FireInvestLLM: Leveraging Large Language Models for Improved Fire Investigation Inquiries

Fire incidents occur frequently worldwide, and reliable fire investigations are crucial for uncovering the causes of fires, assessing damages, and improving prevention and control measures. To enhance the efficiency and accuracy of fire investigations, we propose an automated inquiry simulation pipeline based on LLMs, named FireInvestLLM. Specifically, we firstly performed Supervised Fine-Tuning on the model using our custom fire investigation instruction dataset, equipping it with domain knowledge. Secondly, we optimised the prompt design by retrieving historical similar case inquery templates, thus improving the model’s capabilities in role-playing, event comprehension and instruction execution. Finally, we enhance the dialogue flow through a question validation and self-reflection to ensure that the model maintains consistent professionalism and consistency across multiple rounds of dialogue. We utilized a custom fire incident dataset to evaluate our method. Experimental results show that FireInvestLLM significantly enhances the quality of inquiry dialogues in fire investigations, showcasing its substantial potential in practical applications and providing fire departments with an efficient AI assistant.