Objective Addressing the challenges of mixed tabular-textual data question answering in the rail transit field, a large language model based on a dynamic chain reasoning framework （hereinafter referred to as the DCLL （dynamic chain large language） model） is proposed, aiming to tackle the collaborative comprehension of structured tabular data and unstructured textual information and to provide frontline engineers with intelligent question-answering tools for scenarios such as equipment maintenance and fault diagnosis.

Method The proposed DCLL model achieves semantic collaboration and precise parsing of complex data by constructing a multi-stage reasoning mechanism encompassing scenario analysis, information filtering, logic construction, intelligent verification, and engine execution. Instruction fine-tuning is performed using the DeepSeek R1 model, combining the field-specific professional dataset TRTDD （transportation table domain dataset） to complete the deep integration of field knowledge, while employing LoRA （low-rank adaptation）-based parameter-efficient adaptation technology to reduce training costs. Based on the HybridQA （hybrid table and text question answering） benchmark dataset, the TATQA （financial table and text question answering） dataset, and TRTDD, a comparative analysis is conducted to evaluate the performance differences between the DCLL models with parameter sizes of 32B and 70B and general-purpose large language models such as GLM4-9B-Chat, Qwen2.5-72B-Instruct, and Llama-3.3-70B-Instruct.

Result & Conclusion  Experimental results demonstrate that the proposed DCLL models with parameter sizes of 32B and 70B significantly outperform general-purpose large language models in both EM （exact math） and F₁-score metrics. This approach validates the superiority of the dynamic chain reasoning framework in complex operational scenarios, providing an efficient multimodal data question-answering solution for the rail transit field.