Fine-tuning a Large Language Model for the Indian Legal System

This paper introduces an application based on a large language model customized for the Indian legal system, leveraging the LLama 3.1 8B foundational model. The model is pre-trained on a diverse corpus of legal texts and subsequently fine-tuned with curated Indian legal data to enhance accuracy and contextual relevance in legal responses. Advanced techniques such as Low-Rank Adaptation and Quantized Low-Rank Adaptations are employed to optimize the model’s efficiency while minimizing computational costs during fine-tuning. Pruning, as a compression method is utilized to enhance the model’s performance further and enable its deployment in resource-constrained environments. Additionally, the Retrieval Augmented Generation module is strategically implemented for document-specific queries, ensuring contextually accurate responses when processing legal documents. This research work is backed up by extensive experiments measuring the effectiveness across many precision metrics. The application is also tested against HaluEval, a Hallucination Evaluation Benchmark for factual reliability, and has demonstrated significant improvements in the model’s effectiveness as a resource for the legal domain. This AI-driven tool is the first step in simplifying the legal advisory services and decision-support systems in the Indian judiciary. It also goes a long way in enhancing the legal services experience.