Comparison of Fine-Tuned Large Language Models for Structural Classification of Indian Legal Judgements

This study explores how fine-tuned Large Language Models (LLMs) can help automate the structural classification of Indian legal judgments—organizing them into meaningful sections such as facts, arguments, and verdicts. Using a supervised learning approach, several LLMs were fine-tuned on a labeled dataset sourced from Indian Kanoon, enabling the models to better understand the unique language and structure of Indian legal texts. We conducted experiments using several state-of-the-art models, such as BART (base and large), Flan-T5, Google-T5, Tiny Llama, Tiny-Mistral, and XLNet. These models were assessed under two settings: full fine-tuning and Low-Rank Adaptation (LoRA), a parameter-efficient method designed to reduce the computational overhead of model adaptation. Among the evaluated models, Tiny Llama achieved the highest performance, with a weighted average accuracy of 0.829, demonstrating its robustness and scalability despite its relatively small size. To the best of our knowledge, this work represents one of the earliest systematic attempts to employ domain-adapted LLMs for the structural classification of Indian legal documents, paving the way for improved efficiency within the Indian judicial framework.