| { | |
| "best_metric": null, | |
| "best_model_checkpoint": null, | |
| "epoch": 19.0, | |
| "eval_steps": 500, | |
| "global_step": 1539, | |
| "is_hyper_param_search": false, | |
| "is_local_process_zero": true, | |
| "is_world_process_zero": true, | |
| "log_history": [ | |
| { | |
| "epoch": 1.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.17153996101364524, | |
| "precision": 0.25287356321839083, | |
| "recall": 0.12979351032448377, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.08383233532934131, | |
| "precision": 1.0, | |
| "recall": 0.04375, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8469620831429876, | |
| "precision": 0.7427884615384616, | |
| "recall": 0.9851222104144527, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.5420081967213116, | |
| "precision": 0.6803858520900321, | |
| "recall": 0.4504044274159217, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8343351767251743, | |
| "precision": 0.8141717503519474, | |
| "recall": 0.8555226824457594, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9004323458767015, | |
| "precision": 0.8590809093131264, | |
| "recall": 0.9459659511472983, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9988389078242631, | |
| "precision": 0.9991008541885209, | |
| "recall": 0.9985770987793006, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.8856789444749863, | |
| "eval_loss": 0.28347474336624146, | |
| "eval_macro avg": { | |
| "f1-score": 0.6254212866619178, | |
| "precision": 0.76405734152864, | |
| "recall": 0.629876554361031, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0448, | |
| "eval_samples_per_second": 16.056, | |
| "eval_steps_per_second": 2.18, | |
| "eval_weighted avg": { | |
| "f1-score": 0.8748148382495958, | |
| "precision": 0.8768575102351055, | |
| "recall": 0.8856789444749863, | |
| "support": 36380.0 | |
| }, | |
| "step": 81 | |
| }, | |
| { | |
| "epoch": 2.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.5137931034482759, | |
| "precision": 0.6182572614107884, | |
| "recall": 0.43952802359882004, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.7622641509433963, | |
| "precision": 0.9619047619047619, | |
| "recall": 0.63125, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8853910477127397, | |
| "precision": 0.8241758241758241, | |
| "recall": 0.9564293304994687, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.5413588470388762, | |
| "precision": 0.6279853891542568, | |
| "recall": 0.47573435504469985, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.7949260042283298, | |
| "precision": 0.856492027334852, | |
| "recall": 0.7416173570019724, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9007064868336545, | |
| "precision": 0.8614779777627618, | |
| "recall": 0.9436780835744566, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9988389947942025, | |
| "precision": 0.9990260713215463, | |
| "recall": 0.9986519883172321, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.8864211105002748, | |
| "eval_loss": 0.2809496521949768, | |
| "eval_macro avg": { | |
| "f1-score": 0.7710398049999251, | |
| "precision": 0.8213313304378275, | |
| "recall": 0.7409841625766643, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0579, | |
| "eval_samples_per_second": 16.014, | |
| "eval_steps_per_second": 2.175, | |
| "eval_weighted avg": { | |
| "f1-score": 0.8798131143596267, | |
| "precision": 0.8787439244541853, | |
| "recall": 0.8864211105002748, | |
| "support": 36380.0 | |
| }, | |
| "step": 162 | |
| }, | |
| { | |
| "epoch": 3.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6401137980085349, | |
| "precision": 0.6181318681318682, | |
| "recall": 0.6637168141592921, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.7918781725888324, | |
| "precision": 0.6666666666666666, | |
| "recall": 0.975, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8866442199775534, | |
| "precision": 0.93935790725327, | |
| "recall": 0.8395324123273114, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6254390633315594, | |
| "precision": 0.6255056418990845, | |
| "recall": 0.6253724989357173, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.76480605487228, | |
| "precision": 0.6205096714768191, | |
| "recall": 0.9965483234714004, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9049205625504156, | |
| "precision": 0.9449897450922942, | |
| "recall": 0.868111163447951, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9988758992805756, | |
| "precision": 0.9995500562429697, | |
| "recall": 0.9982026510896428, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.8894997251236944, | |
| "eval_loss": 0.318760484457016, | |
| "eval_macro avg": { | |
| "f1-score": 0.8018111100871074, | |
| "precision": 0.7735302223947104, | |
| "recall": 0.852354837633045, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0862, | |
| "eval_samples_per_second": 15.925, | |
| "eval_steps_per_second": 2.163, | |
| "eval_weighted avg": { | |
| "f1-score": 0.8920666591013163, | |
| "precision": 0.9012548868310761, | |
| "recall": 0.8894997251236944, | |
| "support": 36380.0 | |
| }, | |
| "step": 243 | |
| }, | |
| { | |
| "epoch": 4.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.7034883720930232, | |
| "precision": 0.6934097421203438, | |
| "recall": 0.7138643067846607, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.8958990536277602, | |
| "precision": 0.9044585987261147, | |
| "recall": 0.8875, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9051172707889126, | |
| "precision": 0.9080213903743316, | |
| "recall": 0.9022316684378321, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6932699265998139, | |
| "precision": 0.6739698492462312, | |
| "recall": 0.7137079608343976, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8904009720534629, | |
| "precision": 0.877815045519885, | |
| "recall": 0.903353057199211, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9127480293558032, | |
| "precision": 0.9218310342461052, | |
| "recall": 0.9038422717179194, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9981619715668254, | |
| "precision": 0.9999248459341651, | |
| "recall": 0.9964053021792856, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9113523914238593, | |
| "eval_loss": 0.2517726719379425, | |
| "eval_macro avg": { | |
| "f1-score": 0.8570122280122289, | |
| "precision": 0.8542043580238825, | |
| "recall": 0.8601292238790438, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.062, | |
| "eval_samples_per_second": 16.002, | |
| "eval_steps_per_second": 2.173, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9122886837716081, | |
| "precision": 0.9134709768686261, | |
| "recall": 0.9113523914238593, | |
| "support": 36380.0 | |
| }, | |
| "step": 324 | |
| }, | |
| { | |
| "epoch": 5.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6590538336052201, | |
| "precision": 0.7372262773722628, | |
| "recall": 0.5958702064896755, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9078947368421054, | |
| "precision": 0.9583333333333334, | |
| "recall": 0.8625, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9066802651708312, | |
| "precision": 0.8715686274509804, | |
| "recall": 0.944739638682253, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6513589503280225, | |
| "precision": 0.7243355914538823, | |
| "recall": 0.5917411664538101, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8981975120588981, | |
| "precision": 0.9256933542647828, | |
| "recall": 0.8722879684418146, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9158848058378942, | |
| "precision": 0.8877170824123777, | |
| "recall": 0.9458986609245676, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9994384336041331, | |
| "precision": 0.9992513849378649, | |
| "recall": 0.9996255523103422, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9121220450797142, | |
| "eval_loss": 0.2739432156085968, | |
| "eval_macro avg": { | |
| "f1-score": 0.8483583624924435, | |
| "precision": 0.8720179501750691, | |
| "recall": 0.8303804561860663, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0932, | |
| "eval_samples_per_second": 15.904, | |
| "eval_steps_per_second": 2.16, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9087400479269403, | |
| "precision": 0.9081638580455985, | |
| "recall": 0.9121220450797142, | |
| "support": 36380.0 | |
| }, | |
| "step": 405 | |
| }, | |
| { | |
| "epoch": 6.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6675461741424802, | |
| "precision": 0.60381861575179, | |
| "recall": 0.7463126843657817, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.887608069164265, | |
| "precision": 0.8235294117647058, | |
| "recall": 0.9625, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8737316798196167, | |
| "precision": 0.9303721488595438, | |
| "recall": 0.8235919234856536, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6529236868186323, | |
| "precision": 0.6109050445103857, | |
| "recall": 0.7011494252873564, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8689320388349515, | |
| "precision": 0.7863418530351438, | |
| "recall": 0.9709072978303748, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.8937734066010992, | |
| "precision": 0.9312910284463894, | |
| "recall": 0.8591615638247763, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9988009592326139, | |
| "precision": 0.9994750656167979, | |
| "recall": 0.9981277615517112, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.8944749862561847, | |
| "eval_loss": 0.37784042954444885, | |
| "eval_macro avg": { | |
| "f1-score": 0.8347594306590942, | |
| "precision": 0.8122475954263937, | |
| "recall": 0.8659643794779507, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0734, | |
| "eval_samples_per_second": 15.966, | |
| "eval_steps_per_second": 2.168, | |
| "eval_weighted avg": { | |
| "f1-score": 0.8971820344797716, | |
| "precision": 0.9033144340485707, | |
| "recall": 0.8944749862561847, | |
| "support": 36380.0 | |
| }, | |
| "step": 486 | |
| }, | |
| { | |
| "epoch": 6.17, | |
| "grad_norm": 21.778270721435547, | |
| "learning_rate": 1.7530864197530865e-05, | |
| "loss": 0.2481, | |
| "step": 500 | |
| }, | |
| { | |
| "epoch": 7.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6489859594383774, | |
| "precision": 0.6887417218543046, | |
| "recall": 0.6135693215339233, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.8852459016393444, | |
| "precision": 0.9310344827586207, | |
| "recall": 0.84375, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9011898603207449, | |
| "precision": 0.8780241935483871, | |
| "recall": 0.9256110520722636, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.626564114338193, | |
| "precision": 0.6800398704211313, | |
| "recall": 0.5808854831843338, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8813559322033898, | |
| "precision": 0.9049350649350649, | |
| "recall": 0.8589743589743589, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9073595340161655, | |
| "precision": 0.8831698305516626, | |
| "recall": 0.9329116479375547, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9981994148098132, | |
| "precision": 1.0, | |
| "recall": 0.9964053021792856, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9030786146234194, | |
| "eval_loss": 0.4109443426132202, | |
| "eval_macro avg": { | |
| "f1-score": 0.8355572452522898, | |
| "precision": 0.8522778805813102, | |
| "recall": 0.8217295951259599, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0454, | |
| "eval_samples_per_second": 16.054, | |
| "eval_steps_per_second": 2.18, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9003265559019435, | |
| "precision": 0.8992988510674564, | |
| "recall": 0.9030786146234194, | |
| "support": 36380.0 | |
| }, | |
| "step": 567 | |
| }, | |
| { | |
| "epoch": 8.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.7029972752043596, | |
| "precision": 0.6531645569620254, | |
| "recall": 0.7610619469026548, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9134328358208955, | |
| "precision": 0.8742857142857143, | |
| "recall": 0.95625, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8914728682170544, | |
| "precision": 0.930635838150289, | |
| "recall": 0.8554729011689692, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6790487421383647, | |
| "precision": 0.6307046367287331, | |
| "recall": 0.7354193273733504, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8939393939393939, | |
| "precision": 0.836340206185567, | |
| "recall": 0.9600591715976331, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.8993180625983564, | |
| "precision": 0.9362166885102665, | |
| "recall": 0.8652176838705337, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9979819119515658, | |
| "precision": 0.9960462513987318, | |
| "recall": 0.9999251104620684, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.902363936228697, | |
| "eval_loss": 0.4763801097869873, | |
| "eval_macro avg": { | |
| "f1-score": 0.8540272985528559, | |
| "precision": 0.8367705560316182, | |
| "recall": 0.8762008773393156, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0791, | |
| "eval_samples_per_second": 15.948, | |
| "eval_steps_per_second": 2.166, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9048169208096876, | |
| "precision": 0.9101018951943243, | |
| "recall": 0.902363936228697, | |
| "support": 36380.0 | |
| }, | |
| "step": 648 | |
| }, | |
| { | |
| "epoch": 9.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6884779516358464, | |
| "precision": 0.6648351648351648, | |
| "recall": 0.7138643067846607, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.8860759493670887, | |
| "precision": 0.8974358974358975, | |
| "recall": 0.875, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8984416980118215, | |
| "precision": 0.908695652173913, | |
| "recall": 0.8884165781083954, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6685829873309911, | |
| "precision": 0.6687965921192758, | |
| "recall": 0.6683695189442316, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8819011041766683, | |
| "precision": 0.8592142188961647, | |
| "recall": 0.9058185404339251, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9072658065820057, | |
| "precision": 0.9083367057871308, | |
| "recall": 0.9061974295134917, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9971465044679734, | |
| "precision": 0.9998494089300505, | |
| "recall": 0.9944581741930653, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9054700384826828, | |
| "eval_loss": 0.43875375390052795, | |
| "eval_macro avg": { | |
| "f1-score": 0.8468417145103421, | |
| "precision": 0.843880520025371, | |
| "recall": 0.8503035068539672, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0863, | |
| "eval_samples_per_second": 15.925, | |
| "eval_steps_per_second": 2.163, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9056589487317133, | |
| "precision": 0.9059462683069767, | |
| "recall": 0.9054700384826828, | |
| "support": 36380.0 | |
| }, | |
| "step": 729 | |
| }, | |
| { | |
| "epoch": 10.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6910569105691057, | |
| "precision": 0.6390977443609023, | |
| "recall": 0.7522123893805309, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9046153846153846, | |
| "precision": 0.8909090909090909, | |
| "recall": 0.91875, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8888888888888888, | |
| "precision": 0.9213226909920182, | |
| "recall": 0.8586609989373007, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6645316253002403, | |
| "precision": 0.6271250472232717, | |
| "recall": 0.7066836951894423, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.886980737990253, | |
| "precision": 0.8377904427882508, | |
| "recall": 0.9423076923076923, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.8983343700324833, | |
| "precision": 0.9233501456276195, | |
| "recall": 0.8746383150528229, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9975225225225225, | |
| "precision": 1.0, | |
| "recall": 0.9950572904965176, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.8995601979109401, | |
| "eval_loss": 0.5381875038146973, | |
| "eval_macro avg": { | |
| "f1-score": 0.8474186342741253, | |
| "precision": 0.8342278802715933, | |
| "recall": 0.8640443401949011, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.099, | |
| "eval_samples_per_second": 15.885, | |
| "eval_steps_per_second": 2.157, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9017669881051834, | |
| "precision": 0.9056169116577473, | |
| "recall": 0.8995601979109401, | |
| "support": 36380.0 | |
| }, | |
| "step": 810 | |
| }, | |
| { | |
| "epoch": 11.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.7103825136612023, | |
| "precision": 0.6615776081424937, | |
| "recall": 0.7669616519174042, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9032258064516129, | |
| "precision": 0.850828729281768, | |
| "recall": 0.9625, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8955223880597015, | |
| "precision": 0.9331797235023042, | |
| "recall": 0.8607863974495218, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.685932169375878, | |
| "precision": 0.6488230827638573, | |
| "recall": 0.7275436355896125, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8921899617375647, | |
| "precision": 0.820703933747412, | |
| "recall": 0.9773175542406312, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9036754507628294, | |
| "precision": 0.932183990271121, | |
| "recall": 0.8768588924029338, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9971084156370874, | |
| "precision": 1.0, | |
| "recall": 0.9942335055792706, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9051951621770203, | |
| "eval_loss": 0.5869538187980652, | |
| "eval_macro avg": { | |
| "f1-score": 0.8554338150979823, | |
| "precision": 0.8353281525298509, | |
| "recall": 0.880885948168482, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0923, | |
| "eval_samples_per_second": 15.906, | |
| "eval_steps_per_second": 2.16, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9071963301332511, | |
| "precision": 0.9114149044954519, | |
| "recall": 0.9051951621770203, | |
| "support": 36380.0 | |
| }, | |
| "step": 891 | |
| }, | |
| { | |
| "epoch": 12.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6758409785932722, | |
| "precision": 0.7015873015873015, | |
| "recall": 0.6519174041297935, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.916923076923077, | |
| "precision": 0.9030303030303031, | |
| "recall": 0.93125, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8989473684210527, | |
| "precision": 0.8905109489051095, | |
| "recall": 0.9075451647183846, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6533197602081212, | |
| "precision": 0.6970794110547912, | |
| "recall": 0.6147296722009365, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9200480192076831, | |
| "precision": 0.8965839962564343, | |
| "recall": 0.9447731755424064, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9082653940478558, | |
| "precision": 0.8948605759811924, | |
| "recall": 0.922077922077922, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9988389947942025, | |
| "precision": 0.9990260713215463, | |
| "recall": 0.9986519883172321, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9089059923034635, | |
| "eval_loss": 0.5580677390098572, | |
| "eval_macro avg": { | |
| "f1-score": 0.8531690845993235, | |
| "precision": 0.854668372590954, | |
| "recall": 0.852992189569525, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0574, | |
| "eval_samples_per_second": 16.016, | |
| "eval_steps_per_second": 2.175, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9068749310769137, | |
| "precision": 0.9057713940131727, | |
| "recall": 0.9089059923034635, | |
| "support": 36380.0 | |
| }, | |
| "step": 972 | |
| }, | |
| { | |
| "epoch": 12.35, | |
| "grad_norm": 0.45772114396095276, | |
| "learning_rate": 1.506172839506173e-05, | |
| "loss": 0.0314, | |
| "step": 1000 | |
| }, | |
| { | |
| "epoch": 13.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6760563380281691, | |
| "precision": 0.72, | |
| "recall": 0.6371681415929203, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9171974522292993, | |
| "precision": 0.935064935064935, | |
| "recall": 0.9, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.905503634475597, | |
| "precision": 0.8852791878172589, | |
| "recall": 0.926673751328374, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6522588803310726, | |
| "precision": 0.7090727318170458, | |
| "recall": 0.6038739889314602, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9228121927236971, | |
| "precision": 0.9200980392156862, | |
| "recall": 0.9255424063116371, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9102441910145594, | |
| "precision": 0.8896312475909033, | |
| "recall": 0.9318350043738645, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9992880428673137, | |
| "precision": 1.0, | |
| "recall": 0.9985770987793006, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9106102253985706, | |
| "eval_loss": 0.5264647603034973, | |
| "eval_macro avg": { | |
| "f1-score": 0.8547658188099584, | |
| "precision": 0.8655923059294041, | |
| "recall": 0.8462386273310795, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0562, | |
| "eval_samples_per_second": 16.02, | |
| "eval_steps_per_second": 2.176, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9080379636257123, | |
| "precision": 0.9070294280758862, | |
| "recall": 0.9106102253985706, | |
| "support": 36380.0 | |
| }, | |
| "step": 1053 | |
| }, | |
| { | |
| "epoch": 14.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6980609418282548, | |
| "precision": 0.6579634464751958, | |
| "recall": 0.7433628318584071, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9226006191950465, | |
| "precision": 0.9141104294478528, | |
| "recall": 0.93125, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.8921568627450981, | |
| "precision": 0.9150837988826815, | |
| "recall": 0.8703506907545164, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6821321785751109, | |
| "precision": 0.6614677064587082, | |
| "recall": 0.7041294167730949, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9124910093502757, | |
| "precision": 0.8880074661689221, | |
| "recall": 0.9383629191321499, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9052323204796293, | |
| "precision": 0.9166609175577786, | |
| "recall": 0.894085189421977, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9980114808839531, | |
| "precision": 1.0, | |
| "recall": 0.9960308544896278, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9075865860362837, | |
| "eval_loss": 0.6585939526557922, | |
| "eval_macro avg": { | |
| "f1-score": 0.8586693447224812, | |
| "precision": 0.850470537855877, | |
| "recall": 0.8682245574899676, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0606, | |
| "eval_samples_per_second": 16.006, | |
| "eval_steps_per_second": 2.174, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9086880683615599, | |
| "precision": 0.9102351141829325, | |
| "recall": 0.9075865860362837, | |
| "support": 36380.0 | |
| }, | |
| "step": 1134 | |
| }, | |
| { | |
| "epoch": 15.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6920731707317073, | |
| "precision": 0.7160883280757098, | |
| "recall": 0.6696165191740413, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.90282131661442, | |
| "precision": 0.9056603773584906, | |
| "recall": 0.9, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9076600209863588, | |
| "precision": 0.8963730569948186, | |
| "recall": 0.9192348565356004, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6644877718922575, | |
| "precision": 0.7061077844311378, | |
| "recall": 0.6275010642826735, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9016511127063891, | |
| "precision": 0.8758716875871687, | |
| "recall": 0.9289940828402367, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9134082844161443, | |
| "precision": 0.9005362280931206, | |
| "recall": 0.9266536572236054, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9987253505286046, | |
| "precision": 0.9999249305607687, | |
| "recall": 0.9975286452482588, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9114623419461243, | |
| "eval_loss": 0.6387954950332642, | |
| "eval_macro avg": { | |
| "f1-score": 0.854403861125126, | |
| "precision": 0.8572231990144593, | |
| "recall": 0.8527898321863451, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0835, | |
| "eval_samples_per_second": 15.934, | |
| "eval_steps_per_second": 2.164, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9096653222792656, | |
| "precision": 0.9087293921765042, | |
| "recall": 0.9114623419461243, | |
| "support": 36380.0 | |
| }, | |
| "step": 1215 | |
| }, | |
| { | |
| "epoch": 16.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6602254428341385, | |
| "precision": 0.7269503546099291, | |
| "recall": 0.6047197640117994, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.8903225806451613, | |
| "precision": 0.92, | |
| "recall": 0.8625, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9055441478439425, | |
| "precision": 0.8758689175769613, | |
| "recall": 0.9373007438894793, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6361776387050188, | |
| "precision": 0.7319302132373304, | |
| "recall": 0.5625798212005109, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9097614949594295, | |
| "precision": 0.9073075036782736, | |
| "recall": 0.9122287968441815, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9127908863717504, | |
| "precision": 0.8816300940438871, | |
| "recall": 0.9462351120382209, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9993256911665542, | |
| "precision": 0.9997751293006522, | |
| "recall": 0.9988766569310268, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9103353490929081, | |
| "eval_loss": 0.6975058317184448, | |
| "eval_macro avg": { | |
| "f1-score": 0.844878268932285, | |
| "precision": 0.8633517446352906, | |
| "recall": 0.8320629849878884, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0731, | |
| "eval_samples_per_second": 15.967, | |
| "eval_steps_per_second": 2.168, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9060232340867683, | |
| "precision": 0.9056723119856793, | |
| "recall": 0.9103353490929081, | |
| "support": 36380.0 | |
| }, | |
| "step": 1296 | |
| }, | |
| { | |
| "epoch": 17.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6964285714285714, | |
| "precision": 0.7027027027027027, | |
| "recall": 0.6902654867256637, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.903954802259887, | |
| "precision": 0.8247422680412371, | |
| "recall": 1.0, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9045822102425876, | |
| "precision": 0.9179431072210066, | |
| "recall": 0.8916046758767269, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6845814977973568, | |
| "precision": 0.7092651757188498, | |
| "recall": 0.6615581098339719, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.8985378821444396, | |
| "precision": 0.8157683024939663, | |
| "recall": 1.0, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9151271572832584, | |
| "precision": 0.9186330349877949, | |
| "recall": 0.9116479375546733, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9988753936122357, | |
| "precision": 1.0, | |
| "recall": 0.9977533138620535, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9136888400219901, | |
| "eval_loss": 0.692401111125946, | |
| "eval_macro avg": { | |
| "f1-score": 0.8574410735383337, | |
| "precision": 0.8412935130236511, | |
| "recall": 0.878975646264727, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0854, | |
| "eval_samples_per_second": 15.928, | |
| "eval_steps_per_second": 2.163, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9128097974394099, | |
| "precision": 0.9132839716585025, | |
| "recall": 0.9136888400219901, | |
| "support": 36380.0 | |
| }, | |
| "step": 1377 | |
| }, | |
| { | |
| "epoch": 18.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6927899686520376, | |
| "precision": 0.7391304347826086, | |
| "recall": 0.6519174041297935, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9192546583850932, | |
| "precision": 0.9135802469135802, | |
| "recall": 0.925, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9089016137428423, | |
| "precision": 0.8908163265306123, | |
| "recall": 0.9277364505844846, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6788553259141495, | |
| "precision": 0.7276046738072055, | |
| "recall": 0.6362281822051937, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9119420989143546, | |
| "precision": 0.8927727916863486, | |
| "recall": 0.9319526627218935, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.9172837873578418, | |
| "precision": 0.9016052511860662, | |
| "recall": 0.9335172599421304, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9990254872563717, | |
| "precision": 1.0, | |
| "recall": 0.9980528720137797, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.915915338097856, | |
| "eval_loss": 0.6629185676574707, | |
| "eval_macro avg": { | |
| "f1-score": 0.861150420031813, | |
| "precision": 0.866501389272346, | |
| "recall": 0.8577721187996108, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0772, | |
| "eval_samples_per_second": 15.954, | |
| "eval_steps_per_second": 2.167, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9138987233713269, | |
| "precision": 0.9130176425817779, | |
| "recall": 0.915915338097856, | |
| "support": 36380.0 | |
| }, | |
| "step": 1458 | |
| }, | |
| { | |
| "epoch": 18.52, | |
| "grad_norm": 34.05133056640625, | |
| "learning_rate": 1.2592592592592593e-05, | |
| "loss": 0.0105, | |
| "step": 1500 | |
| }, | |
| { | |
| "epoch": 19.0, | |
| "eval_B-Claim": { | |
| "f1-score": 0.6967340590979781, | |
| "precision": 0.7368421052631579, | |
| "recall": 0.6607669616519174, | |
| "support": 339.0 | |
| }, | |
| "eval_B-MajorClaim": { | |
| "f1-score": 0.9244712990936558, | |
| "precision": 0.8947368421052632, | |
| "recall": 0.95625, | |
| "support": 160.0 | |
| }, | |
| "eval_B-Premise": { | |
| "f1-score": 0.9071840587309911, | |
| "precision": 0.8954451345755694, | |
| "recall": 0.9192348565356004, | |
| "support": 941.0 | |
| }, | |
| "eval_I-Claim": { | |
| "f1-score": 0.6769585253456222, | |
| "precision": 0.7378201908588649, | |
| "recall": 0.6253724989357173, | |
| "support": 4698.0 | |
| }, | |
| "eval_I-MajorClaim": { | |
| "f1-score": 0.9125809435707678, | |
| "precision": 0.8593205574912892, | |
| "recall": 0.972879684418146, | |
| "support": 2028.0 | |
| }, | |
| "eval_I-Premise": { | |
| "f1-score": 0.916641828117238, | |
| "precision": 0.9025042389461327, | |
| "recall": 0.9312293923692887, | |
| "support": 14861.0 | |
| }, | |
| "eval_O": { | |
| "f1-score": 0.9990254872563717, | |
| "precision": 1.0, | |
| "recall": 0.9980528720137797, | |
| "support": 13353.0 | |
| }, | |
| "eval_accuracy": 0.9158603628367235, | |
| "eval_loss": 0.69818514585495, | |
| "eval_macro avg": { | |
| "f1-score": 0.8619423144589463, | |
| "precision": 0.8609527241771824, | |
| "recall": 0.8662551808463499, | |
| "support": 36380.0 | |
| }, | |
| "eval_runtime": 5.0766, | |
| "eval_samples_per_second": 15.956, | |
| "eval_steps_per_second": 2.167, | |
| "eval_weighted avg": { | |
| "f1-score": 0.9134424215990907, | |
| "precision": 0.912854812597098, | |
| "recall": 0.9158603628367235, | |
| "support": 36380.0 | |
| }, | |
| "step": 1539 | |
| } | |
| ], | |
| "logging_steps": 500, | |
| "max_steps": 4050, | |
| "num_input_tokens_seen": 0, | |
| "num_train_epochs": 50, | |
| "save_steps": 500, | |
| "total_flos": 2723540890581000.0, | |
| "train_batch_size": 4, | |
| "trial_name": null, | |
| "trial_params": null | |
| } | |