Upload README.md with huggingface_hub

README.md

@@ -56,9 +56,12 @@ The released safetensors are BitNet QAT master weights — they only produce sen
 
 ## Validation (onnxruntime CPU, greedy, vs ternarized PyTorch reference)
 
-fp32 ONNX: 4/4 transcripts byte-identical. q4 / q2: 3/4 byte-identical; the FLEURS-fr clip
-differs by two words (int8-encoder-class deviation, same class as the reference GGUF engine).
-Full transcripts in `conversion_report.json`.
+fp32 ONNX: 4/4 transcripts byte-identical. q4 / q4f16 / q2: 3/4 byte-identical; the FLEURS-fr
+clip differs by two words ("géographe"→"géologue", "des"→"les"). For comparison, the official
+ggml engine (VibeASR.cpp, I8_S+I2_S) transcribes that same clip as *"aux gens ouverts fiscales,
+mais assètent sur les toits de douane"* — these ONNX builds are strictly closer to the fp
+reference than the original GGUF engine on every tested clip. Full transcripts in
+`conversion_report.json`.
 
 ## Files / sizes