piotrekwitek/ppocr-v6-small-mlx

PP-OCRv6 small — MLX runtime weights

Project-hosted MLX weight snapshot consumed by pii-redactor (macOS arm64 only).

Contents

File	Task	Origin
PP-OCRv6_det_small_mlx.safetensors	text detection (DB)	Offline safetensors conversion of the pinned PP-OCRv6 small ONNX detector used by the OCR runtimes evaluation; tensor names/layout follow the released Paddle checkpoint. Verified numerically against onnxruntime (identical argmax; iteration 63/69).
PP-OCRv6_rec_small_mlx.safetensors	multilingual text recognition (SVTR)	Offline safetensors conversion of the pinned PP-OCRv6 small ONNX recognizer; verified numerically against onnxruntime on product crops (max
PP-LCNet_x0_25_textline_ori_mlx.safetensors	0°/180° textline orientation	Project-derived re-expression of PaddlePaddle/PP-LCNet_x0_25_textline_ori_onnx @ aea1f18d97338aaf84b463f90192f2820524a00a (inference.onnx, sha256 94a6a0a0…f11f2ff3). Extracted from the export's graph constants, BatchNorm folded, layout frozen by spikes/ocr_runtimes_eval/extract_ori_weights.py. Numeric parity vs onnxruntime: max
ppocr_keys_v1.txt	recognizer vocabulary	Extracted verbatim from the pinned recognizer export's character metadata (18,708 entries); the runtime no longer parses the container.

Licence and attribution

Upstream PaddleOCR / PP-OCR models are Apache-2.0. The orientation model is from the official PaddlePaddle namespace (Apache-2.0 per its model card). Conversion and derivation were performed by this project; the derived files are not byte-verbatim upstream objects. See pii_redactor/resources/licenses/PP-OCR-ATTRIBUTION.md in the product repo.

No training data, private documents, or personal data are included.