Abstract: Coal mine emergencies are characterized by high risk, complexity, and sudden occurrence, which makes it difficult for traditional rule-based or single-model approaches to provide rapid and accurate decision support. Existing methods also face limitations in semantic understanding, process coordination, and structured output. To address these issues, this study proposes an intelligent emergency management decision-support model that integrates accident reports and emergency response plans.A large-language-model-based multi-task fine-tuning framework, referred to as LLaMA-MT, is developed based on the LLaMA model to enhance intelligence throughout the emergency management process. Through dual-modal semantic modeling, accident descriptions are aligned and fused with relevant emergency plan content, enabling structured text generation across the full emergency lifecycle, including response, situation control, and recovery. This framework improves the model's domain adaptability and its ability to generate well-structured emergency decision outputs.Experimental results show that LLaMA-MT achieves a BLEU score of 42.3, a ROUGE-L score of 55.6, and a BERTScore of 0.802, demonstrating competitive performance compared with mainstream large language models. Further validation using virtual cases shows that LLaMA-MT can automatically generate structured emergency plans aligned with standard procedures, effectively supporting practical emergency management needs.