{
  "protocol": {
    "dataset": "OpenAI GSM8K official train/test JSONL",
    "task": "six-candidate policy selection; not free-form generation",
    "train_examples": 512,
    "validation_examples": 128,
    "steps": 300,
    "group_size": 4,
    "learning_rate": 0.08,
    "seed": 42
  },
  "baseline": {
    "accuracy": 0.1640625,
    "mean_reward": 0.31261065795890286
  },
  "results": {
    "dpo": {
      "accuracy": 0.8046875,
      "mean_reward": 0.8346500627127349,
      "kl_from_reference": 0.06833317338028673,
      "reward_model_parameters": 0
    },
    "kto": {
      "accuracy": 0.8359375,
      "mean_reward": 0.855992353047812,
      "kl_from_reference": 0.014338344537671776,
      "desirable_utility": 0.5029538721680035,
      "undesirable_utility": 0.49721555455573463,
      "reference_kl_ema": 0.0074314044704069495
    },
    "orpo": {
      "accuracy": 0.84375,
      "mean_reward": 0.8617960843780659,
      "kl_from_reference": 0.8973084851689165,
      "sft_nll": 0.07153727844510231,
      "log_odds_margin": 6.350187804446428,
      "reference_model_parameters": 0
    },
    "grpo": {
      "accuracy": 0.78125,
      "mean_reward": 0.8168955128271573,
      "kl_from_reference": 1.0400522058527373,
      "value_model_parameters": 0,
      "rollout_policy_refreshes": 18
    },
    "dapo": {
      "accuracy": 0.7578125,
      "mean_reward": 0.7996058785056733,
      "kl_from_reference": 1.0870483587827842,
      "rollout_policy_refreshes": 18,
      "clip_low": 0.2,
      "clip_high": 0.28,
      "clip_fraction": 0.25,
      "mean_pseudo_tokens": 2.25,
      "overlong_penalty": 0.0125
    },
    "gspo": {
      "accuracy": 0.828125,
      "mean_reward": 0.8508578717725304,
      "kl_from_reference": 0.7016685190362502,
      "rollout_policy_refreshes": 18,
      "sequence_ratio_mean": 1.0130756461379398,
      "sequence_ratio_std": 0.01168781319879202
    },
    "ppo-rlhf": {
      "accuracy": 0.8125,
      "mean_reward": 0.8335089378910975,
      "kl_from_reference": 0.8731202191948979,
      "critic_updates": 300,
      "rollout_policy_refreshes": 18
    },
    "rloo": {
      "accuracy": 0.828125,
      "mean_reward": 0.8508578717725304,
      "kl_from_reference": 0.5707328389167291,
      "leave_one_out_samples": 4,
      "value_model_parameters": 0
    },
    "remax": {
      "accuracy": 0.703125,
      "mean_reward": 0.7553879371297694,
      "kl_from_reference": 0.79386672912852,
      "value_model_parameters": 0
    }
  },
  "schema_version": 2,
  "manifest_ref": "experiments:classic-post-training-gsm8k-seed42",
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/experiments/classic-post-training-gsm8k-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}