# QARI-OCR: High-Fidelity Arabic Text Recognition through Multimodal Large Language Model Adaptation

**Authors:** Ahmed Wasfy, Omer Nacar, Abdelakreem Elkhateb, Mahmoud Reda, Omar Elshehy, Adel Ammar, Wadii Boulila  
**arXiv:** [2506.02295](https://arxiv.org/abs/2506.02295)  
**Date:** 2025-06-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.02295v1/

## Key points

- Cursive nature
- Diacritical marks (tashkeel)
- Varied typography
- Focus on optimizing OCR for Arabic text
- Leading model: QARI v0.2 with impressive benchmarks
- Demonstrates proficiency in handling script complexities
- Resilience to optical degradation and accurate transcription from varied inputs
- Diacritics, ligatures, variant letterforms, etc.
- Structural document understanding
- Handwritten text recognition capabilities
- Significant improvement in Arabic OCR accuracy and efficiency
- Open-source models and datasets for further research opportunities

## Summary

, , , , 

In the realm of Optical Character Recognition (OCR), the complexities inherent in Arabic script have long posed challenges due to its cursive nature, diacritical marks (tashkeel), and varied typography. To address these challenges, a series of vision-language models known as Qari-OCR has been developed, with a focus on optimizing OCR specifically for Arabic text. Through iterative fine-tuning on specialized synthetic datasets, the leading model, QARI v0.2, has achieved impressive quantitative benchmarks with a Word Error Rate (WER) of 0.160, Character Error Rate (CER) of 0.061, and BLEU score of 0.737 on diacritically-rich texts. aside, qualitative analysis is essential to understand the practical capabilities of the model. Visual illustrations provided by Figure 2 showcase Qari-OCR's proficiency in handling various complexities inherent in Arabic script, supporting its strong quantitative performance. Moreover, the resilience of the model to optical degradation and its ability to accurately transcribe text from varied inputs were tested. Figure 3 demonstrates that Qari-OCR, particularly QARI v0.3 trained on more complex layouts, can robustly detect and transcribe Arabic text even from low-resolution images with small sizes and tightly cropped boundaries. The qualitative assessment further delves into the nuances of Arabic script that pose challenges for OCR systems such as diacritics (tashkeel), ligatures like Lam-Alif (*), variant letterforms, classical language structures, embedded punctuation and numerals, diverse orthographic forms of Hamza (ح), and features like Maddah. Additionally, an in-depth analysis reveals how Qari-OCR excels in structural document understanding and handwritten text recognition through models like QARI v0.3. This work signifies a significant improvement in Arabic OCR accuracy and efficiency while also providing all models and datasets as open-source resources to foster further research in this domain. Overall, the refined detailed summary emphasizes not only the  but also highlights the qualitative strengths of Qari-OCR in handling complex Arabic script intricacies with precision and robustness across various document layouts and image resolutions.