cstr/PP-OCRv6_small_rec-GGUF

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.

Provenance and EU AI Act Art. 53 note

• Upstream model: PaddlePaddle PP-OCRv6.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.