| To address the issues that general large language models generate factual hallucinations when applied to the vertical domain of Huizhou archways—a domain characterized by deep clan ethics and complex structural forms—and considering that full-parameter fine-tuning incurs high costs and may lead to the forgetting of general capabilities, while relying solely on prompts fails to stably internalize long-tail knowledge, this study takes Huizhou archways as the research object to construct an instruction fine-tuning dataset covering architectural structures and cultural connotations based on the Self-Instruct paradigm. On this basis, a Low-Rank Adaptation fine-tuning strategy applied to all linear layers is adopted, and rank parameters are optimized to enhance the model"s capability in modeling complex components and historical semantics, thereby mitigating the small-sample overfitting phenomenon. Five mainstream base models are evaluated on the self-constructed Huizhou archway dataset. Experimental results demonstrate that the accuracy of all models improved after fine-tuning; notably, the BLEU-4 score of the base model Qwen2.5-7B increased from 5.65 to 36.72. Regarding the BERTScore-F1 semantic evaluation metric, which reflects the deep comprehension ability of the models, Qwen2.5-7B achieved an absolute improvement of 1.11 percentage points, with the remaining models also exhibiting a consistent upward trend. This demonstrates that the proposed fine-tuning strategy mitigates, to a certain extent, the shortcomings of traditional n-gram metrics that focus solely on literal matching. Furthermore, while enhancing the question-answering accuracy of general LLMs regarding Huizhou archways, this study also provides a highly generalizable, lightweight fine-tuning strategy for models with specific architectures, offering certain reference value for the digital protection of ancient architecture and the application of vertical large models. |
