": 0.9237283475391805, + "precision": 0.9374930237749749, + "recall": 0.9103620203771949, + "support": 9226.0 }, - "eval_accuracy": 0.9460298507462687, - "eval_loss": 0.49525222182273865, + "eval_accuracy": 0.9455085267388392, + "eval_loss": 0.43638840317726135, "eval_macro avg": { - "f1-score": 0.9244257922515735, - "precision": 0.9223126672965584, - "recall": 0.9272976180663707, - "support": 33500.0 - }, - "eval_runtime": 4.7487, - "eval_samples_per_second": 17.057, - "eval_steps_per_second": 2.316, + "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": 0.9457692710078085, - "precision": 0.9460288590900406, - "recall": 0.9460298507462687, - "support": 33500.0 + "f1-score": 0.9453527490407726, + "precision": 0.9453961496579408, + "recall": 0.9455085267388392, + "support": 27619.0 }, "step": 3078 }, { "epoch": 39.0, "eval_B": { - "f1-score": 0.8914011647824598, - "precision": 0.8796484110885734, - "recall": 0.9034722222222222, - "support": 1440.0 + "f1-score": 0.8896321070234113, + "precision": 0.8866666666666667, + "recall": 0.8926174496644296, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9630648600316389, - "precision": 0.9533817521561507, - "recall": 0.9729466808727475, - "support": 21587.0 + "f1-score": 0.9625824964131994, + "precision": 0.9584571428571429, + "recall": 0.9667435158501441, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9199569976544176, - "precision": 0.9421479331398258, - "recall": 0.8987873579681085, - "support": 10473.0 + "f1-score": 0.9277944793659471, + "precision": 0.9358253390671518, + "recall": 0.9199002818122697, + "support": 9226.0 }, - "eval_accuracy": 0.9467761194029851, - "eval_loss": 0.4956269860267639, + "eval_accuracy": 0.9482964625801079, + "eval_loss": 0.44264644384384155, "eval_macro avg": { - "f1-score": 0.9248076741561722, - "precision": 0.9250593654615167, - "recall": 0.9250687536876927, - "support": 33500.0 - }, - "eval_runtime": 4.7588, - "eval_samples_per_second": 17.021, - "eval_steps_per_second": 2.312, + "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": 0.9465077148425208, - "precision": 0.9467003253592766, - "recall": 0.9467761194029851, - "support": 33500.0 + "f1-score": 0.9482068310592221, + "precision": 0.9481860074636412, + "recall": 0.9482964625801079, + "support": 27619.0 }, "step": 3159 }, { "epoch": 40.0, "eval_B": { - "f1-score": 0.8935439560439561, - "precision": 0.8838315217391305, - "recall": 0.9034722222222222, - "support": 1440.0 + "f1-score": 0.8865485878410723, + "precision": 0.8852772466539197, + "recall": 0.887823585810163, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.962365961745999, - "precision": 0.9541067200874158, - "recall": 0.9707694445731228, - "support": 21587.0 + "f1-score": 0.9624967964233846, + "precision": 0.9512016660100185, + "recall": 0.9740634005763689, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9186346593952379, - "precision": 0.9373012718600954, - "recall": 0.9006970304592762, - "support": 10473.0 + "f1-score": 0.9265749778172139, + "precision": 0.9486713604360664, + "recall": 0.905484500325168, + "support": 9226.0 }, - "eval_accuracy": 0.9459701492537314, - "eval_loss": 0.48434382677078247, + "eval_accuracy": 0.9478981860313552, + "eval_loss": 0.4583560824394226, "eval_macro avg": { - "f1-score": 0.9248481923950643, - "precision": 0.9250798378955473, - "recall": 0.9249795657515404, - "support": 33500.0 - }, - "eval_runtime": 4.7973, - "eval_samples_per_second": 16.884, - "eval_steps_per_second": 2.293, + "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": 0.9457360627092687, - "precision": 0.9458321008364828, - "recall": 0.9459701492537314, - "support": 33500.0 + "f1-score": 0.9476291806512029, + "precision": 0.947866868638148, + "recall": 0.9478981860313552, + "support": 27619.0 }, "step": 3240 }, { "epoch": 41.0, "eval_B": { - "f1-score": 0.8901893287435455, - "precision": 0.8825938566552901, - "recall": 0.8979166666666667, - "support": 1440.0 + "f1-score": 0.886558627264061, + "precision": 0.8815165876777251, + "recall": 0.8916586768935763, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9599889178768498, - "precision": 0.9569179784589892, - "recall": 0.9630796312595543, - "support": 21587.0 + "f1-score": 0.9625846436359462, + "precision": 0.9584047537424294, + "recall": 0.9668011527377521, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.915215089981715, - "precision": 0.9224949073624988, - "recall": 0.9080492695502721, - "support": 10473.0 + "f1-score": 0.927930883639545, + "precision": 0.9363275215184286, + "recall": 0.9196835031432907, + "support": 9226.0 }, - "eval_accuracy": 0.9430746268656717, - "eval_loss": 0.4890231490135193, + "eval_accuracy": 0.9482240486621528, + "eval_loss": 0.45645976066589355, "eval_macro avg": { - "f1-score": 0.92179777886737, - "precision": 0.920668914158926, - "recall": 0.9230151891588311, - "support": 33500.0 - }, - "eval_runtime": 4.7614, - "eval_samples_per_second": 17.012, - "eval_steps_per_second": 2.31, + "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": 0.9429910758500526, - "precision": 0.942961573712993, - "recall": 0.9430746268656717, - "support": 33500.0 + "f1-score": 0.9481376786914271, + "precision": 0.9481263619938463, + "recall": 0.9482240486621528, + "support": 27619.0 }, "step": 3321 }, { "epoch": 42.0, "eval_B": { - "f1-score": 0.8924028875902373, - "precision": 0.8835942818243703, - "recall": 0.9013888888888889, - "support": 1440.0 + "f1-score": 0.8838095238095238, + "precision": 0.8779564806054873, + "recall": 0.8897411313518696, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9607680236314964, - "precision": 0.9572775350655323, - "recall": 0.9642840598508362, - "support": 21587.0 + "f1-score": 0.9627247894377418, + "precision": 0.9507137237270991, + "recall": 0.9750432276657061, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9167108242208198, - "precision": 0.9250437487847559, - "recall": 0.9085266876730641, - "support": 10473.0 + "f1-score": 0.9265310659108592, + "precision": 0.9507299270072993, + "recall": 0.9035334923043572, + "support": 9226.0 }, - "eval_accuracy": 0.9441492537313433, - "eval_loss": 0.493211567401886, + "eval_accuracy": 0.9479343929903328, + "eval_loss": 0.4557858109474182, "eval_macro avg": { - "f1-score": 0.9232939118141844, - "precision": 0.9219718552248861, - "recall": 0.9247332121375963, - "support": 33500.0 - }, - "eval_runtime": 4.7587, - "eval_samples_per_second": 17.022, - "eval_steps_per_second": 2.312, + "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": 0.9440558789948568, - "precision": 0.9440331073525219, - "recall": 0.9441492537313433, - "support": 33500.0 + "f1-score": 0.947654297555007, + "precision": 0.9479715421451187, + "recall": 0.9479343929903328, + "support": 27619.0 }, "step": 3402 }, { "epoch": 43.0, "eval_B": { - "f1-score": 0.8943758573388203, - "precision": 0.8834688346883469, - "recall": 0.9055555555555556, - "support": 1440.0 + "f1-score": 0.8847795163584636, + "precision": 0.875234521575985, + "recall": 0.8945349952061361, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620124858900229, - "precision": 0.956832554302997, - "recall": 0.9672488071524529, - "support": 21587.0 + "f1-score": 0.9635252309879175, + "precision": 0.9506872370266479, + "recall": 0.9767146974063401, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9189842805320435, - "precision": 0.9311899627524015, - "recall": 0.9070944333046882, - "support": 10473.0 + "f1-score": 0.9281497159407374, + "precision": 0.9546287809349221, + "recall": 0.9030999349663993, + "support": 9226.0 }, - "eval_accuracy": 0.9457910447761194, - "eval_loss": 0.48890525102615356, + "eval_accuracy": 0.9490206017596582, + "eval_loss": 0.46096372604370117, "eval_macro avg": { - "f1-score": 0.9251242079202955, - "precision": 0.9238304505812485, - "recall": 0.9266329320042322, - "support": 33500.0 - }, - "eval_runtime": 4.7372, - "eval_samples_per_second": 17.099, - "eval_steps_per_second": 2.322, + "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": 0.945653347387699, - "precision": 0.945662446316296, - "recall": 0.9457910447761194, - "support": 33500.0 + "f1-score": 0.9487344607868312, + "precision": 0.949154506003899, + "recall": 0.9490206017596582, + "support": 27619.0 }, "step": 3483 }, { "epoch": 43.21, - "grad_norm": 0.36121609807014465, + "grad_norm": 0.008026141673326492, "learning_rate": 2.7160493827160496e-06, - "loss": 0.0015, + "loss": 0.0008, "step": 3500 }, { "epoch": 44.0, "eval_B": { - "f1-score": 0.8921703296703297, - "precision": 0.8824728260869565, - "recall": 0.9020833333333333, - "support": 1440.0 + "f1-score": 0.8856053384175404, + "precision": 0.8805687203791469, + "recall": 0.8906999041227229, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9620328986776021, - "precision": 0.956918282231083, - "recall": 0.9672024829758651, - "support": 21587.0 + "f1-score": 0.9627508804077075, + "precision": 0.9565341070717415, + "recall": 0.9690489913544669, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9190600522193212, - "precision": 0.9309432853364679, - "recall": 0.9074763678029218, - "support": 10473.0 + "f1-score": 0.9280184483610608, + "precision": 0.9403582953154557, + "recall": 0.9159982657706481, + "support": 9226.0 }, - "eval_accuracy": 0.9457313432835821, - "eval_loss": 0.49400997161865234, + "eval_accuracy": 0.9483688764980629, + "eval_loss": 0.4500958025455475, "eval_macro avg": { - "f1-score": 0.9244210935224176, - "precision": 0.9234447978848358, - "recall": 0.9255873947040402, - "support": 33500.0 - }, - "eval_runtime": 4.7323, - "eval_samples_per_second": 17.116, - "eval_steps_per_second": 2.324, + "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": 0.9455953846379588, - "precision": 0.9455977568781024, - "recall": 0.9457313432835821, - "support": 33500.0 + "f1-score": 0.9482353578197027, + "precision": 0.9482619054140469, + "recall": 0.9483688764980629, + "support": 27619.0 }, "step": 3564 }, { "epoch": 45.0, "eval_B": { - "f1-score": 0.8950808393532852, - "precision": 0.8868438991138378, - "recall": 0.9034722222222222, - "support": 1440.0 + "f1-score": 0.8907963757749165, + "precision": 0.8861480075901328, + "recall": 0.8954937679769894, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9623735050597976, - "precision": 0.9556479239939707, - "recall": 0.9691944225691388, - "support": 21587.0 + "f1-score": 0.9618886240271116, + "precision": 0.9586123991069895, + "recall": 0.9651873198847263, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9191287051860476, - "precision": 0.9342209072978304, - "recall": 0.9045163754416118, - "support": 10473.0 + "f1-score": 0.9268638794891386, + "precision": 0.9334872471416007, + "recall": 0.9203338391502276, + "support": 9226.0 }, - "eval_accuracy": 0.9461492537313433, - "eval_loss": 0.49951982498168945, + "eval_accuracy": 0.9475723234005576, + "eval_loss": 0.4541914463043213, "eval_macro avg": { - "f1-score": 0.92552768319971, - "precision": 0.925570910135213, - "recall": 0.9257276734109908, - "support": 33500.0 - }, - "eval_runtime": 4.7702, - "eval_samples_per_second": 16.98, - "eval_steps_per_second": 2.306, + "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": 0.9459614385614344, - "precision": 0.9459917167794611, - "recall": 0.9461492537313433, - "support": 33500.0 + "f1-score": 0.9475040515214316, + "precision": 0.9474829225732714, + "recall": 0.9475723234005576, + "support": 27619.0 }, "step": 3645 }, { "epoch": 46.0, "eval_B": { - "f1-score": 0.8926862611073137, - "precision": 0.8788694481830417, - "recall": 0.9069444444444444, - "support": 1440.0 + "f1-score": 0.8897338403041826, + "precision": 0.882186616399623, + "recall": 0.8974113135186961, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9626744265976216, - "precision": 0.9524392455567646, - "recall": 0.9731319775790985, - "support": 21587.0 + "f1-score": 0.9623334864492421, + "precision": 0.9587003775311749, + "recall": 0.9659942363112392, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.918604082032206, - "precision": 0.9423579031934123, - "recall": 0.8960183328559153, - "support": 10473.0 + "f1-score": 0.9276581794339416, + "precision": 0.9353239312472454, + "recall": 0.9201170604812486, + "support": 9226.0 }, - "eval_accuracy": 0.9461791044776119, - "eval_loss": 0.5005789995193481, + "eval_accuracy": 0.9480792208262429, + "eval_loss": 0.45604071021080017, "eval_macro avg": { - "f1-score": 0.9246549232457139, - "precision": 0.9245555323110729, - "recall": 0.9253649182931527, - "support": 33500.0 - }, - "eval_runtime": 4.7467, - "eval_samples_per_second": 17.064, - "eval_steps_per_second": 2.317, + "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": 0.9458884063904681, - "precision": 0.9461251438615541, - "recall": 0.9461791044776119, - "support": 33500.0 + "f1-score": 0.9480087167815329, + "precision": 0.9480021282724854, + "recall": 0.9480792208262429, + "support": 27619.0 }, "step": 3726 }, { "epoch": 47.0, "eval_B": { - "f1-score": 0.8930774503084304, - "precision": 0.881596752368065, - "recall": 0.9048611111111111, - "support": 1440.0 + "f1-score": 0.8893111638954868, + "precision": 0.8813559322033898, + "recall": 0.8974113135186961, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9616694275722437, - "precision": 0.9552955158385519, - "recall": 0.9681289665076204, - "support": 21587.0 + "f1-score": 0.9612492437120056, + "precision": 0.961000057607005, + "recall": 0.9614985590778098, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9176447764089631, - "precision": 0.9324790537210449, - "recall": 0.9032750883223527, - "support": 10473.0 + "f1-score": 0.9265089014329135, + "precision": 0.9279191128506197, + "recall": 0.925102969867765, + "support": 9226.0 }, - "eval_accuracy": 0.945134328358209, - "eval_loss": 0.4959483742713928, + "eval_accuracy": 0.9469205981389623, + "eval_loss": 0.46677273511886597, "eval_macro avg": { - "f1-score": 0.9241305514298791, - "precision": 0.9231237739758873, - "recall": 0.9254217219803614, - "support": 33500.0 - }, - "eval_runtime": 4.7667, - "eval_samples_per_second": 16.993, - "eval_steps_per_second": 2.308, + "f1-score": 0.9256897696801353, + "precision": 0.9234250342203382, + "recall": 0.9280042808214236, + "support": 27619.0 + }, + "eval_runtime": 4.5969, + "eval_samples_per_second": 17.403, + "eval_steps_per_second": 2.175, "eval_weighted avg": { - "f1-score": 0.9449577076052906, - "precision": 0.9449945299533534, - "recall": 0.945134328358209, - "support": 33500.0 + "f1-score": 0.9469277326103895, + "precision": 0.9469418506075343, + "recall": 0.9469205981389623, + "support": 27619.0 }, "step": 3807 }, { "epoch": 48.0, "eval_B": { - "f1-score": 0.8924657534246575, - "precision": 0.8804054054054054, - "recall": 0.9048611111111111, - "support": 1440.0 + "f1-score": 0.8897338403041826, + "precision": 0.882186616399623, + "recall": 0.8974113135186961, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.961560580616963, - "precision": 0.9550356424785231, - "recall": 0.968175290684208, - "support": 21587.0 + "f1-score": 0.961359055571552, + "precision": 0.9605293440736479, + "recall": 0.9621902017291066, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9173661992333446, - "precision": 0.9326164167324388, - "recall": 0.902606702950444, - "support": 10473.0 + "f1-score": 0.9265377091936535, + "precision": 0.9289605578557419, + "recall": 0.9241274658573596, + "support": 9226.0 }, - "eval_accuracy": 0.944955223880597, - "eval_loss": 0.4958418905735016, + "eval_accuracy": 0.9470292190158949, + "eval_loss": 0.464282363653183, "eval_macro avg": { - "f1-score": 0.9237975110916551, - "precision": 0.9226858215387891, - "recall": 0.9252143682485877, - "support": 33500.0 - }, - "eval_runtime": 4.7381, - "eval_samples_per_second": 17.096, - "eval_steps_per_second": 2.322, + "f1-score": 0.9258768683564625, + "precision": 0.9238921727763376, + "recall": 0.9279096603683875, + "support": 27619.0 + }, + "eval_runtime": 4.6, + "eval_samples_per_second": 17.391, + "eval_steps_per_second": 2.174, "eval_weighted avg": { - "f1-score": 0.9447741833815136, - "precision": 0.9448188038927311, - "recall": 0.944955223880597, - "support": 33500.0 + "f1-score": 0.947022300395537, + "precision": 0.9470254124826993, + "recall": 0.9470292190158949, + "support": 27619.0 }, "step": 3888 }, { "epoch": 49.0, "eval_B": { - "f1-score": 0.8926979773740144, - "precision": 0.8815165876777251, - "recall": 0.9041666666666667, - "support": 1440.0 + "f1-score": 0.8886774500475738, + "precision": 0.8819641170915958, + "recall": 0.8954937679769894, + "support": 1043.0 }, "eval_I": { - "f1-score": 0.9613878265677557, - "precision": 0.9562766396260141, - "recall": 0.9665539445036364, - "support": 21587.0 + "f1-score": 0.961645645990749, + "precision": 0.958698516354471, + "recall": 0.9646109510086456, + "support": 17350.0 }, "eval_O": { - "f1-score": 0.9174445035546742, - "precision": 0.9295374362994904, - "recall": 0.9056621789363124, - "support": 10473.0 + "f1-score": 0.9265099023405532, + "precision": 0.9327694166758211, + "recall": 0.9203338391502276, + "support": 9226.0 }, - "eval_accuracy": 0.9448358208955224, - "eval_loss": 0.49691277742385864, + "eval_accuracy": 0.9472102538107824, + "eval_loss": 0.4589785635471344, "eval_macro avg": { - "f1-score": 0.923843435832148, - "precision": 0.9224435545344098, - "recall": 0.9254609300355385, - "support": 33500.0 - }, - "eval_runtime": 4.7348, - "eval_samples_per_second": 17.107, - "eval_steps_per_second": 2.323, + "f1-score": 0.9256109994596254, + "precision": 0.9244773500406293, + "recall": 0.9268128527119542, + "support": 27619.0 + }, + "eval_runtime": 4.6098, + "eval_samples_per_second": 17.354, + "eval_steps_per_second": 2.169, "eval_weighted avg": { - "f1-score": 0.9446973249332784, - "precision": 0.9447036798873807, - "recall": 0.9448358208955224, - "support": 33500.0 + "f1-score": 0.9471531517192171, + "precision": 0.9471392328153709, + "recall": 0.9472102538107824, + "support": 27619.0 }, "step": 3969 } @@ -1974,7 +1974,7 @@ "num_input_tokens_seen": 0, "num_train_epochs": 50, "save_steps": 500, - "total_flos": 7023665406454200.0, + "total_flos": 7045545984044400.0, "train_batch_size": 4, "trial_name": null, "trial_params": null