{ "best_metric": null, "best_model_checkpoint": null, "epoch": 3.0, "eval_steps": 500, "global_step": 243, "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": 0.32421052631578945, "precision": 0.42777777777777776, "recall": 0.26101694915254237, "support": 295.0 }, "eval_B-MajorClaim": { "f1-score": 0.192090395480226, "precision": 0.8095238095238095, "recall": 0.10897435897435898, "support": 156.0 }, "eval_B-Premise": { "f1-score": 0.8566265060240965, "precision": 0.7620578778135049, "recall": 0.9779917469050894, "support": 727.0 }, "eval_I-Claim": { "f1-score": 0.5001390820584144, "precision": 0.5900886117492615, "recall": 0.43398503499879315, "support": 4143.0 }, "eval_I-MajorClaim": { "f1-score": 0.776470588235294, "precision": 0.6649333766655834, "recall": 0.93296853625171, "support": 2193.0 }, "eval_I-Premise": { "f1-score": 0.8861458374413376, "precision": 0.8781460059402845, "recall": 0.8942927644670858, "support": 12563.0 }, "eval_O": { "f1-score": 0.9987647926046939, "precision": 0.9977708781147998, "recall": 0.9997606892150607, "support": 12536.0 }, "eval_accuracy": 0.8713396498328887, "eval_loss": 0.3092103600502014, "eval_macro avg": { "f1-score": 0.647778246879979, "precision": 0.732899762512146, "recall": 0.6584271542806628, "support": 32613.0 }, "eval_runtime": 4.965, "eval_samples_per_second": 16.314, "eval_steps_per_second": 2.215, "eval_weighted avg": { "f1-score": 0.8639627599323451, "precision": 0.8662078389010613, "recall": 0.8713396498328887, "support": 32613.0 }, "step": 81 }, { "epoch": 2.0, "eval_B-Claim": { "f1-score": 0.6186291739894552, "precision": 0.6423357664233577, "recall": 0.5966101694915255, "support": 295.0 }, "eval_B-MajorClaim": { "f1-score": 0.8113879003558718, "precision": 0.912, "recall": 0.7307692307692307, "support": 156.0 }, "eval_B-Premise": { "f1-score": 0.8871181938911022, "precision": 0.8575096277278562, "recall": 0.9188445667125172, "support": 727.0 }, "eval_I-Claim": { "f1-score": 0.656253149249219, "precision": 0.5633217993079584, "recall": 0.7859039343470915, "support": 4143.0 }, "eval_I-MajorClaim": { "f1-score": 0.8368159203980099, "precision": 0.9206349206349206, "recall": 0.7669858641130871, "support": 2193.0 }, "eval_I-Premise": { "f1-score": 0.8831027253668763, "precision": 0.933020288827855, "recall": 0.8382551938231314, "support": 12563.0 }, "eval_O": { "f1-score": 0.9998006141085457, "precision": 0.9996013077107089, "recall": 1.0, "support": 12536.0 }, "eval_accuracy": 0.8880814399165977, "eval_loss": 0.2581149935722351, "eval_macro avg": { "f1-score": 0.8133010967655829, "precision": 0.8326319586618081, "recall": 0.8053384227509405, "support": 32613.0 }, "eval_runtime": 5.0296, "eval_samples_per_second": 16.105, "eval_steps_per_second": 2.187, "eval_weighted avg": { "f1-score": 0.8933836566327042, "precision": 0.9064023864086126, "recall": 0.8880814399165977, "support": 32613.0 }, "step": 162 }, { "epoch": 3.0, "eval_B-Claim": { "f1-score": 0.6576271186440678, "precision": 0.6576271186440678, "recall": 0.6576271186440678, "support": 295.0 }, "eval_B-MajorClaim": { "f1-score": 0.8673139158576051, "precision": 0.8758169934640523, "recall": 0.8589743589743589, "support": 156.0 }, "eval_B-Premise": { "f1-score": 0.8838487972508591, "precision": 0.8832417582417582, "recall": 0.8844566712517193, "support": 727.0 }, "eval_I-Claim": { "f1-score": 0.6468460321049062, "precision": 0.6083351052519669, "recall": 0.6905623944001931, "support": 4143.0 }, "eval_I-MajorClaim": { "f1-score": 0.8695855771482088, "precision": 0.8936477382098171, "recall": 0.8467852257181943, "support": 2193.0 }, "eval_I-Premise": { "f1-score": 0.8926256077795786, "precision": 0.9090533960551291, "recall": 0.8767810236408501, "support": 12563.0 }, "eval_O": { "f1-score": 0.9998803589232302, "precision": 0.9997607464710104, "recall": 1.0, "support": 12536.0 }, "eval_accuracy": 0.8965749854352559, "eval_loss": 0.2399493306875229, "eval_macro avg": { "f1-score": 0.8311039153869222, "precision": 0.8324975509054003, "recall": 0.8307409703756262, "support": 32613.0 }, "eval_runtime": 5.0024, "eval_samples_per_second": 16.192, "eval_steps_per_second": 2.199, "eval_weighted avg": { "f1-score": 0.8986385496019461, "precision": 0.9016737059631544, "recall": 0.8965749854352559, "support": 32613.0 }, "step": 243 } ], "logging_steps": 500, "max_steps": 4050, "num_input_tokens_seen": 0, "num_train_epochs": 50, "save_steps": 500, "total_flos": 430032772197000.0, "train_batch_size": 4, "trial_name": null, "trial_params": null }