ChatTower: Tower Crane Expert System Based on ChatGLM

To address the limitations of general large language models (LLMs) in specialized fields, this paper investigates a tower crane expert system based on the open-source model ChatGLM. Recognizing the challenges of poor sample quality due to inadequate filtering in existing dataset construction methods, a dataset construction and screening algorithm for LLMs is proposed. The method includes two components: topic text segmentation and dataset filtering. The topic text segmentation algorithm determines the number of topics and segmentation points by evaluating the similarity between topic words and sentence keywords, combined with optimization. This approach reduces the proportion of text blocks containing multiple topics during Q&A sample generation, thus improving sample quality. The dataset filtering algorithm generates Q&A samples by setting construction standards and utilizing the Zhipu model. Sample evaluation criteria are established based on accuracy, comprehensiveness, and validity, and the 01 model is used to assess sample validity. Experimental results show that the topic text segmentation algorithm reduces the Pk error rate by 8.2% to 37.7% compared to classical algorithms. The dataset filtering algorithm improves precision by 9.4% to 89.2% compared to the case without the algorithm. These findings demonstrate the effectiveness of the proposed algorithms in improving the quality of the tower crane Q&A dataset and significantly enhancing the performance of the tower crane expert system.