README.md

---
pretty_name: Satellite Civilian Conflict Disruption Reporter v1
license: cc-by-nc-4.0
tags:
- satellite-imagery
- vision-language
- conflict-disruption
- civilian-infrastructure
- paired-image
- evidence-first
- blackline-atlas
task_categories:
- image-to-text
size_categories:
- n<1K
configs:
- config_name: flat
  data_files:
  - split: train
    path: train_flat.jsonl
  - split: eval
    path: eval_flat.jsonl
  - split: calibration
    path: calibration_flat.jsonl
- config_name: sft
  data_files:
  - split: train
    path: train_sft.jsonl
  - split: eval
    path: eval_sft.jsonl
  - split: calibration
    path: calibration_sft.jsonl
---

# Satellite Civilian Conflict Disruption Reporter v1

Dataset ID: `ChrisRPL/satellite-civilian-conflict-disruption-reporter-v1`

## Status

This is a valid **diagnostic reporter-schema dataset**, not the current Blackline Atlas canonical model gate. The canonical compact calibration/gold dataset remains `ChrisRPL/satellite-disruption-triage-aux-v2-2`.

Use this dataset for future schema-simplification experiments only after respecting the mixed source licenses. Do not treat the associated diagnostic LoRA `ChrisRPL/lfm25-vl-civilian-conflict-reporter-lora-v1` as an accepted or HF-Jobs-completed model.

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 final `recommended_action` is derived from visible evidence and uncertainty.

## Output schema

Assistant responses are strict JSON with exactly these keys:

```json
{
  "visible_change_summary": "string",
  "civilian_disruption_evidence": ["collapsed_building | roof_loss | burn_scar | ..."],
  "negative_evidence": ["no_visible_change | low_visibility | sar_speckle_or_modality_artifact | ..."],
  "uncertainty_factors": ["string"],
  "severity_hint": "none | low | medium | high",
  "recommended_action": "discard | defer | downlink_now",
  "confidence": 0.0,
  "short_rationale": "string"
}
```

## Files

- `train_flat.jsonl`, `eval_flat.jsonl`, `calibration_flat.jsonl`
- `train_sft.jsonl`, `eval_sft.jsonl`, `calibration_sft.jsonl`
- compatibility aliases: `eval_calibration_flat.jsonl`, `eval_calibration_sft.jsonl`
- `images/baseline/*.png`, `images/current/*.png`
- `metadata.json`, `validation_report.md`, `source_audit.md`

## Counts

| Split | Rows |
|---|---:|
| train | 121 |
| eval | 79 |
| calibration | 6 |
| total | 206 |

## Recommended-action balance

```json
{
  "calibration": {"discard": 1, "downlink_now": 5},
  "eval": {"defer": 21, "discard": 20, "downlink_now": 38},
  "train": {"defer": 30, "discard": 33, "downlink_now": 58}
}
```

## Modality balance

```json
{"optical-to-SAR": 144, "optical-to-optical": 60, "optical-to-optical-cloudy": 2}
```

## Source balance

```json
{
  "Blackline Atlas Sentinel-2 paired capture": 15,
  "GabeT29/BRIGHT-XView2Format via ChrisRPL/satellite-disruption-triage-aux-v2-2": 144,
  "xBD-Ukraine local materialization from sda-kr/xbd-ukraine lineage": 47
}
```

## License table

| Source/license | Rows | Notes |
|---|---:|---|
| CC-BY-NC-4.0 | 144 | Non-commercial restriction applies |
| Copernicus Sentinel data terms / internal derived capture; verify for downstream redistribution | 15 | Respect upstream source terms |
| MIT | 47 | Respect upstream source terms |

Composite dataset license is `cc-by-nc-4.0` because BRIGHT-derived rows are CC-BY-NC-4.0. Users must also respect per-source terms, especially Sentinel/Copernicus attribution and xBD-Ukraine lineage.

## Split policy

- BRIGHT explosion events are event-held-out: Bata appears in train; Beirut appears in eval.
- xBD-Ukraine rows are location-held-out by city: Mariupol train, Rubizhne eval, Marinka calibration.
- Blackline Sentinel-2 captures are case/event-held-out; no capture case is in more than one split.
- Image-pair SHA-256 checks prevent exact image-pair duplication across splits.

## Known limitations

- This is a compact diagnostic dataset, not a comprehensive global conflict-damage benchmark.
- Some labels are inherited or rule-derived from BRIGHT/xBD annotations and Blackline scenario labels; they are not all expert human VLM annotations.
- BRIGHT rows are optical-to-SAR and non-commercial (CC-BY-NC-4.0), so the composite dataset is non-commercial.
- Sentinel-2 rows are macro-scale and may miss small roof-level damage; use uncertainty fields and analyst review.
- The dataset intentionally excludes military asset detection, route intelligence, targeting, and tactical ranking tasks.

## Intended use

Research and demo-supporting analyst assistance for humanitarian/civilian-infrastructure disruption triage. Outputs are evidence summaries, not operational commands. Human review is required before any external reporting.