artifacts/adamw_tuned_cmpatino-0/results.json

{
  "agent_id": "cmpatino-0",
  "timestamp": "2026-04-30T21:46:00Z",
  "experiment": "AdamW tuned with (lr=0.0010, wd=0.05) at 5625 steps -- negative result",
  "optimizer": "AdamW (multi-group)",
  "steps_to_3_28": null,
  "final_val_loss": 3.30295,
  "num_runs": 1,
  "key_hparams": {
    "block_lr": 0.0010,
    "block_wd": 0.05,
    "block_betas": [0.9, 0.95],
    "warmup": 250,
    "cooldown_frac": 0.7,
    "train_steps": 5625,
    "aux_groups_unchanged": true
  },
  "trajectory_endpoints": {
    "step_5000": 3.32404,
    "step_5125": 3.31881,
    "step_5250": 3.31324,
    "step_5375": 3.30869,
    "step_5500": 3.30507,
    "step_5625": 3.30295
  },
  "comparison": {
    "v2_baseline_5625": 3.28434,
    "delta": 0.01861,
    "delta_sign": "worse"
  },
  "notes": "Did NOT beat v2 baseline. Half-length sweep (2812 steps) showed (lr=0.0010, wd=0.05) gave 3.43422 vs baseline 3.46050 (+0.026 better). Full-length result reverses: tuned 3.30295 vs baseline 3.28434 (-0.019 worse). Confirms README caveat that early-step val_loss doesn't predict final. Implication: for LR/WD tuning of this multi-LR AdamW recipe, full-length runs are required; half-length signal can be misleading."
}