AI Summary & Key Takeaways
AI InsightWe observed balanced precision (0.8826) and recall (0.8836), with an overall classification accuracy of 88.3\%.
📌 3 Key Takeaways
Abstract
Occupational health and safety management requires systematic hazard identification and comprehensive risk assessment. Traditional approaches rely on manual processes that can be time-consuming and inconsistent across evaluators. In this study, we present an automated artificial intelligence system for hazard identification and risk assessment using fine-tuned large language models. We leverage Qwen3-32B as the base model. This model was adapted via Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, to acquire domain-specific knowledge of workplace hazards, risk assessment methodologies, and safety control measures. We trained our model on a dataset of hazard scenarios prepared by occupational safety experts. Each sample contains structured JSON outputs, including the hazard name, description, probability score, severity score, and control measures. Our evaluation on a held-out test set shows promising results. The fine-tuned model achieved an F1-score of 0.8830, which represents a 32.1\% improvement over the base model. We observed balanced precision (0.8826) and recall (0.8836), with an overall classification accuracy of 88.3\%. One particularly noteworthy finding is that no extreme misclassifications occurred between the high- and low-risk categories. This pattern indicates conservative and safety-conscious predictions. Our findings demonstrate that large language models can be effectively adapted to specialized occupational safety tasks through parameter-efficient fine-tuning. This approach offers significant potential as a decision support tool to enhance consistency and efficiency in safety management practices.