NuTonic/firewatch-sft-v1

FireWatch SFT

Temporal Sentinel-2 image pairs (pre/post around a wildfire-relevant event_date) over a fixed AOI. dNBR-style burn signal drives a change mask; each tile emits a change caption row and, when regions exist, a grounding row with normalized bounding boxes.

Record counts (this build)

Split	JSONL lines
train	200
validation	0
test	48
total	248

Tiles processed (caption rows ≈ this; grounding rows added when regions exist): 161.

Inputs

• Events: JSON or CSV with event_id, lat, lon, event_date (see data/events/fire_smoke_events.json in the nutonic repo for examples).
• CLI: build_lfm_vl_firewatch_sft.py --events … --max-events 0 processes every event row; tune --bbox-half-km, --native-tile, --stride, --max-cloud-pct, --pre-window-days, --post-window-days.

Dataset layout

• data/train.jsonl, data/validation.jsonl, data/test.jsonl — VLM SFT samples (messages list with system/user/assistant content; images referenced by relative paths under this folder).
• images/ — PNG chips (typically pre/post pair (*_t0.png, *_t1.png)) consumed by the JSONL.
• metadata/ — one JSON sidecar per tile/sample (scene ids, bbox, optional regions, profile tag).

Splits

Train / validation / test are assigned from a stable hash of event_id (same event_id always lands in the same split).

Regenerating locally

Run from the nutonic repository root (paths relative to that root):

python data/scripts/build_lfm_vl_firewatch_sft.py <see --help for required args>

Upload to the Hub (requires a token with write access to the target dataset repo):

export HF_TOKEN="hf_..."   # or HUGGING_FACE_HUB_TOKEN
python data/scripts/build_lfm_vl_firewatch_sft.py ... --upload-repo NuTonic/firewatch-sft-v1 --hf-token "$HF_TOKEN"

Skip upload with --no-upload. Resume partial STAC downloads with --skip-existing.

Orchestrator (all PRO profiles)

python data/scripts/run_pro_sft_orchestrator.py \
  --events-per-profile 0 \
  --hf-org YOUR_ORG \
  --upload-repo-firewatch YOUR_ORG/firewatch-sft-v1 \
  --hf-token "$HF_TOKEN"

Use --upload-repo-<profile> per profile, or --hf-org to upload to YOUR_ORG/<default-repo-suffix> for each builder. See run_pro_sft_orchestrator.py --help.

Environment

Variable	Purpose
HF_TOKEN or HUGGING_FACE_HUB_TOKEN	Hub upload if --hf-token is not passed.
HTTP_PROXY / HTTPS_PROXY	Optional; if your network requires a proxy to reach STAC and Hugging Face.

STAC: scenes are resolved from Element84 Earth Search (sentinel-2-l2a by default). No Earth Engine or Mapbox keys are required for this pipeline.

Hub target for this artifact

Dataset repo id: NuTonic/firewatch-sft-v1

Card URL: https://huggingface.co/datasets/NuTonic/firewatch-sft-v1

Limitations

• Optical-only Sentinel-2; cloud cover and revisit gaps can drop events or leave empty change masks.
• Detections and captions are heuristic (indices / thresholds), not operational truth labels.
• Assistant text follows conservative wording (candidates, confidence qualifiers).