ChrisRPL/satellite-disruption-triage-v0

这是一个小型高质量辅助数据集，包含来自11个全球灾难事件的成对卫星图像（基线+当前），并带有结构化JSON分类输出，用于训练或评估视觉语言模型（VLMs）在宏观尺度民用破坏检测上的性能。数据集共120个示例（92个训练，28个评估），每个示例包含：1）灾难事件前的基线卫星图像；2）事件后的当前卫星图像；3）结构化JSON输出，包含动作（discard、defer、downlink_now）、类别（如earthquake_building_damage）、理由和归一化边界框。数据集来源于BRIGHT数据集，采用CC-BY-NC-4.0许可。

This is a small, high-quality auxiliary dataset of paired satellite images (baseline + current) from 11 global disaster events, annotated with structured JSON triage outputs for training or evaluating vision-language models (VLMs) on macro-scale civilian disruption detection. The dataset contains 120 examples total (92 train, 28 eval), each including: 1) a baseline satellite image of a location before a disaster event; 2) a current satellite image of the same location after the event; 3) a structured JSON output with fields like action (discard, defer, downlink_now), category (e.g., earthquake_building_damage), rationale, and normalized bounding box. The dataset is derived from the BRIGHT dataset and licensed under CC-BY-NC-4.0.