{
  "paper": {
    "title": "Data Mixing Laws: Optimizing Data Mixtures by Predicting LM Performance"
  },
  "dataset": {
    "name": "WikiText-2 + public narrative domain",
    "tokens": 240000
  },
  "baseline": {
    "name": "uniform mixture",
    "validation_loss": 5.672924328383141
  },
  "method": {
    "name": "predicted optimal mixture",
    "validation_loss": 5.475767325192497
  },
  "relative": {
    "validation_loss_percent": -3.4754033683159733
  },
  "stages": {
    "pilot_mixtures": 10,
    "law": "L_i(p)=c_i+exp(a_i p+b_i)",
    "selected_domain_0_weight": 0.99,
    "unseen_grid_points": 99
  },
  "paper_results": {
    "equivalent_extra_training_percent": 48.0
  },
  "scope": "实际用公开文本域的十组 pilot mixture 拟合逐域指数 mixing law，并在未训练配比网格上选择最优配比；小型 unigram proxy 不等同于 RedPajama 多模型 scaling curve。",
  "schema_version": 2,
  "manifest_ref": "reproduction:data-mixing-laws",
  "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/2403.16952-data-mixing-laws/metrics/public-seed42.json",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}