{ "best_metric": null, "best_model_checkpoint": null, "epoch": 4.0, "eval_steps": 500, "global_step": 324, "is_hyper_param_search": false, "is_local_process_zero": true, "is_world_process_zero": true, "log_history": [ { "epoch": 1.0, "eval_Claim": { "f1-score": 0.5385059461955894, "precision": 0.6267132491265789, "recall": 0.4720647773279352, "support": 4940.0 }, "eval_MajorClaim": { "f1-score": 0.7814784727863526, "precision": 0.9588039867109635, "recall": 0.6595063985374772, "support": 2188.0 }, "eval_O": { "f1-score": 0.9938608715302625, "precision": 0.9996969237763298, "recall": 0.9880925634688834, "support": 13353.0 }, "eval_Premise": { "f1-score": 0.8968837690149166, "precision": 0.8455112497215416, "recall": 0.954902824076986, "support": 15899.0 }, "eval_accuracy": 0.8837548103353491, "eval_loss": 0.2860471308231354, "eval_macro avg": { "f1-score": 0.8026822648817803, "precision": 0.8576813523338533, "recall": 0.7686416408528204, "support": 36380.0 }, "eval_runtime": 4.9572, "eval_samples_per_second": 16.34, "eval_steps_per_second": 2.219, "eval_weighted avg": { "f1-score": 0.8768739289107891, "precision": 0.8792073379911217, "recall": 0.8837548103353491, "support": 36380.0 }, "step": 81 }, { "epoch": 2.0, "eval_Claim": { "f1-score": 0.6469459868753155, "precision": 0.6453172205438067, "recall": 0.648582995951417, "support": 4940.0 }, "eval_MajorClaim": { "f1-score": 0.8435970110796187, "precision": 0.9669226225634967, "recall": 0.7481718464351006, "support": 2188.0 }, "eval_O": { "f1-score": 0.998877245508982, "precision": 0.9983541557567143, "recall": 0.9994008836965476, "support": 13353.0 }, "eval_Premise": { "f1-score": 0.9099026477336144, "precision": 0.8972179761540813, "recall": 0.922951129001824, "support": 15899.0 }, "eval_accuracy": 0.903243540406817, "eval_loss": 0.24885737895965576, "eval_macro avg": { "f1-score": 0.8498307227993828, "precision": 0.8769529937545246, "recall": 0.8297767137712222, "support": 36380.0 }, "eval_runtime": 5.0097, "eval_samples_per_second": 16.169, "eval_steps_per_second": 2.196, "eval_weighted avg": { "f1-score": 0.9028656814459712, "precision": 0.9043261520711511, "recall": 0.903243540406817, "support": 36380.0 }, "step": 162 }, { "epoch": 3.0, "eval_Claim": { "f1-score": 0.7040630102775911, "precision": 0.6698958142935478, "recall": 0.7419028340080972, "support": 4940.0 }, "eval_MajorClaim": { "f1-score": 0.8888411094388303, "precision": 0.8392204628501827, "recall": 0.9446983546617916, "support": 2188.0 }, "eval_O": { "f1-score": 0.998989104796136, "precision": 0.9988769092542678, "recall": 0.9991013255448213, "support": 13353.0 }, "eval_Premise": { "f1-score": 0.914582593823615, "precision": 0.9390987408880053, "recall": 0.8913139191144097, "support": 15899.0 }, "eval_accuracy": 0.9137987905442551, "eval_loss": 0.24298855662345886, "eval_macro avg": { "f1-score": 0.8766189545840432, "precision": 0.8617729818215009, "recall": 0.8942541083322799, "support": 36380.0 }, "eval_runtime": 4.9876, "eval_samples_per_second": 16.24, "eval_steps_per_second": 2.205, "eval_weighted avg": { "f1-score": 0.9154289662937856, "precision": 0.9184781183611038, "recall": 0.9137987905442551, "support": 36380.0 }, "step": 243 }, { "epoch": 4.0, "eval_Claim": { "f1-score": 0.6981611893583725, "precision": 0.6754352763058289, "recall": 0.7224696356275304, "support": 4940.0 }, "eval_MajorClaim": { "f1-score": 0.901231802911534, "precision": 0.8836187966622749, "recall": 0.9195612431444241, "support": 2188.0 }, "eval_O": { "f1-score": 0.9989130842172332, "precision": 0.9998499399759904, "recall": 0.9979779824758481, "support": 13353.0 }, "eval_Premise": { "f1-score": 0.9146224912392481, "precision": 0.9266670970240785, "recall": 0.9028869740235235, "support": 15899.0 }, "eval_accuracy": 0.9142935678944475, "eval_loss": 0.23980534076690674, "eval_macro avg": { "f1-score": 0.8782321419315969, "precision": 0.8713927774920431, "recall": 0.8857239588178316, "support": 36380.0 }, "eval_runtime": 4.9726, "eval_samples_per_second": 16.289, "eval_steps_per_second": 2.212, "eval_weighted avg": { "f1-score": 0.9153623106642749, "precision": 0.9168247558035478, "recall": 0.9142935678944475, "support": 36380.0 }, "step": 324 } ], "logging_steps": 500, "max_steps": 4050, "num_input_tokens_seen": 0, "num_train_epochs": 50, "save_steps": 500, "total_flos": 573368735469600.0, "train_batch_size": 4, "trial_name": null, "trial_params": null }