%0 Dataset
%T An AI-Ready Multimodal Dataset for Fine-Tuning Large Models in Road Ice and Snow Hazard Warning
%J National Cryosphere Desert Data Center
%I National Cryosphere Desert Data Center(www.ncdc.ac.cn)
%U http://www.ncdc.ac.cn/portal/metadata/e181e032-0e0b-42f6-bf3b-8a365b8c7cdc
%W NCDC
%R 10.12072/ncdc.llm.db7721.2026
%A LIU Jingqi
%A zhang zhi xing
%A Zhang Yaonan
%A REN Yanrun
%A Kang Jianfang
%K artificial intelligence;Disaster Prevention and Mitigation
%X Road ice and snow hazards in the Tianshan Mountains are jointly influenced by complex terrain, localized meteorological conditions, and road infrastructure, and are characterized by sudden onset, continuous evolution, and pronounced spatial variability. Existing datasets are primarily designed for isolated tasks, such as image-based road-surface recognition or structured risk assessment, and therefore provide limited support for the continuous operational chain from environmental perception and risk assessment to traffic control and information dissemination. To address this gap, this study developed an AI-Ready multimodal instruction-tuning dataset for road ice and snow hazard warning in cold mountainous regions. The textual data were derived from 6,496 publicly available road weather hazard records released by the Xinjiang Public Travel Information Service Network between March 2013 and January 2026. After data standardization, semantic review, and construction of a master event table, 5,536 valid events were retained and used to derive eight instruction-tuning tasks: fact extraction, risk assessment, traffic-control recommendation, public advisory generation, recovery-time assessment, emergency-response review, initial road-event report generation, and cross-source consistency checking. These tasks yielded 39,909 text instruction-tuning samples. The image data were subjected to perceptual deduplication, algorithm-based screening, and manual review, resulting in 808 road images covering six road-surface conditions: dry, snow-covered, ice-covered, snow-blowing, snow-melting, and wet. Single-image multimodal instruction-tuning samples were then constructed following a visible-evidence-first principle. For 264 snowmelt and wet-road images, deterministically generated scenario parameters for air temperature, precipitation, and wind speed were further incorporated, and reproducible icing-time supervision labels were generated using a fourth-order polynomial model. The