ABSTRACT As large language models (LLMs) continue to expand, their effective adaptation to specialized fields remains a critical challenge. This work presents an initial step toward the development of HydroLLM, a domain-specific LLM for hydrology. We construct a dataset of approximately 8,800 hydrology-focused question–answer pairs, each with a supporting context passage drawn from textbooks and scientific articles. The dataset includes four instructional formats: multiple-choice, true/false, fill-in-the-blank, and open-ended. Using this corpus, we fine-tune several LLMs of varying type and scale – from compact (1.5B) to large (32B) parameter counts using parameter-efficient LoRA (low-rank adaptation) methods. Our methodology compares different fine-tuned models and evaluates performance using accuracy and cosine similarity metrics across task types. Results show that the 8B-DeepSeek-Llama variant achieved the strongest overall performance, while the 32B model overfitted and the 1.5B model underperformed – demonstrating that larger size is not always advantageous and highlighting the need to match model capacity to dataset size. This work demonstrates that effective domain adaptation requires careful consideration of architecture, parameter count, and task complexity. By establishing performance and identifying the limits of current fine-tuning approaches, we took a concrete step toward building HydroLLM as a robust, domain-specific language model for hydrological analysis and decision support.
Kizilkaya et al. (2025) studied this question.