Title |
Development of AI Prototype for Generating Construction Safety Guidelines Through Fine-Tuning of Large-Scale Language Model |
Authors |
Lee, Jungwon ; Ahn, Seungjun |
DOI |
https://dx.doi.org/10.6106/KJCEM.2025.26.2.020 |
Keywords |
Construction Safety; Large Language Model; Fine-tuning; Construction Safety Guideline; LoRA |
Abstract |
In small to medium-sized construction sites, many accidents occur due to non-compliance with safety regulations, some of which lead to fatalities. Against this backdrop, this research aims to develop a construction safety guideline chatbot that generates appropriate safety guidelines based on job conditions and the work environment using a large-scale language model. This study created a question-answer dataset based on construction safety guidelines published by the Korean Occupational Safety and Health Agency and fine-tuned the language model using this dataset. To efficiently manage computational resources during the fine-tuning process, PEFT (Parameter-Efficient Fine-Tuning) and LoRA (Low Rank Adaptation) methods were employed. The fine-tuned model was designed to provide accurate and specific responses in delivering construction safety information. Validation of the fine-tuned model was conducted using qualitative and quantitative evaluation metrics, comparing its responses with those of large-scale language models like GPT-4, Palm2, and KoAlpaca. The fine-tuned model provided more specific and accurate information on construction safety knowledge than these large-scale language models. Additionally, in terms of quantitative evaluation, the fine-tuned model scored 7.5% higher on the BLEU score and 7.21% higher on the BERT Similarity score compared to the GPT-4 model. These scores demonstrate that the fine-tuned language model can significantly contribute to construction site safety management and is expected to respond effectively and promptly to various situations and problems that may arise in construction site environments. |