PP-OCRv6 fine-tuned recognizers for Hebrew + Indic
This is a fine-tune of PP-OCRv6 'small', one for Hebrew, one for (Bengali, Gujarati, Kannada, Malayalam). Both have Latin as well.
Hebrew does not do Niqqud.
Trained exclusively on synthetic data, evaluated against 3 pictures, was better than Tesseract.
- Input strip height is 48; output is already softmax (per-char confidence = max prob).
- Emits glyphs in visual (left-to-right) order (need reversal logic for Hebrew)
Training code
scripts/rec_model/ in translator-rs.
License
Fine-tune of PP-OCRv6 (Apache-2.0). Synthetic training data rendered with mixed-license fonts (Culmus, Google Fonts OFL, SIL) over Leipzig corpora.