AbdoTarek/Baseer-OCR-V1.0

Baseer OCR V1.0 - Arabic Document OCR Model

Baseer OCR V1.0 is a fine-tuned version of Qwen/Qwen2-VL-2B-Instruct specifically designed for Optical Character Recognition (OCR) on complex Arabic legal documents.

🎯 Model Overview

Attribute	Value
Base Model	Qwen2-VL-2B-Instruct
Fine-tuning Method	LoRA (Rank: 48)
Training Framework	LlamaFactory
Model Type	Vision-Language Model (VLM)
License	Apache 2.0
Release Date	March 2025

📋 Model Description

Baseer OCR (بصير - meaning "reader" in Arabic) is specialized for extracting text from Arabic legal documents with complex layouts. This model builds upon Qwen2-VL's strong vision-language capabilities and is fine-tuned to handle:

• 📜 Arabic Legal Documents - Contracts, court documents, official papers
• 🔤 Complex Arabic Typography - Various fonts, sizes, and styles
• 📄 Multi-column Layouts - Documents with complex structuring
• 🖼️ Mixed Content - Documents with tables, stamps, and annotations
• 🔍 Handwritten & Printed Text - Both typed and handwritten Arabic

Key Features

• JSON Structured Output - Returns extracted text in structured JSON format
• Arabic Language Support - Optimized for Arabic script recognition
• Vision-Language Understanding - Combines visual understanding with language generation
• Instruction Following - Responds to detailed extraction prompts

🚀 Quick Start

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
from PIL import Image

# Load the model
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "AbdoTarek/Baseer-OCR-V1.0",
    torch_dtype="auto",
    device_map="auto"
).eval()

processor = AutoProcessor.from_pretrained("AbdoTarek/Baseer-OCR-V1.0")

# Prepare the image
image_path = "path_to_your_arabic_document.jpg"
image = Image.open(image_path)

# Create the prompt
prompt = """Extract ALL visible text from the document image.

Return the result strictly as JSON with this structure:

{
 "subject": "",
 "keywords": [],
 "full_text": ""
}

Rules:
- Do not repeat lines.
- Preserve original order of text.
- Do not add explanations."""

# Prepare messages
messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "You are a helpful assistant."}]
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": prompt}
        ]
    }
]

# Process and generate
text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
image_inputs, _ = process_vision_info(messages)

inputs = processor(
    text=[text],
    images=image_inputs,
    padding=True,
    return_tensors="pt"
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(**inputs, max_new_tokens=1024)
    output_text = processor.batch_decode(
        output_ids[:, inputs.input_ids.shape[1]:],
        skip_special_tokens=True
    )[0]

print(output_text)

📊 Training Details

Dataset

• Training Samples: Custom Arabic legal documents dataset
• Data Format: ShareGPT format with image paths and OCR annotations
• Preprocessing: Images resized to 512x512 pixels

Training Configuration

Parameter	Value
Stage	SFT (Supervised Fine-tuning)
Fine-tuning Type	LoRA
LoRA Rank	48
LoRA Dropout	0.05
LoRA Target	all
Learning Rate	1e-4
Epochs	8
Batch Size	1
Gradient Accumulation	32
Warmup Ratio	0.1
Scheduler	Cosine
Precision	BF16

Hardware Requirements

• GPU: Recommended 16GB+ VRAM (tested with ~8GB)
• RAM: 16GB+ system RAM
• Storage: 5GB+ for model files

💻 Usage Examples

Example 1: Extracting Text from Legal Contract

# Input: Image of Arabic legal contract
# Output: Structured JSON with subject, keywords, and full text

Example 2: Batch Processing

# Process multiple document images in a directory
import os
from glob import glob

image_files = glob("documents/*.jpg")
for img_path in image_files:
    # Process each image...

🔬 Model Performance

The model is optimized for Arabic legal document OCR and provides:

• High accuracy on printed Arabic text
• Structured JSON output for easy integration
• Keyword extraction for document classification
• Subject identification for document categorization

📝 Prompt Templates

The model responds well to detailed prompts. Recommended prompt structure:

Extract ALL visible text from the document image.

Return the result strictly as JSON with this structure:

{
 "subject": "",
 "keywords": [],
 "full_text": ""
}

Rules:
- Do not repeat lines.
- Preserve original order of text.
- Do not add explanations.

📚 Related Models

• Qwen/Qwen2-VL-2B-Instruct - Base model
• AbdoTarek/ocr-models-Qwen2-VL-2B-Instruct-V3.0 - LoRA checkpoints

📄 License

This model is released under the Apache 2.0 license.

🙏 Acknowledgments

• Qwen Team for the base Qwen2-VL model
• LlamaFactory for the fine-tuning framework
• Hugging Face for the model hub infrastructure

📧 Contact

• Author: Abdelrhman Tarek
• HF ID: AbdoTarek
• Model Page: Baseer-OCR-V1.0

---

Made with ❤️ for Arabic OCR