{
 "id": "qwen3.mix",
 "kind": "comparison",
 "unit": "ΔBPB vs FineWeb-Edu baseline (code_py)",
 "entries": [
  {
   "name": "dclm-edu (100%)",
   "value": 0.7034377813053776,
   "ci": [
    0.6639233630446848,
    0.7429521995660704
   ],
   "seeds": [
    0.7124500661615261,
    0.6984721465231007,
    0.6993911312315064
   ],
   "significant": true
  },
  {
   "name": "50/50 mix",
   "value": 0.5904307914737172,
   "ci": [
    0.5531974784151051,
    0.6276641045323293
   ],
   "seeds": [
    0.5896871828270531,
    0.5733911502259641,
    0.6082140413681341
   ],
   "significant": true
  }
 ],
 "direction": "higher_is_better",
 "generated_by": "scripts/export_site_results.py",
 "generated_at": "2026-07-07T19:00:23Z",
 "notes": "Data-composition curve: the 50/50 dclm+FineWeb mix keeps 84% of the pure-dclm code win while holding English (mix wikitext-2 +0.0161 [-0.0005, 0.0327] n.s. vs dclm -0.0097 n.s./worse). Best-of-both; this mix is the data carried into mid-training. 3 seeds, Welch-t 95% CI, seeds[] are paired per-seed deltas.",
 "derived_from": [
  "Qwen3-0.6B/experiments/2026-06-26_qwen3-0.6b_data-mix-composition/verdict.json",
  "Qwen3-0.6B/experiments/2026-06-26_qwen3-0.6b_data-mix-composition/cohort_bpb.json"
 ]
}