Automating construction contract question answering using large language model and fine-tuning

Effective contract management is a critical factor in the success of construction projects. However, current contract management practices often rely heavily on expert experience, suffer from low efficiency, and lack convenient access to domain knowledge, which makes them inadequate for meeting the practical demands of construction projects. Large language models (LLMs), due to their powerful natural language understanding and generation capabilities, have demonstrated significant advantages in semantic comprehension and information extraction, offering a new approach to addressing the aforementioned problems. However, considering issues such as the privacy of construction contract data and the cost of model deployment, general-purpose LLMs are difficult to implement directly within construction companies. This paper proposes a domain adaptation method that combines supervised fine-tuning (SFT) with reinforcement learning (RL), enabling low-cost development of a contract Question Answering (QA) model tailored to the construction domain by fine-tuning a small open-source model. To support this, we constructed a high-quality QA dataset based on the FIDIC standard contract conditions and related interpretive materials, and conducted model training and evaluation. The results show that the fine-tuned model achieves performance comparable to general-purpose LLMs across multiple key dimensions. We also demonstrate a data construction framework that provides a transferable strategy for datasets generation in professional domains where QA data is scarce.