%0 Journal Article
%@ 1438-8871
%I JMIR Publications
%V 27
%N 
%P e70481
%T How to Design, Create, and Evaluate an Instruction-Tuning Dataset for Large Language Model Training in Health Care: Tutorial From a Clinical Perspective
%A Nazar,Wojciech
%A Nazar,Grzegorz
%A Kamińska,Aleksandra
%A Danilowicz-Szymanowicz,Ludmila
%+ Department of Allergology, Faculty of Medicine, Gdańsk Medical University, Smoluchowskiego 17, Gdansk, 80-214, Poland, 48 585844300, wojciech.nazar@gumed.edu.pl
%K generative artificial intelligence
%K large language models
%K instruction-tuning datasets
%K tutorials
%K evaluation framework
%K health care
%D 2025
%7 18.3.2025
%9 Tutorial
%J J Med Internet Res
%G English
%X High-quality data are critical in health care, forming the cornerstone for accurate diagnoses, effective treatment plans, and reliable conclusions. Similarly, high-quality datasets underpin the development and performance of large language models (LLMs). Among these, instruction-tuning datasets (ITDs) used for instruction fine-tuning have been pivotal in enhancing LLM performance and generalization capabilities across diverse tasks. This tutorial provides a comprehensive guide to designing, creating, and evaluating ITDs for health care applications. Written from a clinical perspective, it aims to make the concepts accessible to a broad audience, especially medical practitioners. Key topics include identifying useful data sources, defining the characteristics of well-designed datasets, and crafting high-quality instruction-input-output examples. We explore practical approaches to dataset construction, examining the advantages and limitations of 3 primary methods: fully manual preparation by expert annotators, fully synthetic generation using artificial intelligence (AI), and an innovative hybrid approach in which experts draft the initial dataset and AI generates additional data. Moreover, we discuss strategies for metadata selection and human evaluation to ensure the quality and effectiveness of ITDs. By integrating these elements, this tutorial provides a structured framework for establishing ITDs. It bridges technical and clinical domains, supporting the continued interdisciplinary advancement of AI in medicine. Additionally, we address the limitations of current practices and propose future directions, emphasizing the need for a global, unified framework for ITDs. We also argue that artificial general intelligence (AGI), if realized, will not replace empirical research in medicine. AGI will depend on human-curated datasets to process and apply medical knowledge. At the same time, ITDs will likely remain the most effective method of supplying this knowledge to AGI, positioning them as a critical tool in AI-driven health care.
%R 10.2196/70481
%U https://www.jmir.org/2025/1/e70481
%U https://doi.org/10.2196/70481