FireSeek-Sci : a large language model with fire science knowledge for smart building safety

Abstract

Fire science knowledge is fundamental to fire science education and the subsequent learning and practicing fire protection engineering. The fragmented knowledge and time-consuming information retrieval make it difficult to obtain reliable answers in time, which seriously affects learning and practical application. To address this challenge, this study builds an expert-level fire domain knowledge model, FireSeek-Sci, based on the advanced pretrained large language model (LLM). Firstly, a specific fire domain-knowledge (FDK) database was established covering key areas, including fundamentals of combustion, fire dynamics, fire protection engineering, etc., organized in 6750 Q&A pairs. Secondly, the base LLM was fine-tuned using FDK and achieved significant improvements with over 30 points in generative semantic accuracy. Thirdly, retrieval augmented generation (RAG) technology was further embedded with the fine-tuned fire-specific LLM, further enhancing the accuracy and richness of the generated corpus over 69 %. Based on these developments, an LLM-powered fire science knowledge chatbot named FireSeek-Sci was deployed within a user-friendly web interface. Internal testing with participants demonstrated positive evaluations, with an overall satisfaction score of 3.3 (1–5 scale) and 83 % of participants expressing willingness to use the chatbot again. To further evaluate its engineering applicability, FireSeek-Sci was applied in illustrative case studies to interpret AI-generated simulation outputs into natural-language safety assessments and recommendations, which showed good quantitative grounding, consistent reasoning, and responses that were closer to practical fire engineering judgment. In conclusion, FireSeek-Sci has strong potential as an intelligent assistant for fire science education, research, and building safety applications on fire risk and emergency management.