sat-image-boundingbox-sft-full

NU-TONIC raw SFT Full Satellite imagery** and aligned **land-cover** outputs packaged as **image–text** rows for fine-tuning in SFT format. JSONL user prompts name the modality (satellite imagery vs. overhead context) where it matters. Provenance **Locations:** GeoGuessr-style POIs (source: `stochastic/random_streetview_images_pano_v0.0.2`) **Optical:** Sentinel-2 multispectral optical COGs from a public STAC catalog, **blue/green/red** or **visual** preview, percentile-stretched to uint8. **Labels:** per-pixel land-cover raster from Earth Engine, **reprojected** to the same 10 m grid as the RGB stack, then tiled and **nearest-neighbor** downsampled with the mask. **Context:** optional geographic overhead still per POI (token-based static map API) under `mapbox_stills/`. **Rows:** global caption + grounding per tile of satellite imagery, plus **per–land-cover-class** caption . **Downstream format:** `messages[]` with `image` + text, aligned with conventions. Layout `images/*.png` — RGB chips (downsampled, e.g. 224×224). `overlays/*.png` — optional bbox visualization per JSONL row (``--write-bbox-overlays``). `mapbox_stills/*.png` — overhead context still per POI (when API token set). `data/train.jsonl`, `data/validation.jsonl`, `data/test.jsonl` — splits hashed by `poi_id`. `metadata/*.json` — per-tile sidecars (coordinates, scene id, `land_cover_label_meta` counts/shapes only). Licenses Respect the **STAC** data provider terms, **Earth Engine** terms for label rasters, the **MapBox** ToS for optional stills, and the **MIT** HF pano dataset license for source coordinates. Hub layout (sharded) This snapshot was processed so that ``images/``, ``mapbox_stills/``, ``overlays/``, and ``metadata/`` use at most **8000** files per leaf directory (Hub git limit: 10k files per directory). JSONL paths may include ``sNNNNN/`` shard segments where needed.