MEDFIT-LLM: Medical Enhancements through Domain-Focused Fine Tuning of Small Language Models
Aditya Karnam Gururaj Rao, Arjun Jaggi, Sonam Naidu · 2025
This research explores the efficacy of fine-tuning small Large Language Models (LLMs) for AI-based healthcare chatbots. We present a novel approach to dataset creation using synthetic data generation and fine-tuning of various LLMs. The study compares the performance of base models against their fine-tuned counterparts using a carefully curated dataset of healthcare-related questions and answers. Our methodology involves the use of phi 4 for synthetic data generation, supplemented with domain-specific questions derived from existing research. We employ LORA fine-tuning on MLX for four distinct models: Gemma 2 9B 4bit, LLama 3.2 3B Instruct, Mistral 7B, and Qwen2 7B Instruct 8 bit. The results demonstrate the potential of fine-tuning in enhancing the performance of LLMs in specialized healthcare contexts, contributing to the ongoing development of more accurate and reliable AI-powered healthcare chatbots. The code is available at https://github.com/adityak74/medfit-llm.