A Comparative Evaluation of QLoRA and AdaLoRA for Parameter-Efficient Fine-Tuning of Large Language Models on Medical Textbook Question Answering

Abstract

Parameter-efficient fine-tuning methods have emerged as practical solutions for adapting large language models to specialized domains while minimizing computational overhead. This study presents a systematic comparison of two prominent approaches, QLoRA and AdaLoRA, for fine-tuning instruction-tuned language models on medical textbook question answering. We evaluated both methods using two backbone architectures, Llama-3-8B-Instruct and Qwen2-7B-Instruct, on a dataset comprising 6,500 question-answer pairs derived from 13 authoritative medical textbooks spanning diverse clinical and biomedical disciplines. Our experiments demonstrate that QLoRA consistently outperforms AdaLoRA under single-epoch training conditions, achieving validation perplexity values of 1.085 and 1.086 for Llama-3 and Qwen2, respectively, compared to AdaLoRA’s 1.125 and 1.169. These results correspond to relative validation loss reductions of 30.8% for Llama-3 and 47.5% for Qwen2 when using QLoRA over AdaLoRA. Both methods maintained comparable trainable parameter counts, approximately 167 million for Llama-3 and 161 million for Qwen2, representing roughly 3.5% of total model parameters. Our findings indicate that QLoRA provides more stable convergence behavior within limited training budgets, while AdaLoRA’s adaptive rank allocation mechanism may require extended training schedules to realize its theoretical advantages. These results offer practical guidance for deploying parameter-efficient fine-tuning in medical natural language processing applications where computational resources are constrained.