Glimpse: Multi-Modal OCR

## Inspiration

"While modern VLMs are masters of English and Chinese, they often treat Right-to-Left (RTL) scripts like Arabic and Persian as an afterthought. We built Glimpse to prove that the PaddleOCR-VL architecture can be 'globalized' through specialized fine-tuning, bringing high-precision OCR to under-represented, low-resource languages."

## Challenges we ran into

"The primary hurdle was 'Script Bias.' Base models are pre-trained on Left-to-Right (LTR) data. When presented with Arabic/Persian, the model's visual grounding and sequential logic were initially misaligned (Base CER: ~58%). We had to implement a training regime that focused on Unicode-aware alignment to handle the complex ligatures and cursive nature of the RTL script."

## Accomplishments that we're proud of

"We successfully reduced the Character Error Rate (CER) from a chaotic baseline to an industrial-grade 6.9% in just 500 steps. We achieved a near-perfect validation loss of 0.25, demonstrating that Glimpse has successfully mastered the linguistic 'joiner' logic of the Arabic script, producing human-legible transcriptions from diverse visual fonts."

## What we learned

"We learned that even a model with a billion parameters needs 'linguistic discipline.' Fine-tuning isn't just about showing more images; it’s about aligning the model's internal probability map to the specific grammar of the target language. In the RTL domain, Inference Normalization is just as important as training loss."

## Credit & Aknowledgement

Data Acknowledgements: We would like to express our gratitude to Mohammad Reza Hajesmaeili for providing the Persian_Arabic_TextLine_Image_Ocr_Medium dataset. This open-source contribution was essential in training our model to handle the nuances of Persian and Arabic scripts.

Built With

• paddleocr
• unsloth