yunusserhat/firerisk-siglip2-base

FireRisk SigLIP2 ViT-B/16 full fine-tuning

This model assigns one of seven aerial hazard labels to an RGB image. It uses the SigLIP2 ViT-B/16 image encoder and the FireRisk Bench classifier wrapper. The checkpoint contains the fine-tuned vision encoder and classifier.

These are initial development weights from training seed 42. The selected checkpoint is epoch 7. It was selected by validation macro F1. The test partition has not been evaluated.

Validation metric	Value
Accuracy	63.05%
Macro F1	58.94%
Balanced accuracy	58.11%

The validation partition also selected the checkpoint and fitted the saved temperature. Its metrics do not estimate independent test performance. Single-seed results do not establish variation across training seeds. The scores of releases with different recipe settings do not form a controlled causal ablation of adaptation mode.

Loading

Use FireRisk Bench. These files use its VisionClassifier architecture. They do not provide a Transformers AutoModel or pipeline checkpoint.

git clone https://github.com/yunusserhat/firerisk.git
cd firerisk
export FIRERISK_HOME="$PWD/.artifacts"
bash scripts/bootstrap.sh
source .venv/bin/activate
hf download yunusserhat/firerisk-siglip2-base --local-dir ./downloaded-model
firerisk predict --run-dir ./downloaded-model --image /path/to/aerial_image.png

For CPU installation, follow the repository's docs/installation.md. Prediction does not download FireRisk. It downloads the pinned backbone on the first load, even for the full release, because the current builder starts from the pretrained architecture. Subsequent loads can reuse that cache. best.pt loads with torch.load(..., weights_only=True).

Data and training

The input is blanchon/FireRisk revision 234b2e7fe6be2da773472e83bd4d42cc9815a630. Its mirror contains 70,331 images and only the source training partition. This project defines a new image-level split with 49,231 training, 10,552 validation and 10,548 test examples. It is not the original paper's benchmark split.

The split seed is 2026. Pixel duplicate auditing found 0 exact duplicate rows. The split hash is 0ab5619091f80f73f8229634a38194ad31eeddf9dfe70978d1b664fe8fb598cd.

The backbone is timm/vit_base_patch16_siglip_224.v2_webli revision 4c3661e5ac879a276ddc5ddc6d3f0ecc78fd5d82. The classifier has 92,891,143 total parameters and 92,891,143 trainable parameters. Input size is 224 pixels. config.json and preprocessing.json retain the recipe, normalization, resize policy and augmentation settings. provenance.json preserves the recorded training environment. The recorded artifacts do not identify a recoverable training Git commit. Training used a dirty working tree.

calibration.json holds a temperature fitted on validation logits. The CLI uses it for its probability output. Class order is stored in labels.json and summary.json. validation_metrics.json includes per-class scores, confusion matrices and exploratory bootstrap intervals.

Intended use

The model supports research on WHP-derived aerial hazard classes. Its class probabilities do not represent a calibrated probability of future wildfire occurrence. It is not an active-fire detector.

The mirror has no coordinates, timestamps or scene identifiers. Exact pixel duplicate auditing does not exclude nearby or overlapping imagery across partitions. Geographic or temporal transfer has not been established. Label ambiguity, image acquisition changes and unknown pretraining overlap also limit interpretation.

Files and licenses

The release contains an inference checkpoint, portable configuration, class mapping, preprocessing, validation results and source provenance. It excludes source images, dataset shards, prediction arrays, optimizer state and credentials. SHA256SUMS lists file checksums.

This release's fine-tuning additions and classifier use GNU GPL v3. The original pretrained model retains its Apache 2.0 terms and attribution. See LICENSE, UPSTREAM_LICENSE, UPSTREAM_NOTICE.md and the unmodified upstream_model_card.md. The frozen release does not redistribute its unchanged backbone weights. The dataset card lists its license as unknown. The code or model release does not license the training imagery.