Dataset for "LLM-Driven Diagnostic Reasoning for Distributed Hydrological Model Calibration" (MADR, Huai River Basin)

This dataset accompanies the paper "LLM-Driven Diagnostic Reasoning for Distributed Hydrological Model Calibration", which introduces MADR (Multi-Agent Diagnostic Reasoning), an LLM-driven framework for calibrating distributed hydrological models in nested-basin systems. Study area: catchment upstream of the Bengbu gauging station, Huai River Basin, China (121,330 km²), with five nested gauging stations, Xixian (XX), Wangjiaba (WJB), Zhoukou (ZK), Jiangjiaji (JJJ), and Bengbu (BB). Hydrological model: VIC 5.1.0 coupled with the Lohmann routing model. Simulation period 1980 to 1996 (1980 spin-up, 1981 to 1990 calibration, 1991 to 1996 validation). The record contains four archives plus a top-level README: inputs.zip, model-ready VIC and Lohmann routing inputs for all five sub-basins (derived from CMFD, HWSD, AVHRR, and SRTM; no raw source datasets are redistributed) MADR.zip, main MADR results, 10 independent runs each with Gemini 3 Flash Preview and Claude Sonnet 4.6 ablations.zip, the SADR single-agent ablation (10 runs per LLM), MADR without the Exploration Stage (3 runs), and MADR blank (3 runs) benchmark.zip, SCE-UA and NSGA-III baselines with driver scripts and outputs Observed daily streamflow at the five stations is not included. It was digitized from the hydrological yearbooks of the Huai River Basin Water Resources Commission, whose data policy does not permit public redistribution. The file inputs/obs/README.md inside inputs.zip documents the exact layout and format so that users with yearbook access can rebuild the directory and run the code unchanged. Code repository: https://github.com/jovequli/MADR (private during peer review, public upon acceptance). See README.md in this record for the full dataset layout, file formats, and reproduction instructions.