Grid2Text Dataset Aligning Meteorological Grid Data with Expert Reasoning Chains

Abstract Accurate meteorological forecasting is fundamental to public safety and economic stability, but it heavily relies on numerical weather prediction (NWP) models, whose complex outputs are often inaccessible to non-experts. Although large language models (LLMs) show potential in translating structured data into natural language to automate forecast reporting, progress has been constrained by a critical scarcity of datasets that bridge the modality gap and capture the implicit reasoning of human experts. To address this, we present Grid2Text , a rigorously developed dataset aligning ERA5 meteorological grid features with expert-verified textual forecasts, encompassing explicit reasoning chains for temperature trends, wind vector transitions, humidity ranges, and precipitation types. Furthermore, we establish validation benchmarks derived from a human-in-the-loop workflow to ensure the physical consistency and logical accuracy of the data. We anticipate that this open-access dataset will catalyze the development of interpretable LLM approaches for weather forecast discussions, ultimately advancing the field of scientific text generation.