Thinking-enhanced LoRA finetuning large language models for gas pipeline leak resolution

Given the frequency of gas pipeline leaks due to aging and corrosion, robust emergency response plans are critical. Currently, such plans are predominantly static and lack structured knowledge, which limits their utility in supporting rapid decision-making and providing real-time professional guidance. To address this issue, this paper proposes a fine-tuned large language model for gas pipeline leak resolution (LLM4GPL). Specifically, to instantiate the Thinking-enhanced framework, we constructed a domain-specific dataset based on gas pipeline leak emergency plans. We employed data distillation techniques to generate QA pairs embedded with explicit reasoning steps (Chain-of-Thought), rather than simple fact-retrieval pairs. Subsequently, we utilized the Low-Rank Adaptation (LoRA) strategy to fine-tune a Large Language Model (LLM). In this process, we specifically injected trainable low-rank matrices into the model’s Attention mechanisms and Multi-Layer Perceptron (MLP) modules. This design enables the LLM to capture both the specialized knowledge and the multi-step cognitive logic required for emergency resolution. On the self-built dataset, the fine-tuned model attains the highest reported BLEU and ROUGE-L scores and the lowest test loss and perplexity among the evaluated configurations. An independent evaluation by three domain experts on 50 stratified test cases indicated higher plan quality for the proposed model than for the compared baselines, and GPT-4o scores corresponded closely with the expert ratings. All 648 source records originate from a single municipal gas enterprise; the reported results therefore establish internal performance on one organizational dataset and do not demonstrate external generalization or system-level operational reliability.