{
  "model": "models/decider/4b/mlx/q4",
  "model_sha256": "d0ba3ca5953871030be574146a6eef2d8073d556041756647753921112afaa7f",
  "config_sha256": "93890b32ac4a985e98ee67132c2f95ad0b3f4087b3941d10b6ec14f41e353a2a",
  "tokenizer_sha256": "06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523",
  "hardware": "Apple M5 Max",
  "memory_bytes": 137438953472,
  "versions": {
    "mlx": "0.31.2",
    "mlx-lm": "0.31.3",
    "transformers": "5.9.0",
    "numpy": "2.4.6",
    "scipy": "1.17.1",
    "evalplus": "0.3.1"
  },
  "precision": "MLX affine 4-bit, group size 64",
  "model_identity": "Mapika/decider-4b v2.1 local MLX conversion; Qwen3.5-4B-Base derivative",
  "mode": "Native plain state-first direct letter-logit slot. No chat template or generated reasoning; specialized classifier mode.",
  "training": "No pilot training; different model from planned Qwen2.5-Coder-3B distillation",
  "prompt_template": "Context:\n{context}\n\nQuestion: {question}\nOptions:\n(A) {option_0}\n(B) {option_1}\nAnswer: (",
  "answer_token_ids": [32, 33, 34],
  "load_ms": 1227.91,
  "timing": "Warm resident model; sequential uncached rotations; synchronized MLX extraction + tokenization. No network/cloud or load time."
}