31B real rollouts · 480 (exact stack, 32 lanes, 13-Nov)

{
  "adapter": "rank-16 LoRA loaded with all-zero B weights, so it adds nothing (base model)",
  "evaluation_set": "60 SWE tasks x 8 rollouts (rollout31b-real test kit)",
  "model": "Gemma 4 31B it, QAT W4A16 (gemma-4-31b-it-qat-w4a16-ct)",
  "name": "31B real rollouts · 480 (exact stack, 32 lanes, 13-Nov)",
  "notes": "The first box was taken back by Spot after 207 rollouts; the restart reran all 480. Only the restart's rows are shown (last row per run_key).",
  "result_kind": "local rollout bench (not a Kaggle score)",
  "row_selection": "last row per run_key",
  "rows_kept": 480,
  "settings": {
    "faucet_job": "rollout31b-480-1 (4e4aea5c876851a0), attempt 3",
    "gpu": "1x NVIDIA RTX PRO 6000 Blackwell (faucet box g4-standard-48, Spot)",
    "lanes": 32,
    "session_cap_minutes": 7.5,
    "stack": "exact (lossless) rollout stack, vLLM 0.19.1, torch 2.10.0+cu128",
    "tool_call_budget": 30
  },
  "source": "gs://cellensml-training-jobs/jobs/son/4e4aea5c876851a0/s0/a3/working/r31/rollouts_X",
  "source_rows": 687,
  "source_sha256": "03479861121e195ca3eb79ae47872470b8bfc369ef52cb0bb26bc0b56feb583d",
  "source_timestamp": "2026-11-13 22:19 to 2026-11-14 00:07 New York time (the complete second attempt)"
}