Virtual Assistance for Nepali Law

Abstract

Large-scale LLM models are powerful and efficient due to their training on vast datasets, enabling them to handle a wide range of generic queries. However, they often fall short when addressing queries related to specific topics, such as Nepali law. To overcome these limitations, we propose developing a small-scale instructiontuned LLM model tailored specifically for Nepali law, using a custom dataset. This chatbot will be deployable on any embedded system. We have fine-tuned the opensource LLM model, �meta-llama/Meta-Llama-3.1-8B-Instruct�, with our curated dataset using the LORA technique, achieving a BLEU score of 0.53 and a ROUGE score of 0.40 There is some tradeoff in accuracy due to the quality of our dataset. We utilized the Stanford Alpaca format for our instruction-tuned system in the response and instruction pairs. Our model is a compact, fine-tuned solution designed to serve as a smart reply bot, particularly focused on government rules, policies, services, and system-related public information for citizens.