Wildfire-FM

ICLR Paper Code

The ICLR package is tied to the manuscript snapshot and PDF hash in its manifest.json. Start with its README and artifact map. Full model reruns still require upstream data and comparator assets; unresolved result lineages are documented rather than treated as verified experiments.

The model card, checkpoints, previews, and task snapshot below describe the earlier California release. They are retained for that release and are not a replacement for the ICLR package's manuscript-specific results.

FireEye Platform Overview

For a concise description of the public-facing FireEye platform, its current FireWx-FM integration, and the proposed flood/multi-hazard extension, see docs/FIRE_EYE_PLATFORM_OVERVIEW.md.

Release Navigation

Why WildFIRE-FM

Wildfire Earth-FM transfer scores depend strongly on the contract used for comparison: task form, metric, matching rule, spatial scope, and head-selection criterion. The paper studies these choices by holding outputs or features fixed and then changing only the evaluation contract.

WildFIRE-FM is the in-region reference model used in those comparisons. It is trained for 12-hour gridded wildfire occupancy on a California 5 km grid, then evaluated under the same task-specific contracts as the transferred Earth-FM backbones.

The released weights are California regional checkpoints, not nationwide-trained weights. For nationwide use, the release includes a CONUS cache and training recipe under training/NATIONWIDE_RETRAINING.md.

This Hub repository is model-first. It provides the released weights, model-loading code, compact result summaries, paper previews, and source-data access notes needed to inspect the release boundary before running larger local reproductions.

Key Features

• Wildfire-specialized reference model: Compact U-Net for 12-hour occupancy prediction on a projected California grid.
• Five seeded checkpoints: Seeds 1, 7, 42, 99, and 123 are released with manifest hashes.
• Fixed-contract artifacts: Compact summaries cover matching-rule, head-selection, and task-form comparisons from the final paper.
• Source-data aware release: Provider links and data roles are documented for each public resource used by the study.
• Compiled paper PDF: The readable manuscript is available directly from the Hub release.
• Paper-aligned previews: Figure assets summarize selection regret, supporting-task rank changes, and primary-task rank changes.

Visual Tour

Release contents. The Hub repository is organized around weights, model code, source-data notes, compact artifacts, and final-paper visual previews.

Head-selection regret. Choosing a lightweight head by a ranking metric can lose decision performance under the same frozen features.

Supporting-task rank map. Backbone ordering changes across burned area, analog retrieval, smoke PM2.5, and extreme heat task contracts.

Primary-task rank changes. Primary-task rankings change under fixed matching-rule and task-form choices.

At A Glance

Input Channel Contract

The pretrained checkpoints expect a fixed 16-channel tensor in [channel, y, x] order. Channel order matters. Inputs use native source units with no mean/std normalization, min/max scaling, or unit conversion. Invalid or missing values are zero-filled, with validity masks included as channels 10 and 11.

FIRMS provides the occupancy target, not an input channel. WFIGS and MTBS are used for supporting event-level tasks and are not part of the 16-channel pretrained occupancy input. Full channel-level details are in data_sources/DATA_SOURCES.md; the same contract is available as models/wildfire_fm/input_channels.json. The cape channel is the HRRR surface instantaneous CAPE field selected with typeOfLevel=surface and stepType=instant. For validity channels, firewx_valid is 1.0 everywhere in the released California regional cache and is not a per-variable HRRR missing-data mask; static_valid is the fraction of the four static layers valid after reprojection at each grid cell.

Hugging Face Release

This Hub repository is a model release with paper-aligned reproducibility artifacts.

Full local reruns require users to prepare the provider-hosted raw data and comparator assets described in data_sources/.

Quick Start

Clone the Hub repository or download the files you need:

git clone https://huggingface.co/RAI-Lab/Wildfire-FM
cd Wildfire-FM

Load a seeded checkpoint:

import torch
from models.wildfire_fm.modeling_unet import UNetSmallFlex

model = UNetSmallFlex(
    in_ch=16,
    base=32,
    dropout=0.1,
    norm_type="group",
    norm_groups=8,
    use_aux_spatial_head=True,
)
checkpoint = torch.load(
    "models/wildfire_fm/checkpoints/seed_1/best_firms_prauc.pt",
    map_location="cpu",
)
state = checkpoint.get("model", checkpoint)
model.load_state_dict(state)
model.eval()

Verify the public release artifacts:

python3 scripts/reproduce_paper_outputs.py

The checkpoint expects the same fixed-order 16-channel gridded input described above and in data_sources/DATA_SOURCES.md.

Task-Contract Snapshot

The card below reports the best final-paper mean for each displayed task contract, with the winning backbone named in the final column.

The release provides sanitized CSV/JSON summaries used to audit displayed values.

Data Sources

The study uses public or provider-hosted resources. Source roles and access paths are documented in data_sources/DATA_SOURCES.md.

• NOAA HRRR fields for regional weather inputs.
• NASA FIRMS active-fire detections for occupancy supervision.
• LANDFIRE fuel and canopy layers for static landscape context.
• Wildfire Risk to Communities housing density and LandScan population for exposure context.
• WFIGS and MTBS event-level resources for burned-area and analog tasks.
• External Earth-FM and backbone assets for comparator features.

Repository Layout

WildFIRE-FM/
  README.md
  LICENSE
  requirements.txt
  models/wildfire_fm/
    modeling_unet.py
    checkpoint_manifest.json
    checkpoints/seed_*/best_firms_prauc.pt
  training/
    build_phase1_cache_regional_hrrr.py
    train_cold_tiled_mainline.py
    configs/
  spatial_serving/
    grid_to_polygons.py
  data_downloader/
    hrrr_downloader.py
  paper/
    wildfire_fm_evaluation_contracts.pdf
  paper_reproduction/iclr2027/
    README.md
    manifest.json
    scripts/
    evaluation/
    data/
    reference_outputs/
    docs/
    slurm/
  assets/
    wildfire_fm_model_card.svg
    release_contents.svg
    selection_regret_final.png
    supporting_rank_map_final.png
    primary_rank_change_final.png
  paper_outputs/figures/
    matching.pdf
    fig_task_contract_tiles.pdf
    fig_selection_regret_scatter.pdf
    fig_rank_heatmap1.pdf
  artifacts/
    manifests/
    results/
  data_sources/
    DATA_SOURCES.md
  experiments/
    raw_reference/
    slurm/
  scripts/
    audit_release.py
    reproduce_paper_outputs.py

How To Cite

If you use WildFIRE-FM, the released checkpoints, the fixed-contract evaluation artifacts, or the paper-aligned scripts, please cite:

@misc{wildfire_fm_evaluation_contracts_2026,
  title = {Does Your Wildfire Prediction Model Actually Work, or Just Score Well?},
  author = {Yangshuang Xu and Yuyang Dai and Liling Chang and Qi Wang and Yushun Dong},
  year = {2026},
  note = {WildFIRE-FM model and fixed-contract wildfire evaluation artifacts}
}

The citation will be updated with arXiv metadata after the preprint is public.

News

• 2026-10-11 - Updated the ICLR reproduction package to the single-Pangu manuscript snapshot, with 29 tables, 7 figure groups, portable evidence, and explicit verification coverage.
• 2026-05-18 - The model card was reorganized with a larger navigation panel and final-paper visual previews.
• 2026-05-13 - The Hub release added the compiled paper PDF without checklist pages.
• 2026-05-13 - Five seeded WildFIRE-FM checkpoints were uploaded with SHA-256 metadata.

Scope

WildFIRE-FM is scoped to the paper's California regional grid and fixed-contract wildfire transfer comparisons. Use in a new region or grid should start with local preprocessing, validation, and contract-specific evaluation. The release is designed for research inspection, artifact-level checking, and model-loading experiments rather than operational alerting.

License

This release is provided under the MIT license. External data and comparator assets remain governed by their original providers and terms of use.

WildFIRE-FM is released to make wildfire Earth-FM transfer comparisons easier to inspect, reproduce at the artifact level, and evaluate under explicit contracts.