Construction and Evaluation of Large Language Models for the Cultural Tourism Domain via LoRA-Efficient Fine-Tuning

The cultural and tourism industry plays a crucial role in the digital economy, driving economic growth and fostering cultural exchange. However, existing intelligent methods predominantly rely on general-purpose models, which often lack domain-specific adaptation. To address this, this study explores the feasibility of applying large language models in vertical domains, with a focus on Dali, Yunnan, as the experimental site. Scenic spot data is collected from the Ctrip platform using web crawlers, and a fine-tuning dataset is constructed through automated Question and Answer (Q&A) generation. The DeepSeek-R1-Distill-Qwen-1.5B model is adopted as the base model. Using Low-Rank Adaptation (LoRA) parameter-efficient fine-tuning method and the LLaMA-Factory framework, the model is trained on 3A and above scenic spots, high-rated attractions, and representative travel routes. The fine-tuned model is then evaluated on tasks including self-awareness, domain-specific Q&A, and itinerary planning. The results indicate a reduction in training loss from 3.7 to 0.2444 in the initial phase, stabilizing at 0.0125 after convergence. In professional capability assessment, after 200 training epochs, the fine-tuned model demonstrates significant improvements in three aspects: self-awareness, scenic spot Q&A, and travel itinerary planning. It accurately interprets user queries and provides correct, expectation-aligned information, showing clear advantages over baseline models. This study offers a replicable approach for integrating AI into the cultural tourism sector, lowering the barrier to building intelligent industry solutions.