{
  "status": "COMPLETE",
  "experiment_completed": "2026-11-02 20:58:58 +08:00",
  "training_rounds_completed": 21,
  "training_rounds_total": 21,
  "main_tests_completed": 720,
  "auxiliary_tests_completed": 180,
  "total_completed_answers": 900,
  "gpu_stopped": "2026-11-02 21:01:28 +08:00",
  "models": [
    {
      "id": "N",
      "name": "原模型",
      "sampled": 27,
      "fixed": 15,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "WT",
      "name": "Wide",
      "sampled": 30,
      "fixed": 14,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "DT",
      "name": "Deep",
      "sampled": 30,
      "fixed": 14,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "WDT",
      "name": "原 Wide＋Deep",
      "sampled": 27,
      "fixed": 15,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "R",
      "name": "同题量普通训练",
      "sampled": 28,
      "fixed": 14,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "WDT_OPT",
      "name": "改进 Wide＋Deep",
      "sampled": 30,
      "fixed": 14,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "WDT_REPAIR",
      "name": "改进 Wide＋Deep＋纠错",
      "sampled": 34,
      "fixed": 15,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "WDT_SCALE50",
      "name": "50 道种子版",
      "sampled": 29,
      "fixed": 13,
      "sampled_total": 60,
      "fixed_total": 30
    }
  ],
  "auxiliary": [
    {
      "id": "one",
      "name": "学 1 次",
      "sampled": 28,
      "fixed": 15,
      "sampled_total": 60,
      "fixed_total": 30
    },
    {
      "id": "four",
      "name": "学 4 次",
      "sampled": 29,
      "fixed": 16,
      "sampled_total": 60,
      "fixed_total": 30
    }
  ],
  "test_scope": {
    "questions": 30,
    "sampled_answers_per_question": 2,
    "fixed_answers_per_question": 1,
    "main_models": 8,
    "main_answers": 720,
    "auxiliary_answers": 180,
    "deferred_questions": 20,
    "sampling_temperature": 1.0,
    "max_steps": 24,
    "difficulty_counts": {
      "1": 10,
      "2": 11,
      "3": 8,
      "4": 1
    },
    "sampling_score": "全部随机回答的平均成功率，不取每题两次中的最好成绩"
  },
  "analysis": "analysis-20261103.html",
  "new_experiment_started": false
}