.png){kind=icon}
.png){kind=icon}
Return
FireKnowLLM: Leveraging expert-guided fine-tuning for reliable firefighting knowledge question answering
.png){kind=icon}
PRE
AI
Z
DOI：10.1016/j.displa.2026.103448.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Fire incidents remain a persistent global challenge, requiring accurate and safety-compliant firefighting knowledge to support fire prevention, emergency response, and regulatory supervision. However, generalpurpose large language models (LLMs) often fail to provide reliable and professional answers in the firefighting domain due to insufficient domain specialization and expert guidance. To address this issue, FireKnowLLM is proposed as a domain-adapted large language model for reliable firefighting knowledge question answering. The model is fine-tuned using a parameter-efficient Low-Rank Adaptation (LoRA) approach on a large-scale, expert-annotated firefighting question-answer dataset, enabling regulation-aligned knowledge acquisition. In addition, standardized instruction formats and structured response templates are designed to improve answer accuracy, completeness, and safety compliance. FireKnowLLM is evaluated through automatic metrics, expert human assessment, and safety-oriented adversarial testing. Experimental results demonstrate substantial improvements in professional accuracy and completeness over the base model, with competitive safety performance approaching leading commercial APIs. Notably, FireKnowLLM significantly reduces unsafe following behaviors compared to non-fine-tuned open-source alternatives.
Keywords:
Firefighting knowledge question answering
Qwen3-32B large language model
LoRA fine-tuning
Journal
IF:
3.4
Papers:
2.1K
Citations:
3.2K
Organization
No organization information available
Cited Papers