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

Blackline Atlas Satellite Disruption Triage Aux v2.2

This 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.

Scope

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.

Files

• train_calibration_flat.jsonl: evidence-first flat rows for calibration/training experiments.
• eval_gold_flat.jsonl: event-held-out eval rows for product gating.
• train_calibration_sft.jsonl: chat/SFT version of train calibration rows.
• eval_gold_sft.jsonl: chat/SFT version of eval gold rows.
• images/baseline/: baseline image files.
• images/current/: current image files.
• metadata.json: machine-readable stats.
• validation_report.md: validation checks.
• source_audit.md: source and exclusion notes.

Counts

Split	Rows
train_calibration	93
eval_gold	51
total	144

Class Balance

{
  "eval_gold": {
    "defer": 17,
    "discard": 17,
    "downlink_now": 17
  },
  "total": {
    "defer": 46,
    "discard": 44,
    "downlink_now": 54
  },
  "train_calibration": {
    "defer": 29,
    "discard": 27,
    "downlink_now": 37
  }
}

Schema

Every flat row contains these fields, with triage_action near the end and all outputs structured for deterministic validation:

row_id, example_id, split, baseline_image, current_image, location_name, country, source_event, source_dataset, modality, baseline_date, current_date, visual_evidence_tags, evidence_strength, damage_mechanism, visibility_quality, negative_type, bbox_norm, bbox_quality, change_confidence, civilian_infrastructure_type, rationale, triage_action, provenance, license

The main target fields are visual_evidence_tags, evidence_strength, damage_mechanism, visibility_quality, negative_type, bbox_norm, bbox_quality, change_confidence, civilian_infrastructure_type, rationale, and triage_action.

Split Policy

The split is event-held-out and location-held-out:

• Train calibration: Bata, Equatorial Guinea.
• Eval gold: Beirut, Lebanon.

No source event or location appears in both splits.

Known Limitations

1. The dataset is explosion-focused and does not cover the full global conflict distribution.
2. All rows are optical-to-SAR pairs, so SAR speckle and modality differences can look like change.
3. Labels are inherited from BRIGHT-derived mask statistics and rule-based evidence mapping, not manual expert annotation.
4. License is CC-BY-NC-4.0, so commercial use is restricted.

Intended Use

Use this dataset 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.