With massive training data and sufficient computing resources, large language models (LLMs) have demonstrated impressive capabilities. These models can rapidly respond to questions in almost all domains and are capable of retrieving, synthesizing, and summarizing information. The capabilities demonstrated by LLMs can enhance our livelihood and foster innovation. Nonetheless, in some professional domains, the focus is not only on response speed but also on higher requirements for response reliability. For example, in the medical domain, the reliability of information provided by the model poses a great risk to subsequent diagnosis and treatment, especially when the language is not English. In specific domains, domain-specific knowledge can be used to refine pre-trained LLMs to improve their performance in specific tasks. In this study, we aimed to build an LLM for epilepsy, called EpilepsyLLM. We constructed an epilepsy knowledge dataset in Japanese for LLM fine-tuning, and the dataset contained basic information on epilepsy, common treatment methods and drugs, and important notes on patients lives. Using the constructed dataset, we refined several different pre-trained models with supervised learning. In the evaluation process, we applied multiple metrics to measure the reliability of the LLMs output. The experimental results highlighted that the fine-tuned EpilepsyLLM can provide more reliable and specialized epilepsy responses.
Zhao et al. (Mon,) studied this question.