CyFuture_Hackathon

This repository contains the data and code files required for building a Legal Contract Reviewer for english contracts using the CUAD dataset.

Dataset Used

This project leverages a subset of the Contract Understanding Atticus Dataset (CUAD) to automate legal contract analysis. CUAD is an expert-annotated dataset designed for high-precision clause extraction, featuring:

• 500+ contracts
• 13,000+ labeled clauses
• 41 specialized legal categories

By utilizing CUAD, this system enhances efficiency in contract review and risk assessment, helping users quickly identify key provisions.

For more details, visit CUAD.

Dataset Conversion for LLaMA Models

Since LLaMA models do not natively support direct ingestion of the CUAD dataset, the data has been transformed into an optimized format. The conversion process ensures compatibility and structured input handling for improved legal NLP performance.

• Converted Dataset: alpaca_format_for_llama.json
• Conversion Code: Implemented in data_conversion.ipynb
• Source Dataset: A subset of the CUAD (Contract Understanding Atticus Dataset), containing expert-annotated legal contract clauses.

This adaptation enhances contract clause extraction and facilitates efficient automated legal review using large language models.

📝 Legal Answering using CUAD + LoRA Fine-tuning(LegalAnsweringUsingCUAD.ipnyb)

This repository contains a Google Colab notebook to fine-tune a Large Language Model (LLM) on the CUAD dataset for legal contract understanding and question answering. We use Unsloth and LoRA adapters to efficiently train on long legal texts.

---

🚀 Overview of Steps

1️⃣ Installation

We install all required libraries including Unsloth.

2️⃣ Load and Prepare the Model (Unsloth)

We load a quantized LLaMA-3 model (8B) using Unsloth, with support for long context lengths.

3️⃣ Add LoRA Adapters

We apply LoRA adapters so that only a small subset of parameters are updated, reducing memory footprint.

4️⃣ Data Preparation

We structure CUAD examples using instruction-style prompts to make them suitable for fine-tuning as described earlier.

5️⃣ Train the Model

The model is fine-tuned using the formatted dataset.

KEY FEATURES

✅ Uses LoRA for memory-efficient fine-tuning

✅ Supports long legal documents with 13,500 token context

✅ 4-bit quantization for low resource usage

✅ Can be run on Google Colab