{
  "encoder": "hfl/chinese-roberta-wwm-ext-large",
  "encoder_type": "bert",
  "head_layers": 2,
  "max_len": 512,
  "head_max_len": 192,
  "max_prefixes": 6,
  "act_costs": {
    "escalate": 0.5
  },
  "cost_wrong_act": 3.0,
  "amp_dtype": "bf16",
  "train_dtype": "fp32",
  "temperature": [
    0.467223,
    1.034508,
    0.970277
  ],
  "temperature_by_options": {
    "choice:11+": 0.466635,
    "choice:2": 0.253066,
    "choice:3-5": 0.545042,
    "noul:2": 0.970277,
    "score:3-5": 1.034508
  },
  "model_name": "layazh",
  "training": {
    "recipe": {
      "head_lr": 0.0003,
      "encoder_lr": 2e-05,
      "warmup_frac": 0.06,
      "scheduler": "linear warmup 6% + cosine",
      "unfrozen_encoder_layers": 8,
      "option_order_augmentation": true,
      "grad_checkpointing": true,
      "amp": "fp32 (sm_52/Maxwell has no bf16)",
      "batch_tokens": 4096,
      "optimizer": "AdamW",
      "grad_clip": 1.0
    },
    "effective_steps": "~1250 (stopped near plateau; not the full planned schedule)",
    "probe_best": 0.7991,
    "probe_acc": {
      "choice": 0.7243,
      "score": 0.7744,
      "noul": 0.9638,
      "ALL": 0.7991
    },
    "hardware": "GTX Titan X (Maxwell, sm_52, 12GB), torch 2.7.1+cu128"
  },
  "calibration": {
    "scheme": "bucket",
    "report_ece": 0.2546653607969531,
    "fit_n": 3122,
    "report_n": 3122,
    "by_qtype": [
      0.467223,
      1.034508,
      0.970277
    ],
    "by_bucket": {
      "choice:11+": 0.466635,
      "choice:2": 0.253066,
      "choice:3-5": 0.545042,
      "noul:2": 0.970277,
      "score:3-5": 1.034508
    },
    "all_schemes_report_ece": {
      "identity": 0.3511506951291145,
      "qtype": 0.26843774177473434,
      "bucket": 0.2546653607969531
    }
  },
  "evaluation": {
    "full_eval_n": 6244,
    "full_eval_acc": 0.6951,
    "baselines": {
      "random": 0.2959,
      "majority_class": 0.4734
    },
    "upstream_laya_multilingual_acc": 0.4783,
    "ab_n": 1200
  }
}