Optimising LoRA for Fine-Tuning Lightweight LLM Expert Agents in Cross-Regional Collaboration in the Construction Industry: Evidence from the Greater Bay Area

Under the “One Country, Two Systems, Three Legal Jurisdictions” framework, cross-regional collaboration in the construction industry of the Guangdong–Hong Kong–Macao Greater Bay Area (GBA) is more complex than other urban agglomerations, and this makes traditional methods relying on manual analysis of influencing factors inefficient and subjective. While existing large language models (LLMs) can meet the needs of intelligent applications, they lack specific domain knowledge. Considering the intelligent advantages of LLMs, this research proposes a lightweight expert agent through providing an optimised low-rank adaptation (LoRA) model SVDSR-LoRA, integrating domain knowledge using the fine-tuning method. Experiments on the 1.5b lightweight base model of qwen2.5 and deepseek-r1 show that the proposed SVDSR-LoRA training method can increase the mid-term convergence speed by 36