shreyashsri79/satquery-m1-rs-vqa

SatQuery M1 — remote-sensing VQA adapter

A QLoRA adapter for Qwen/Qwen2-VL-7B-Instruct (revision eed13092ef92e448dd6875b2a00151bd3f7db0ac), trained on VRSBench (CC-BY-4.0) for visual question answering on optical satellite imagery. It is component M1 of SatQuery AI, an entry for Smart India Hackathon 2026, PS26167 (ISRO).

	VRSBench validation, exact match, n = 2,000
Base model, zero-shot	0.527
With this adapter	0.660
Gain	+0.133

Training: 12,000 QA pairs from Annotations_train, 1 epoch, LoRA r=8 / alpha=16 / dropout 0.05 on q_proj k_proj v_proj o_proj, 4-bit NF4 base, fp16 compute, one Kaggle T4. Train and validation images do not overlap.

Serve it with the 4-bit base model and greedy decoding, using the prompt suffix "Answer the question using a single word or short phrase." — a different prompt or precision is a different model from the one measured. pack.json holds the full manifest.

Optical imagery only. Not trained or evaluated on SAR.

Licence: Apache-2.0, as the base model. Data: VRSBench, CC-BY-4.0 (Xia et al.).