wormOCR 1.0.0

First PyPI release of the wormOCR package.

• pip install wormocr
• wormocr download for wormocr-2M
• wormocr download wormocr-9M for the larger model

Weights stay on GitHub Releases (not on PyPI). See the README for measured scores.

wormocr-9M

wormocr-9M

8,776,416 parameters. Same printed-and-handwriting training mix as wormocr-2M, larger CRNN (input height 48, BiGRU 512).

• Best checkpoint: training step 19,000 (run early-stopped at 30,000)
• Mixed validation CER: 0.0742
• ONNX max abs diff vs PyTorch: 1.24e-5
• Does not replace wormocr-2M; download either model

Download

wormocr download wormocr-9M

Measured CER (greedy, 6 Nov 2026, RTX 5060)

Set	wormocr-2M	wormocr-9M
Printed synthetic	3.1%	3.7%
FUNSD (eval only)	11.5%	10.2%
CORD	3.7%	3.1%
Synthetic handwriting	9.2%	7.1%
Handwritten names	5.3%	5.3%
GNHK	23.3%	19.2%
IAM (eval only)	22.4%	18.3%
Bentham	15.1%	8.8%
RxHand case-insensitive (eval only)	37.9%	35.2%

Assets

• wormocr-9M.safetensors (33.5 MB)
• wormocr-9M.json
• wormocr-9M.onnx (33.5 MB)
• SHA256SUMS.txt

wormocr-2M

Printed and handwritten English lines. 2,174,880 parameters. Same network as the previous file on this tag.

This replaces the printed-only weights published earlier under wormocr-2M. If you already downloaded that copy, run wormocr download --force.

Best checkpoint is step 58,000 of a 60,000-step GPU run. Greedy character error on 23 October 2026:

• Printed synthetic: 3.1% (1,000 lines)
• FUNSD forms (evaluation only): 11.5% (8,526)
• CORD receipts: 3.7% (795)
• Synthetic handwriting: 9.2% (1,000)
• Handwritten names: 5.3% (26,170)
• GNHK notes: 23.3% (2,410)
• IAM lines (evaluation only): 22.4% (2,915)
• Bentham: 15.1% (1,135)
• RxHand words, case-insensitive (evaluation only): 37.9% (1,107)

Code: f33d517