{
  "source_model": "fastino/gliner2-privacy-filter-PII-multi",
  "opset": 17,
  "quantization": {
    "scheme": "B_fp16_weights",
    "description": "weight-only compression, fp32 compute: MatMul weights stored fp16 (Cast to fp32 at load), 250112x768 token-embedding table stored int8 with per-row symmetric scales (dequantized after Gather)",
    "max_prob_drift_vs_pytorch": 0.005668938159942627,
    "span_set_parity": "identical to PyTorch engine at thresholds 0.3/0.5/0.7 on the golden suite"
  },
  "fp32_parity": "max prob drift 1.29e-05, span sets identical (see export_to_onnx.py output)",
  "files": {
    "model.quantized.onnx": {
      "sizeBytes": 446391629,
      "sha256": "489edc522a34b8c6a0630c4710998abf0227b11acf45600ba2bce95bcd943930"
    },
    "tokenizer.json": {
      "sizeBytes": 16020604,
      "sha256": "f6df10ec83bea993035b2dd7c39345a3d4fcf23421c2adb6cb4ffc1e6d1bc4b5"
    },
    "tokenizer_config.json": {
      "sizeBytes": 711,
      "sha256": "233beed1f1095cccfc7907cde31a8d90a0c6aa4fdfaf6493f8e55fd162e81ae6"
    }
  }
}