surfiniaburger/unsloth_finetune_ocr_arabic

Glimpse: Multi-Modal OCR for RTL Scripts (Arabic & Persian)

Glimpse is a high-precision OCR model specialized for Right-to-Left (RTL) languages. It was developed as part of the ERNIE AI Developer Challenge 2025 to solve the "Script Bias" typically found in general-purpose vision models.

Model Details

• Developed by: surfiniaburger
• License: apache-2.0
• Finetuned from model: unsloth/PaddleOCR-VL
• Architecture: ERNIE-4.5-0.3B Vision-Language Model

📊 Performance Metrics

After 500 steps of fine-tuning, Glimpse achieved the following results on unseen RTL text lines:

• Character Error Rate (CER): 6.97% (Reduced from ~59% baseline)
• Validation Loss: 0.25 (Near-perfect convergence)

📚 Dataset Credit

This model was trained using the Persian & Arabic Text-Line Image OCR Dataset (Medium) curated by Mohammad Reza Hajesmaeili.

• Dataset Source: mohajesmaeili/Persian_Arabic_TextLine_Image_Ocr_Medium
• Content: Diverse text-line images containing complex ligatures and various Arabic/Persian fonts.

🛠 Training Tools

This model was trained 2x faster using Unsloth and Huggingface's TRL library. We utilized QLoRA (4-bit) to ensure efficient memory usage while maintaining high-fidelity weights.