ChrisRPL/satellite-civilian-conflict-disruption-reporter-v1

这是一个用于微调和评估视觉语言模型的紧凑、质量优先的配对图像数据集，专注于民用基础设施冲突破坏的报告。数据集通过比较基线卫星图像和当前卫星图像，结合位置/日期/上下文文本，要求助手返回关于可见民用破坏证据的严格JSON格式报告。数据集特别强调不涉及战术目标、军事资产检测、路线情报或军事排名等任务。数据集包含训练、评估和校准三个分割，共计206行数据。

This is a compact, quality-first paired-image dataset for fine-tuning and evaluating a vision-language model as a civilian disruption reporter. It compares a baseline satellite image and a current satellite image, uses location/date/context text, and asks the assistant to return concise strict JSON about visible civilian disruption evidence. The task is deliberately narrower than general disaster detection and narrower than policy-action prediction. It does not train tactical targeting, military asset detection, route intelligence, or military ranking. The dataset includes train, eval, and calibration splits, totaling 206 rows.