WildfireIA/Anonymous-WildfireIA

WildfireIA Anonymous Benchmark Release

This anonymous release contains the WildfireIA benchmark data used for review. WildfireIA is an event-level benchmark for predicting whether a reported Natural wildfire escapes initial attack from public information available at fire discovery time.

Contents

• data/canonical/raw_feature_tables/: canonical benchmark tables. These are the primary dataset artifact. They contain event-level tables, source-level feature tables, patch-level canonical tables, labels, splits, and manifests.
• Model-ready caches are not included in this compact release; regenerate them deterministically from the canonical tables with code/dataloader.py.
• code/: anonymous copies of the cache generation, training, and summary scripts.
• metadata/: release manifest and cache generation commands.
• croissant.json: Croissant metadata with Responsible AI fields.

Tasks

Task 1 predicts initial attack failure. The sample unit is one FPA-FOD Natural wildfire event. Events with final size at most 10 ha are labeled 0, events with final size at least 50 ha are labeled 1, and intermediate-size events are excluded from the Task 1 supervised split.

Task 2 predicts remaining time-to-containment as a regression target, log(1 + containment_hours), using the same discovery-time input contract.

Rebuilding Model-Ready Caches

The canonical tables can regenerate all official model-ready caches:

python code/dataloader.py \
  --base_dir . \
  --canonical_dir data/canonical/raw_feature_tables \
  --output_dir data/model_ready \
  --task ia_failure \
  --representation all \
  --weather_days 5 \
  --input_protocol all \
  --overwrite

Additional ablation caches are generated by changing --input_protocol and --weather_days; see metadata/cache_generation_commands.md.

Responsible Use

The benchmark is intended for reproducible scientific comparison and ablation analysis. It should not be used as a standalone operational dispatch system without agency validation. The data are public-source derived, but they include wildfire locations, fire-station locations, roads, population density, and other geospatial context.