How Far Can Synthetic Data Take Thai OCR?
A Thai OCR model trained only on synthetic pages nearly matched real-label performance on printed documents.
The paper isolates what helps synthetic OCR data transfer to real Thai pages: typeface diversity, two-dimensional layout, and real handwriting glyphs mattered; non-text context did not consistently help. Its Wayu-Paxa-OCR-Zero adapts PaddleOCR-VL-1.6 with 45,723 synthetic pages and no OCR labels from real Thai document pages. Median character error rate fell from 6.64% to 1.24% on printed pages, and from 74.87% to 20.55% on handwriting, versus the base checkpoint. The model also beat Typhoon OCR v1 7B across all five evaluation sets. HF Daily Papers' note
The paper isolates what helps synthetic OCR data transfer to real Thai pages: typeface diversity, two-dimensional layout, and real handwriting glyphs mattered; non-text context did not consistently help. Its Wayu-Paxa-OCR-Zero adapts PaddleOCR-VL-1.6 with 45,723 synthetic pages and no OCR labels from real Thai document pages. Median character error rate fell from 6.64% to 1.24% on printed pages, and from 74.87% to 20.55% on handwriting, versus the base checkpoint. The model also beat Typhoon OCR v1 7B across all five evaluation sets. HF Daily Papers' note
score 4