Modelpublic

aanandmodi/satquery-qwen3vl-bigearthnet-txt-lora

sourceHugging Faceapache-2.0updated 20d agoView on Hugging Face

1likes181downloads

Model Card

SatQuery Qwen3-VL BigEarthNet.txt adapter

QLoRA adapter for SIH26167 SatQuery AI single-image remote-sensing VQA, captioning and grounding.

Base and data

—Base: Qwen/Qwen3-VL-2B-Instruct at immutable revision 89644892e4d85e24eaac8bacfd4f463576704203.
—Text: BIFOLD-BigEarthNetv2-0/BigEarthNet.txt.
—Imagery profile: lithuania-summer; 6,000 bounded training rows.
—Boxes are converted from BigEarthNet.txt normalized 0..1 coordinates to Qwen 0..1000.

Intended use

Use on RGB previews derived from optical imagery. Do not pass raw SAR or 12-band Sentinel tensors to this model. SatQuery routes those inputs to its TerraMind fusion specialist.

Limitations

The training subset is European Sentinel imagery. Performance on Indian Cartosat/RISAT imagery is not established. Outputs require validation, evidence display, uncertainty labeling and abstention. Generated-token likelihood is not a calibrated probability of correctness.

Validation summary

json

{
  "rows": 120,
  "vqa_exact_match": 0.6,
  "grounding_mean_iou": 0.3620281495145568,
  "grounding_accuracy_iou_0_5": 0.4,
  "caption_token_f1": 0.35107449654793066,
  "proxy_ece_on_non_caption_rows": 0.24431528830364974
}