pylopdf-ocr-models

A data-only package containing the PP-OCRv6 small detector, recognizer, and multilingual character dictionary used by pylopdf. The models run locally through pylopdf's pure-Rust RTen backend; OCR does not require a system executable, ONNX Runtime, an ONNX parser at runtime, or network access.

Install it through the main package extra:

pip install "pylopdf[ocr]"

The small PP-OCRv6 tier is intended for mobile and desktop use. Its unified model supports 50 languages, including Japanese, Simplified Chinese, Traditional Chinese, and English.

Bundled artifacts

The installed package also includes these values in SHA256SUMS. Release CI builds the wheel and sdist twice and requires byte-identical output. It then installs both formats into isolated environments, verifies distribution metadata and license files, and streams each installed artifact through SHA-256 before publishing.

The source ONNX artifacts and dictionary are the PP-OCRv6 small files listed in RapidOCR's pinned model manifest and downloaded from its versioned v3.9.2 model repository. They originate from PaddleOCR; RapidOCR converted and distributes the ONNX forms.

Reproducing the RTen conversion

After downloading and verifying the source files above, reproduce the bundled model data from the repository root with:

uvx --from rten-convert==0.22.0 rten-convert \
  --no-infer-shapes PP-OCRv6_det_small.onnx PP-OCRv6_det_small.rten
uvx --from rten-convert==0.22.0 rten-convert \
  --no-infer-shapes PP-OCRv6_rec_small.onnx PP-OCRv6_rec_small.rten

rten-convert embeds each source ONNX SHA-256 in the RTen file. Conversion is deterministic: repeating the commands produces the artifact hashes in the first table. Numerical outputs were also checked against the ONNX models before packaging. The character dictionary remains a separate input because RTen's conversion format does not preserve the ONNX recognizer's custom character metadata.

Model copyright is held by Baidu. PaddleOCR and RapidOCR are licensed under Apache License 2.0. See LICENSE and NOTICE.