Adapting General Purpose Large Language Models for Domain Specific Tasks

Abstract

The emergence of Large Language Models (LLMs), such as OpenAI s ChatGPT, has introduced innovative ways to leverage computer-based systems for addressing specific problems across various domains. For example, LLM-based agents can assist in drafting responses to business emails, summarizing documents, and performing tasks that require human intelligence. Their intuitive interfaces, utilising natural language, have made them a preferred option for interacting with computers. However, LLMs face several challenges, including hallucination, the provision of incorrect responses based on outdated information, and limited support for non-English languages. newlineThis thesis explores various methods to adapt general-purpose LLMs for domain-specific, specialised tasks. First, it proposes adaptation techniques to enhance the capabilities of general-purpose LLMs for Marathi language processing tasks, such as sentiment analysis, paraphrasing, and topic mapping. Second, it suggests techniques to deploy LLMs to assist students and teachers in evaluating examinations, generating assessments, and providing student support. newlineThe thesis evaluates the effectiveness of various adaptation techniques, including prompt-based training, Retrieval-Augmented Generation, and fine-tuning, for specialised tasks. These techniques are assessed using standard metrics such as accuracy and recall, and the end user s satisfaction with the responses provided. The findings indicate that a general-purpose LLM s capabilities can be significantly extended by selecting an appropriate adaptation technique. Moreover, the enhanced model performs better in specialised tasks than its general-purpose counterpart.