BalOCR experimental pilot

This LoRA adapter was trained for 40 optimizer steps on 80 unique Balochi OCR examples. It saw 41 real Arabic-script pairs and 39 synthetic Latin-script pairs. Full BalOCR training is pending. The checkpoint contains 626,688 adapter parameters and loads with PaddleOCR-VL-1.6 at the pinned revision.

The early diagnostics covered 12 Arabic images and 12 synthetic Latin images. Character error rate changed from 32.3% to 45.2% for Arabic and from 7.4% to 5.3% for synthetic Latin. One adapted Arabic output reached the token cap. The early split was superseded after a later audit found overlap between cropped references and longer training paragraphs. These figures describe only that diagnostic run. Accuracy on unseen Arabic books and real Latin scans remains unverified.

Try the adapter

Clone the BalOCR source repository and follow its environment setup. Run this from the checkout with a small printed image. The SDK loads the exact base revision and adds the pilot weights from this repository.

python -m balocr.model --image page.png --adapter shah-bakhsh/BalOCR --device cpu --output result.json

The Hugging Face Space currently serves the original pretrained base. The mixed pilot results are recorded in training_summary.json.

Source credits · Prepared dataset

Downloads last month
20
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for shah-bakhsh/BalOCR

Adapter
(7)
this model

Dataset used to train shah-bakhsh/BalOCR