Trade-offs in Medical LLM Adaptation: An Empirical Study in French QA

This study investigates adaptation strategies for medical domain large language models (LLMs) in French question-answering (QA), comparing continual pretraining (CPT), supervised fine-tuning (SFT), and their combination across various model families and sizes. Results indicate that while CPT+SFT generally yields the best performance in multiple-choice QA, the improvements are minimal compared to SFT alone, which remains a cost-effective option; in open-ended QA, CPT enhances overlap metrics but can degrade generation quality with SFT. The findings offer practical guidelines for selecting adaptation methods in resource-constrained environments, emphasizing the potential for effective cross-lingual transfer from French to English benchmarks.