Falcon OCR Arabic
Falcon OCR Arabic adapts the 270M Falcon OCR architecture to Arabic documents. TII trained it on real and synthetic documents, then applied reinforcement learning. The announced public route is the Arabic OCR playground.
Arabic changes more than the character set
The training targets joined letters, dots, optional diacritics, mixed scripts and right-to-left reading order. It keeps the original early-fusion architecture. Output formats include plain text, formula LaTeX and table HTML. TII identifies dense magazine, newspaper and comic pages as harder cases.TII: Falcon OCR Arabic
Check reading order in the result
Practical guidance. Use a page with Arabic text beside Latin names and mixed numeral styles. Check each table row and column in reading order. The base Falcon OCR weight license does not by itself grant access to this separate Arabic adaptation.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Provider product |
| Model identifiers | Falcon-OCR-Arabic |
| Model size | 270M parametersTII: Falcon OCR Arabic |
| Access | Provider playground; downloadable Arabic weights and their license were not verifiedTII: Falcon OCR Arabic |
| Adaptation | Supervised fine-tuning followed by reinforcement learningTII: Falcon OCR Arabic |
Related models
- Falcon OCR 1.5
Technology Innovation Institute (TII) ยท Falcon OCR
Falcon OCR reads document images with a 270M parameter model.
Launch date not confirmedDownloadable