Misraj/Baseer__Nakba
Misraj/Baseer__Nakba is a 3-billion parameter Vision-Language Model (VLM) developed by Misraj AI, specifically fine-tuned for Arabic Handwritten Text Recognition (HTR). It excels at parsing and recognizing highly cursive, historical Arabic manuscripts, achieving 1st place in the NAKBA NLP 2026 competition with a Character Error Rate (CER) of 0.0790. This model is optimized for challenging document degradation and diverse writing styles, making it ideal for digitizing historical Arabic texts.
Loading preview...
Baseer-Nakba HTR: A Specialized VLM for Arabic Handwritten Text Recognition
Misraj/Baseer__Nakba is a 3-billion parameter Vision-Language Model (VLM) developed by Misraj AI, specifically designed for Arabic Handwritten Text Recognition (HTR). This model adapts the Baseer VLM to effectively parse and recognize highly cursive, historical Arabic manuscripts, addressing challenges like varying writer styles and document degradation.
Key Capabilities
- Top-tier HTR Performance: Secured 1st place in the NAKBA NLP 2026 Arabic HTR competition with a Character Error Rate (CER) of 0.0790 and a Word Error Rate (WER) of 0.2440.
- Robust to Historical Documents: Trained with progressive fine-tuning, domain-matched data augmentation, and advanced checkpoint merging to handle the complexities of historical Arabic texts.
- Vision-Language Integration: Utilizes a unified encoder-decoder architecture, with a focus on both vision and language understanding for accurate transcription.
Good for
- Digitizing Historical Arabic Manuscripts: Ideal for researchers and institutions working with old, handwritten Arabic documents.
- High-Accuracy Arabic OCR: Provides state-of-the-art performance for challenging Arabic HTR tasks.
- Applications Requiring Robustness: Suitable for scenarios where document degradation and diverse handwriting styles are prevalent.
For a live demo of the Baseer model for document extraction, visit baseerocr.com.