artifacts/adamw_baseline_v2_cmpatino-0/results.json

{
  "agent_id": "cmpatino-0",
  "timestamp": "2026-04-30T18:04:00Z",
  "experiment": "Multi-LR AdamW baseline reproduction (calibration)",
  "optimizer": "AdamW (multi-group)",
  "steps_to_3_28": null,
  "final_val_loss": 3.28434,
  "num_runs": 1,
  "key_hparams": {
    "block_lr": 0.0015,
    "block_wd": 0.1,
    "block_betas": [0.9, 0.95],
    "aux_groups": {
      "embed.weight": {"lr": 0.3, "wd": 0, "betas": [0.8, 0.95]},
      "proj.weight": {"lr": 0.003125, "wd": 0, "betas": [0.8, 0.95]},
      "ndim<2 (biases, gains)": {"lr": 0.01, "wd": 0, "betas": [0.8, 0.95]}
    },
    "warmup": 250,
    "cooldown_frac": 0.7,
    "train_steps": 5625,
    "init": "only proj zeroed; default torch init elsewhere"
  },
  "trajectory_endpoints": {
    "step_5000": 3.30826,
    "step_5125": 3.30215,
    "step_5250": 3.29597,
    "step_5375": 3.29094,
    "step_5500": 3.28678,
    "step_5625": 3.28434
  },
  "notes": "Calibration of upstream AdamW baseline (3.274 reference) on 2xH100. Reached 3.28434 at step 5625 -- ~0.01 above reference, attributable to numerics differences between 2-GPU and 8-GPU runs (torch fused AdamW + all_reduce ordering). Trajectory shape matches expected. Used as the reference-curve for downstream sweeps; not a leaderboard submission."
}