{
 "id": "qwen3.three_build",
 "kind": "comparison",
 "unit": "in-loop val PPL (n=1, not suite-stamped)",
 "entries": [
  {
   "name": "Faithful baseline (AdamW)",
   "value": 28.65
  },
  {
   "name": "Modernized (IMU-1 bundle)",
   "value": 23.52
  },
  {
   "name": "Exploratory (partial-RoPE 0.25)",
   "value": 29.54
  }
 ],
 "direction": "lower_is_better",
 "generated_by": "scripts/export_site_results.py",
 "generated_at": "2026-07-07T19:00:23Z",
 "notes": "Three builds at 1,189,478,400 tokens (18,150 steps x 65,536), single seed each, trainer-internal eval on held-out FineWeb-Edu. The -17.9% IMU-1 margin is motivation, not a result: the bundle changes 6 things at once (incl. NorMuon + cautious-WD); attribution lives in the de-confound experiments. IMU-1 final = eval@18000 (run DONE at 18,150; no after.txt). Fourth arm (partial-RoPE 0.10) died at step 5,450/18,150.",
 "derived_from": [
  "Qwen3-0.6B/builds/2026-06-08_reproduce-faithful_qwen3-0.6b/results/qwen3_baseline2tpp_after.txt",
  "Qwen3-0.6B/builds/2026-06-08_reproduce-modernized_qwen3-0.6b/results/qwen3_imu1_2tpp_train.log",
  "Qwen3-0.6B/builds/2026-06-08_reproduce-exploratory_qwen3-0.6b/results/qwen3_prope25_2tpp_train.log"
 ]
}