{
  "epoch": 21,
  "step": 127974,
  "source_checkpoint_sha256": "019d665185348b40679911dec3528abd61d1c9f9310036ae3c07d3c4f33d8777",
  "evaluation_scope": "Prepared development EVAL; same split used for training monitoring and checkpoint selection; not an independent test",
  "variants": {
    "fp32": {
      "variant": "fp32",
      "file": "model.onnx",
      "sample_sha256": "20a60e7a747cdf9a698a464c9d324612aad2ca6b43a541db0813554eb9af8694",
      "supported_tags": 14767,
      "binned_mAP": 0.5197223424911499,
      "exact_mAP": 0.5222825981934666,
      "micro_f1_at_0_7927": 0.6176290651268128,
      "threshold_calibration": {
        "method": "Maximum micro-F1 on 200 uniformly spaced thresholds over the same development EVAL set",
        "threshold": 0.6984924674034119,
        "micro_f1": 0.6499312610657317,
        "precision": 0.6762897056253009,
        "recall": 0.6255503855054322,
        "independent_test": false,
        "excluded_indices": [
          0,
          1
        ]
      },
      "providers": [
        "CUDAExecutionProvider",
        "CPUExecutionProvider"
      ]
    }
  },
  "reference_pytorch_bf16_exact_mAP": 0.5219968496971725,
  "fp32_minus_pytorch_bf16_exact_mAP": 0.000285748496294036,
  "versions": {
    "pytorch": "2.14.0+cu130",
    "onnxruntime": "1.29.0",
    "numpy": "2.2.6"
  },
  "note": "ONNX uses an SDPA-derived FP32 graph; the historical PyTorch replay used compiled bf16 Flex Attention. Precision and kernel differences can shift scores slightly."
}