anushkrishna/falcon-ocr-models

Falcon OCR model pack

Pinned English/Arabic OCR artifacts for the Falcon desktop app. This is a distribution pack containing three separate models, not one Transformers model. These are PaddlePaddle models mirrored by Falcon, with no fine-tuning or changes to the ONNX bytes. They run locally using ONNX Runtime.

File	Official upstream	Role
paddleocr/det_small.onnx	PaddlePaddle/PP-OCRv6_small_det_onnx	Text detection
paddleocr/rec_en_mobile.onnx	PaddlePaddle/en_PP-OCRv5_mobile_rec_onnx	English recognition
paddleocr/rec_ar_mobile.onnx	PaddlePaddle/arabic_PP-OCRv5_mobile_rec_onnx	Arabic recognition
paddleocr/en_dict.txt	English recognizer's inference.yml	436 dictionary entries
paddleocr/ar_dict.txt	Arabic recognizer's inference.yml	747 dictionary entries

The ONNX files are renamed from upstream inference.onnx. Dictionaries are the upstream PostProcess.character_dict entries joined with LF, with a trailing LF, encoded as UTF-8. At runtime Falcon prepends the CTC blank class and appends the space class, giving 438 English and 749 Arabic classes. The dictionaries match the small dictionaries compiled into Falcon.

manifest.json records immutable upstream revisions, file sizes, SHA-256 hashes, and transformations. Original model cards and configuration files are preserved in sources/. Attribution: PaddlePaddle / PaddleOCR authors; see LICENSE.md and Apache-2.0. Falcon packaging does not imply endorsement by the original authors.

Use a full release commit SHA in download URLs, for example:

https://huggingface.co/anushkrishna/falcon-ocr-models/resolve/<commit-sha>/paddleocr/det_small.onnx

Verify both file size and SHA-256 before loading, and activate the complete pack together. No HF account or publisher token is needed for public downloads. Do not follow main for application installs. Tested application releases choose new revisions; uploads here do not automatically update existing installations.

This release contains roughly 25.7 MB of weights plus dictionaries. It excludes ONNX Runtime, PDF rasterizers, orientation classifiers, layout models, and user documents. The models have recognition errors, particularly on small text and digits; this mirror makes no new accuracy claims.