| Property | Value |
|---|---|
| Base Model | Qwen2-VL-2B-Instruct |
| Task | Arabic OCR |
| Dataset Size | 5000 records |
| License | Follows Qwen2 VL terms |
| Model URL | huggingface.co/NAMAA-Space/Qari-OCR-0.1-VL-2B-Instruct |
What is Qari-OCR-0.1-VL-2B-Instruct?
Qari-OCR is a specialized Arabic Optical Character Recognition model that represents a significant advancement in Arabic text extraction technology. Fine-tuned from the Qwen2-VL-2B-Instruct base model, it achieves remarkable accuracy with a 93.2% word accuracy rate and 98.1% character accuracy rate, substantially outperforming existing solutions.
Implementation Details
The model leverages the vision-language capabilities of Qwen2 VL architecture, fine-tuned on a carefully curated dataset of 5000 Arabic text samples. The implementation focuses on full-page text recognition and demonstrates exceptional performance metrics compared to traditional OCR solutions like Tesseract and EasyOCR.
- Word Error Rate (WER): 0.068
- Character Error Rate (CER): 0.019
- BLEU Score: 0.860
- 95% reduction in WER compared to base model
- 84% lower WER than Tesseract OCR
Core Capabilities
- Full-page Arabic text recognition
- Support for multiple standard Arabic fonts (Almarai, Amiri, Cairo, Tajawal, NotoNaskhArabic)
- Optimized for 16px font size
- High accuracy in printed document processing
- Efficient processing of complex layouts
Frequently Asked Questions
Q: What makes this model unique?
The model's exceptional performance metrics and specialized focus on Arabic text make it stand out. It achieves a 95% reduction in Word Error Rate compared to the base model and significantly outperforms traditional OCR solutions.
Q: What are the recommended use cases?
The model is ideal for processing printed Arabic documents with standard fonts at 16px size. It's particularly effective for full-page text recognition in business documents, academic papers, and printed materials. However, it's not suitable for handwritten text or documents with heavy use of diacritics.