Large language models (LLMs) have considerable potential for application in the field of automotive maintenance and servicing thanks to their powerful generalization capabilities. This study addresses the current shortage of high-quality structured fine-tuning data in this field by constructing the AMCK (Automotive Maintenance and Care Knowledge Dataset) for fine-tuning large language models. The data sources are classified into three types. First, automotive maintenance information requests posted by real users and question-and-answer data related to automotive maintenance scenarios were collected from domestic automotive websites (such as the AutoHome and the DongCheDi); noise was filtered out and semantic content was refined to ensure the validity of the samples. Second, the publicly available AutoMaster professional corpus in the field of automotive engineering was integrated, which includes knowledge on car brand types, maintenance manuals, maintenance process standards. Third, data augmentation techniques were used to expand the dataset, incorporating more diverse data. In the end, 43,962 high-quality samples were selected and stored in JSON format, comprising three attributes: instruction, input, and output. The AMCK dataset can be used for fine-tuning LLMs in automotive maintenance and repair tasks. Its rich domain-specific knowledge samples enables LLMs to achieve more accurate semantic understanding and more reliable decision-making reasoning in automotive services, thereby enhancinag the level of intelligent automotive services.
Wu et al. (Sun,) studied this question.