{
  "paper": {
    "title": "DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining"
  },
  "dataset": {
    "name": "WikiText-2 + public narrative domain",
    "tokens": 240000
  },
  "baseline": {
    "name": "uniform mixture",
    "validation_loss": 5.672924328383141
  },
  "method": {
    "name": "DoReMi group-DRO mixture",
    "validation_loss": 5.664472991106605
  },
  "relative": {
    "validation_loss_percent": -0.14897673205778875
  },
  "stages": {
    "proxy_steps": 80,
    "final_domain_weights": [
      0.5140855253837019,
      0.4859144746162981
    ],
    "worst_excess_loss_updates": 80,
    "reference_models": 2
  },
  "paper_results": {
    "few_shot_accuracy_points": 6.5,
    "training_speedup_x": 2.6
  },
  "scope": "实际在两个公开文本域上训练 reference unigram proxy，执行 group-DRO excess-loss 指数权重更新，再以所学配比评估；未复刻 The Pile、280M proxy 与 8B target。",
  "schema_version": 2,
  "manifest_ref": "reproduction:doremi",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42
    ],
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2305.10429-doremi/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}