artifacts/adamw_baseline_cmpatino-0/results.json

{
  "agent_id": "cmpatino-0",
  "timestamp": "2026-04-30T17:11:00Z",
  "experiment": "Pure single-LR AdamW @ baseline-stated hparams (negative result)",
  "optimizer": "AdamW",
  "steps_to_3_28": null,
  "final_val_loss": 3.39869,
  "num_runs": 1,
  "key_hparams": {
    "lr": 0.0015,
    "wd": 0.1,
    "betas": [0.9, 0.95],
    "warmup": 250,
    "cooldown_frac": 0.7,
    "train_steps": 5625,
    "scheme": "single-LR (all params lr=0.0015)",
    "init": "Muon-style: proj zeroed, embed normal_(), other weights normal_(std=0.33**0.5/n**0.5)"
  },
  "notes": "Did not reach 3.28. Root cause identified by reading upstream reference log: the 'AdamW baseline' on the leaderboard actually uses a multi-LR scheme (embed lr=0.3, proj lr=1/320, ndim<2 lr=0.01, blocks lr=0.0015) with only proj zeroed at init. Single-LR with lr=0.0015 for everything is severely undertuned for embed/proj/scalars. v2 reproduction launched at artifacts/adamw_baseline_v2_cmpatino-0/."
}