{
  "base_model": "Qwen/Qwen3.5-4B",
  "base_snapshot_revision": "not recorded in retained training configuration",
  "retained_floating_source_sha256": {
    "model-00001-of-00002.safetensors": "5da9935a53daa73bf22809a0f89165e8d4406328ecfd6b8c3ac3f38c137c1a56",
    "model-00002-of-00002.safetensors": "01af619b9a984635330554b09cb45b652673230d55aa568a76d09449f7d066e8"
  },
  "training_method": "LoRA on quantized 4-bit base followed by retained floating-point fusion",
  "native_unquantized_base_precision": false,
  "dequantized_from_public_mlx_6bit": false,
  "training_dataset_public": false,
  "known_limitation": "Arabic context case adds trailing newline and fails production format check",
  "nvidia_cuda_inference_verified": false,
  "amd_rocm_inference_verified": false,
  "mobile_inference_verified": false,
  "tested_hardware": "Apple M5 Pro MacBook Pro, 48 GB",
  "format": "Hugging Face Safetensors",
  "tensor_roundtrip_exact": 426,
  "source_dtype_counts": {
    "mlx.core.bfloat16": 402,
    "mlx.core.float32": 24
  },
  "transformations": [
    "Remove MLX wrapper prefix",
    "Transpose depthwise convolution",
    "Undo offset RMSNorm with exact roundtrip",
    "Set absent MTP auxiliary layers to zero; no tensors added or removed"
  ],
  "runtime_validation": "Native MLX targeted inference; Transformers load without missing or unexpected weights; no full-suite score"
}