Research on Construction Method of Fine-Tuning Dataset for LLMs for Electric Power Science Literature

Facing The adaptability of large language models (LLMs) in highly specialized texts of electric power science literature has become a key factor influencing the effect of engineering deployment. Aiming at the problem that general models have insufficient understanding of professional terms, this paper proposes a systematic construction method for fine-tuning datasets of large language models, which is adapted and optimized for typical tasks in the analysis of electric power science literature. The dataset covers academic papers, patent documents, and internal technical materials of enterprises. Through a unified data preprocessing, a multi-task labeling system, and standardized format operations, it supports tasks such as text classification, named entity recognition, question-answering generation, and summary generation. Fine-tuning experiments based on the Qwen model show that the fine-tuning dataset significantly improves the accuracy and generation quality of the model in professional text processing, providing high-quality data support and technical paths for promoting the integrated application of “electric power + AI”.