LibrePaddleOCRVL0.9b

PaddleOCR-VL weights mirrored for LibreYOLO's ocr task: the PaddleOCR-VL-0.9B recognizer at the repository root and the PP-DocLayoutV2 layout detector under layout/, so one snapshot holds the whole two-stage document parser.

Source

• Root: mirrored from PaddlePaddle/PaddleOCR-VL at commit 7fa00a8c55b735ba51ba49a9058f3f9c57a99a11.
• layout/: mirrored from PaddlePaddle/PP-DocLayoutV2_safetensors at commit 880e8971b88938518611c54fc0f59ad57849c9d4 (config.json, model.safetensors, preprocessor_config.json).

Copyright (c) PaddlePaddle Authors. Licensed under the Apache License 2.0. The recognizer's language model is initialized from ERNIE-4.5-0.3B (Baidu, Apache-2.0), as stated by the upstream model card.

Modifications

No learned parameters were changed. Files are byte-identical to the upstream snapshots. Not mirrored: the upstream remote-code Python files (LibreYOLO loads the native transformers classes), the Paddle-format PP-DocLayoutV2/ folder (replaced by its safetensors release under layout/), and the upstream model card, replaced by this one.

Usage

from libreyolo import LibreVLM

model = LibreVLM("paddleocr-vl")
result = model.predict("page.png")
print(result.ocr.markdown)

Requires pip install 'libreyolo[vlm]' with transformers >= 5.3.

License

Apache License 2.0. See LICENSE and NOTICE.