validation_report.json

{
  "method": "4-bit block-32 weight-only (MatMulNBits)",
  "file": "model_q4.onnx",
  "cosine_mean": 0.9518644660711288,
  "cosine_min": 0.928617000579834,
  "top1_agreement": "4/4",
  "top1_ok": true,
  "spearman_mean": 0.886904761904762,
  "gate": "top-1 doc agreement on all queries + mean cosine >= 0.90 vs fp32 PyTorch",
  "queries": 4,
  "docs": 8
}