sft ·craftax ·local ·deferred

Local MLX SFT on Craftax

Hypothesis

LoRA SFT through synth-mlx-rl plus Craftax fixture eval is a valid local-training cookbook.

Verdict

Proved only on FakeEngine plus a scripted craftax_engine fixture. Workshop-to-local-MLX has not passed.

Confidence

high

Why

Fixture seeds 0–1 scored 0.5; SFT cross_entropy + checkpoint succeeded. Not the public gold eval matrix.

Measured receipt

{
  "cookbook_id": "mlx-sft-craftax",
  "proved_at": "2026-08-19T00:54:41Z",
  "craftax_eval": [
    {
      "seed": 0,
      "reward": 0.5,
      "status": "scored",
      "environment_ref": "env:craftax_fixture"
    },
    {
      "seed": 1,
      "reward": 0.5,
      "status": "scored",
      "environment_ref": "env:craftax_fixture"
    }
  ],
  "sft": {
    "cross_entropy_supported": true,
    "loss": 0.25,
    "training_version": 1,
    "checkpoint": "sft-craftax-smoke",
    "http": [
      200,
      200,
      200
    ]
  },
  "claim": "scripted Craftax fixture eval + LoRA SFT step on synth-mlx-rl FakeEngine",
  "not_claimed": "Apple-silicon base-model fine-tune or /evals/craftax gold matrix"
}

Source: cookbooks/mlx-sft-craftax. Reproduce from the public cookbooks worktree; do not treat not_claimed fields as results.