Legal Case Analysis and Summarization System for Arabic 🔍

A comprehensive system for collecting, analyzing, and summarizing legal cases in Arabic using Large Language Models fine tuning and knowledge graph techniques.

📋 Table of Contents

• Overview
• System Architecture
• Key Features
• Prerequisites
• Installation
• Usage Pipeline
• Project Structure
• Performance
• Created Knowledge Graph

🔍 Overview

This project provides an end-to-end solution for processing legal cases, from data collection to serving structured summaries via an API. The system uses state-of-the-art language models, specifically a fine-tuned LLaMA model, to generate comprehensive case summaries and builds a knowledge graph for advanced legal analysis.

🤗 Hugging Face Model

The fine-tuned model is publicly available on the Hugging Face Hub: ahmadsakor/Llama3.2-3B-Instruct-Legal-Summarization

🏗️ System Architecture

The system consists of five main components:

1. Data Collection & Preparation: Automated crawling and OCR processing of legal documents
2. Model Fine-tuning: Custom LLaMA model adaptation for legal summarization
3. Evaluation Framework: Comprehensive metrics for model assessment
4. Knowledge Graph: RDF-based graph database for legal case analysis
5. Inference API: FastAPI service for generating case summaries

✨ Key Features

• Automated legal case collection from the Moroccan Judicial Portal
• Advanced OCR processing with error correction
• Fine-tuned LLaMA model for Arabic legal text
• Comprehensive evaluation framework
• RDF-based knowledge graph
• FastAPI-based inference service
• Multi-language support (Arabic/English)

💻 Prerequisites

• Python 3.10 or higher
• CUDA-capable GPU (12GB+ VRAM recommended)
• 32GB+ RAM recommended
• 50GB+ storage space
• Internet connection
• GraphDB instance (for knowledge graph)

🚀 Installation

Each component has its own dependencies. To set up the complete system:

1. Clone the repository:

git clone [repository-url]
cd [repository-name]

2. Follow the instructions in each directory.

📝 Usage Pipeline

1. Data Collection & Preparation

2. Model Fine-tuning

3. Model Evaluation

4. Knowledge Graph Creation

5. Deploy Inference API

📁 Project Structure

├── fetch_data/          # Data collection and processing
├── fine_tuning/         # Model training and adaptation
├── evaluation/          # Performance assessment
├── knowledge_graph/     # Graph database creation
├── inference/           # API service
├── data/               # Data storage
└── logs/               # System logs

📈 Performance

The fine-tuned model shows significant improvements over the base model:

• Court Information: +4.58% BERTScore F1
• Legal Principles: +10.48% BERTScore F1
• Final Decision: +9.27% BERTScore F1
• JSON Validity: 99% accuracy

Full evaluation metrics available in the evaluation directory.

🌐 Created Knowledge Graph

Below is a sample of the created knowledge graph:

🙏 Acknowledgments

• Moroccan Judicial Portal for providing access to legal cases
• Contributors and maintainers of the OCR libraries used in this project
• Contributors and maintainers of 🤗 Transformers
• Contributors and maintainers of DeepSpeed
• Contributors and maintainers of PEFT (Parameter-Efficient Fine-Tuning)
• Contributors and maintainers of Weights & Biases for experiment tracking
• Contributors and maintainers of Moroccan Judicial Portal for providing access to legal cases
• Contributors and maintainers of Meta AI for the base LLaMA model.