{ "encoder": "jhu-clsp/mmBERT-base", "head_layers": 2, "max_len": 2048, "head_max_len": 256, "max_prefixes": 6, "act_costs": { "escalate": 0.5 }, "cost_wrong_act": 3.0, "amp_dtype": "bf16", "model_name": "rl-agent", "temperature": [ 3.7832961082458496, 1.0, 3.33156156539917 ], "training": { "updates": 15987, "epochs_completed": 4, "hours": 4.97, "world_size": 1, "fine_tuned_from_checkpoint": false, "laya_train": { "epochs": 5, "micro_batch": 2, "grad_accum": 8, "encoder_lr": 2.5e-05, "head_lr": 0.0001, "min_lr": 1e-06, "weight_decay": 0.01, "grad_clip": 1.0, "loss": "rlcd", "rl_samples": 4, "sigma_start": 0.4, "sigma_end": 0.1, "w_sph": 0.75, "w_rps": 1.0, "shuffle_options": [ "choice" ], "option_layout": null, "freeze_encoder": false, "calib_max": 400, "calib_frac": 0.1, "calib_seed": 20260922, "seed": 0, "max_len": 2048, "head_max_len": null, "amp": null, "gradient_checkpointing": null, "log_every": 100, "label_smoothing": 0.0, "text_column": "text", "label_column": "label", "question_id": "label", "instructions": null, "eval_data": "data\\sets\\val.jsonl", "target_error": 0.1, "min_abstain_n": 10 }, "laya_train_calibration": { "choice": { "items": 139, "temperature": 3.7832961082458496, "issues": [], "accuracy": 0.9712230215827338, "mean_confidence": 0.9759824520505122 }, "score": { "items": 44, "temperature": 1.0, "issues": [ "fitted on only 44 calibration items", "the fit came back unchanged at 1.0 (accuracy 1.00, mean confidence 1.000 on the calibration items), which usually means they gave it nothing to correct; this type's confidences are uncalibrated" ], "accuracy": 1.0, "mean_confidence": 0.999999929841763 }, "noul": { "items": 217, "temperature": 3.33156156539917, "issues": [], "accuracy": 0.9953917050691244, "mean_confidence": 0.9961188719812915 } }, "train_report": { "eval_source": "data\\sets\\val.jsonl", "eval_mode": "overlapping_eval", "is_held_out": false, "eval_items": 3118, "note": "Evaluation set contains rows that overlap training data (8/3118 items); metrics do not reflect strictly independent held-out evaluation.", "before": { "items": 3118, "loss": 2.4687, "accuracy": 0.4641, "mean_confidence": 0.8268, "ece": 0.3628, "brier": 0.8545, "brier_top1": 0.3896, "by_type": { "choice": { "items": 1164, "loss": 1.7198, "accuracy": 0.5112, "mean_confidence": 0.7506, "ece": 0.2394, "brier": 0.7207, "brier_top1": 0.2891 }, "score": { "items": 374, "loss": 2.0527, "accuracy": 0.2941, "mean_confidence": 0.7141, "ece": 0.42, "brier": 0.9597, "brier_top1": 0.3878 }, "noul": { "items": 1580, "loss": 3.1189, "accuracy": 0.4696, "mean_confidence": 0.9097, "ece": 0.4401, "brier": 0.9281, "brier_top1": 0.464 } } }, "after": { "items": 3118, "loss": 0.2564, "accuracy": 0.9622, "mean_confidence": 0.9868, "ece": 0.0263, "brier": 0.0724, "brier_top1": 0.0355, "by_type": { "choice": { "items": 1164, "loss": 0.2698, "accuracy": 0.9424, "mean_confidence": 0.972, "ece": 0.0309, "brier": 0.1056, "brier_top1": 0.051 }, "score": { "items": 374, "loss": 0.8551, "accuracy": 0.9465, "mean_confidence": 0.9953, "ece": 0.0581, "brier": 0.1106, "brier_top1": 0.0553 }, "noul": { "items": 1580, "loss": 0.1047, "accuracy": 0.9804, "mean_confidence": 0.9958, "ece": 0.016, "brier": 0.0388, "brier_top1": 0.0194 } } }, "comparison": { "delta_accuracy": 0.4981, "delta_loss": -2.2123, "delta_ece": -0.3365, "delta_brier": -0.7821, "delta_brier_top1": -0.3541, "delta_mean_confidence": 0.16 }, "calibration": { "choice": { "items": 139, "temperature": 3.7832961082458496, "issues": [], "accuracy": 0.9712230215827338, "mean_confidence": 0.9759824520505122 }, "score": { "items": 44, "temperature": 1.0, "issues": [ "fitted on only 44 calibration items", "the fit came back unchanged at 1.0 (accuracy 1.00, mean confidence 1.000 on the calibration items), which usually means they gave it nothing to correct; this type's confidences are uncalibrated" ], "accuracy": 1.0, "mean_confidence": 0.999999929841763 }, "noul": { "items": 217, "temperature": 3.33156156539917, "issues": [], "accuracy": 0.9953917050691244, "mean_confidence": 0.9961188719812915 } }, "abstention_thresholds": { "score:3-5": 1.0, "noul:2": 0.9763, "choice:3-5": 0.9681, "choice:2": 0.9901, "choice:6-10": 0.8453 }, "training": { "epochs": 5, "micro_batch": 2, "grad_accum": 8, "encoder_lr": 2.5e-05, "head_lr": 0.0001, "min_lr": 1e-06, "weight_decay": 0.01, "grad_clip": 1.0, "loss": "rlcd", "rl_samples": 4, "sigma_start": 0.4, "sigma_end": 0.1, "w_sph": 0.75, "w_rps": 1.0, "shuffle_options": [ "choice" ], "option_layout": null, "freeze_encoder": false, "calib_max": 400, "calib_frac": 0.1, "calib_seed": 20260922, "seed": 0, "max_len": 2048, "head_max_len": null, "amp": null, "gradient_checkpointing": null, "log_every": 100, "label_smoothing": 0.0, "text_column": "text", "label_column": "label", "question_id": "label", "instructions": null, "eval_data": "data\\sets\\val.jsonl", "target_error": 0.1, "min_abstain_n": 10 } }, "abstention_thresholds": { "score:3-5": 1.0, "noul:2": 0.9763, "choice:3-5": 0.9681, "choice:2": 0.9901, "choice:6-10": 0.8453 } }, "fine_tuned": true }