CPGA-BOT: A Customized Power Grid Assistant chatBOT Fine-Tuning in Large Language Model
Qing Wang, Ziyin Zhou · 2024
Based on Large Language Models (LLMs), AI chatbots are widely utilized in people's daily life. However, deploying general LLMs directly into domain-specific intelligent chatbots poses challenges due to their lack of training on domain-specific datasets, hindering the ability to effectively play roles within particular domains. Taking the Power Grid (the State Grid Corporation of China, SGCC) domain as an example, existing AI chatbots based on LLMs often struggle with explanations of Power Grid domain-specific terminology and providing unsatisfied responses during interactions concerning Power Grid related inquiries. Addressing these two problems, we propose a novel approach, a customized AI chatbot named CPGA-BOT that customized for Power Grid services. We collect domain-specific datasets that encompasses Power Grid related terminology and residential electricity consultation services, conducts training and fine-tuning on various LLMs. By comparing and evaluating the fine-tuned models with the native, unfine-tuned models, we observe a significant enhancement in accuracy concerning Power Grid-domain knowledge question-answering and notable improvements over the native LLMs.