Basic study on applying QLoRA to large language models for generating explanations of dam operation support AI

Abstract

In recent years, dam operation has become more advanced and complex due to increasingly severe rainfall and environmentally considerate discharge practices. Against the backdrop of a shortage of experienced engineers, expectations are growing for AI-based operational support. In this context, operational support using large language models (LLMs) is promising; however, concerns remain regarding the assurance of faithfulness and validity, including the presence of hallucinations, and these challenges are particularly significant for specialized and atypical tasks such as dam operation. In this study, we constructed a synthetic dataset consisting of forecast rainfall and dam operation plans, and examined changes in explanation quality through the stepwise introduction of retrieval-augmented generation (RAG) and QLoRA. The results confirmed a tendency toward a certain degree of improvement in the quality of operational explanations when QLoRA was used in combination. Overall, these findings suggest that additional training with QLoRA may effectively address domain-specific expressions and judgments that are difficult for RAG alone to capture.