PP-OCRv6 small recognizer — CrispEmbed GGUF
Files: PP-OCRv6_small_rec-f16.gguf, PP-OCRv6_small_rec-f32.gguf,
PP-OCRv6_small_rec-q8-head.gguf, and
PP-OCRv6_small_rec-crispasr-q4_k-policy.gguf.
The policy-q4 container intentionally keeps the complete PP-OCRv6 recognizer graph in F16: quantizing intermediate CNN/SVTR weights caused compounding CTC drift. Source: PaddlePaddle PP-OCRv6, Apache-2.0.
The native F32 conversion matches the reference through logits (cosine 1.0). The published F16 artifact is an inference-speed artifact, not a parity claim: repeated half-precision layers accumulate measurable drift, so production quality should use the Q8/F32-quality policy artifact once published.
q8-head is built from F32 and quantizes only the final SVTR/CTC head;
measured logits cosine is 0.999987.
- Downloads last month
- 4