Fine-tuning Large Language Models with Human-inspired Learning Strategies in Medical Question Answering
Abstract: Training Large Language Models (LLMs) incurs substantial data-related costs, motivating the development of data-efficient training methods through optimised data ordering and selection. Human-inspired learning strategies, such as curriculum learning, offer possibilities for efficient training by ordering data according to common human learning practices. Despite evidence that curriculum learning improves performance of natural language understanding tasks in fine-tuning LLMs, its application to domain-specific question-answering remains underexplored. In this work, we comprehensively examine the effectiveness of human-inspired learning strategies for fine-tuning LLMs in medical question answering. Our work complements previous studies by extending the evaluation to non-curriculum-based learning across multiple language models, using both human-defined and automated data labels. Our results show moderate impact in using human-inspired learning strategies for fine-tuning LLMs, with maximum accuracy gains of 1.77\% per model and 1.81\% per dataset. However, the effectiveness of these learning strategies varies significantly across different model-dataset combinations, suggesting caution in generalising human-inspired strategies for fine-tuning language models. We also find that curriculum learning using LLM-defined question difficulty outperformed human-defined difficulty, highlighting the potential of using model-generated metrics in optimal curriculum design.
Paper Type: Long
Research Area: Special Theme (conference specific)
Research Area Keywords: Efficiency in Model Algorithms, Training, and Inference, Efficient/Low-Resource Methods for NLP, Question Answering
Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency
Languages Studied: English
Submission Number: 5255
Loading