Legalese-Simplifier: Fine-Tuning Llama 3.2 for Legal Accessibility

A professional-grade LLM fine-tuning project that transforms dense, archaic legalese into plain, actionable English using QLoRA and Llama 3.2-3B.

🚀 Project Overview

Legal documents are often inaccessible to the general public. This project demonstrates how to specialize a general-purpose Large Language Model for the legal domain. By fine-tuning on a curated dataset of contract clauses, the model learns to maintain legal intent while drastically improving readability.

Key Technical Features:

• Base Model: Llama 3.2-3B-Instruct
• Fine-tuning Method: QLoRA (4-bit Quantized Low-Rank Adaptation)
• Framework: Unsloth (Optimized for 2x faster training & 70% less VRAM)
• Dataset: 400+ high-quality pairs of complex legal clauses vs. simplified equivalents.

---

📂 Project Structure

legalese-simplifier/
├── .gitignore               # Prevents large model weights from bloating the repo
├── README.md                # Project documentation
├── requirements.txt         # Environment dependencies
├── data/
│   └── legal_train.jsonl    # Specialized legal dataset (JSONL format)
├── notebooks/
│   └── Training_Llama3.ipynb# Colab-ready training pipeline
├── model_cards/
│   └── adapter_config.json  # PEFT/LoRA configuration details
└── app/
    └── main.py              # Gradio-based inference UI



For my project, I have uploaded the ZIP File of the Model that I trained in this Repo, If you would like to run this Repo, Please Unzip the folder and populate the required path to the model in main.py code.

I hope you like this project !!