ChrisRPL/satellite-disruption-triage-aux-v2-2

Blackline Atlas卫星干扰分类辅助数据集v2.2是一个紧凑的校准和黄金评估修复切片，专为Blackline Atlas设计。它用于测试视觉语言模型是否能比较成对的卫星图像，并为爆炸或类似冲突的人为破坏引起的宏观规模民用干扰生成证据优先的JSON。数据集从`ChrisRPL/satellite-disruption-triage-aux-v2-1`开始，仅保留来自两个爆炸事件的真实成对图像行：Bata用于训练校准，Beirut用于保留的评估黄金。它比v2.1更小更干净，因为当前模型的瓶颈是低估了正面干扰案例，而不是缺乏大量行。

允许的范围包括民用基础设施干扰、人道主义/物流透明度、公共问责制和宏观可见损害分类。不包括军事基地、武器系统、部队位置、车队情报、路线开放分析、战术目标、打击计划或破坏支持。

数据集包含多个文件，如训练校准和评估黄金的JSONL文件，以及基线和当前图像文件。数据集的行数和类平衡也有详细说明。每行包含多个字段，如视觉证据标签、证据强度、损害机制等。数据集的分割策略是基于事件和位置的保留，训练校准和评估黄金分别来自不同的地点。

已知的局限性包括数据集主要关注爆炸事件，不覆盖全球冲突分布；所有行都是光学到SAR的配对，SAR斑点和模态差异可能看起来像变化；标签来源于BRIGHT衍生的掩码统计和基于规则的证据映射，而非手动专家注释；许可证为CC-BY-NC-4.0，限制商业使用。

数据集的预期用途是校准和评估一个民用卫星视觉语言模型，该模型应发出结构化的分类动作：`discard`、`defer`或`downlink_now`。在将任何适配器提升到演示关键运行时之前，应将其用作模型门控。

The Blackline Atlas Satellite Disruption Triage Aux v2.2 dataset is a compact calibration and gold-eval repair slice for Blackline Atlas. It is designed to test whether a vision-language model can compare paired satellite images and produce evidence-first JSON for macro-scale civilian disruption caused by explosions or conflict-like human-made damage. It starts from `ChrisRPL/satellite-disruption-triage-aux-v2-1` and keeps only real paired image rows from two explosion events: Bata for train calibration and Beirut for held-out eval gold. It is intentionally smaller and cleaner than v2.1 because the current model bottleneck is under-calling positive disruption cases, not lack of bulk rows.

Allowed scope: civilian infrastructure disruption, humanitarian/logistics transparency, public accountability, and macro-scale visible damage triage. Out of scope: military bases, weapons systems, troop positions, convoy intelligence, route-open analysis, tactical targeting, strike planning, or sabotage support.

The dataset includes several files such as JSONL files for train calibration and eval gold, as well as baseline and current image files. The row counts and class balance are also detailed. Each row contains multiple fields such as visual_evidence_tags, evidence_strength, damage_mechanism, etc. The split policy is event-held-out and location-held-out, with train calibration and eval gold coming from different locations.

Known limitations include the dataset being explosion-focused and not covering the full global conflict distribution; all rows are optical-to-SAR pairs, so SAR speckle and modality differences can look like change; labels are inherited from BRIGHT-derived mask statistics and rule-based evidence mapping, not manual expert annotation; the license is CC-BY-NC-4.0, restricting commercial use.

The intended use is to calibrate and evaluate a civilian satellite VLM that emits structured triage actions: `discard`, `defer`, or `downlink_now`. It should be used as a model gate before any adapter is promoted into a demo-critical runtime.