Alfaxad/vector-100k

VectorOS Vector 100k SimSat VLM数据集是一个高保真度的多模态指令数据集，专为地理空间流行病学任务中的视觉语言模型微调而设计。该数据集包含100,000个聊天式示例，这些示例源自30个关注区域（AOIs）的10,000个地理空间芯片。每个芯片包括一个1024x1024的四面板视觉包（包含SimSat Sentinel-2真彩色视图、假彩色NIR-红-绿视图、Mapbox卫星视图和对齐的证据叠加层）、侧边栏JSON文件（包含数值特征、来源路径等）以及目标风险瓦片JSON文件。数据集覆盖了登革热/伊蚊、疟疾/按蚊和血吸虫病三种疾病模块，并采用分层采样策略以确保模型学习有用的视觉和推理行为，而非仅记忆标签。数据来源包括SimSat Sentinel-2产品、Mapbox卫星图像、CHIRPS降雨数据、JRC全球地表水数据、ESA WorldCover、WorldPop人口数据、OpenStreetMap等。数据集旨在用于公共卫生监测支持，仅用于人口层面的决策支持推理，避免个体健康声明或实地验证。VectorOS Vector 100k SimSat VLM Dataset is a high-fidelity multimodal instruction dataset specifically designed for fine-tuning visual language models (VLMs) in geospatial epidemiological tasks. This dataset contains 100,000 conversational examples derived from 10,000 geospatial chips across 30 Areas of Interest (AOIs). Each chip includes a 1024x1024 four-panel visual package containing SimSat Sentinel-2 true-color view, false-color NIR-red-green view, Mapbox satellite view, and aligned evidence overlays, a sidebar JSON file with numerical features, source paths and other relevant information, and a target risk tile JSON file. The dataset covers three disease modules: dengue/Aedes, malaria/Anopheles, and schistosomiasis, and adopts a stratified sampling strategy to ensure that the model learns useful visual and reasoning behaviors rather than merely memorizing labels. Data sources include SimSat Sentinel-2 products, Mapbox satellite imagery, CHIRPS rainfall data, JRC Global Surface Water data, ESA WorldCover, WorldPop population data, OpenStreetMap and other resources. The dataset is intended for public health surveillance support, and is only used for population-level decision support reasoning, with no individual health claims or field validation permitted.