{
  "eval_source": "/kaggle/input/datasets/anirudhbharadwajpr/laya-router-training-data/eval.jsonl",
  "eval_mode": "held_out",
  "is_held_out": true,
  "eval_items": 39,
  "note": "Metrics reflect independent held-out evaluation.",
  "before": {
    "items": 39,
    "loss": 1.0705,
    "accuracy": 0.7436,
    "mean_confidence": 0.3639,
    "ece": 0.3797,
    "brier": 0.5708,
    "brier_top1": 0.3058,
    "by_type": {
      "choice": {
        "items": 39,
        "loss": 1.0705,
        "accuracy": 0.7436,
        "mean_confidence": 0.3639,
        "ece": 0.3797,
        "brier": 0.5708,
        "brier_top1": 0.3058
      }
    }
  },
  "after": {
    "items": 39,
    "loss": 0.6364,
    "accuracy": 0.9744,
    "mean_confidence": 0.9998,
    "ece": 0.0255,
    "brier": 0.0513,
    "brier_top1": 0.0256,
    "by_type": {
      "choice": {
        "items": 39,
        "loss": 0.6364,
        "accuracy": 0.9744,
        "mean_confidence": 0.9998,
        "ece": 0.0255,
        "brier": 0.0513,
        "brier_top1": 0.0256
      }
    }
  },
  "comparison": {
    "delta_accuracy": 0.2308,
    "delta_loss": -0.4341,
    "delta_ece": -0.3542,
    "delta_brier": -0.5195,
    "delta_brier_top1": -0.2802,
    "delta_mean_confidence": 0.6359
  },
  "calibration": {
    "choice": {
      "items": 16,
      "temperature": 1.0,
      "issues": [
        "fitted on only 16 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": 0.999999883761271
    },
    "score": {
      "items": 0,
      "temperature": 1.0,
      "issues": []
    },
    "noul": {
      "items": 0,
      "temperature": 1.0,
      "issues": []
    }
  },
  "abstention_thresholds": {
    "choice:3-5": 1.0
  },
  "training": {
    "epochs": 4,
    "micro_batch": 4,
    "grad_accum": 4,
    "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": 42,
    "max_len": null,
    "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": "/kaggle/input/datasets/anirudhbharadwajpr/laya-router-training-data/eval.jsonl",
    "target_error": 0.1,
    "min_abstain_n": 10
  }
}