SatQuery AI — LoRA Adapter (Qwen2.5-VL-3B)
Smart India Hackathon 2026 | Problem Statement #167: Earth Observation Foundation Intelligence
Lightweight LoRA adapter (~20MB) fine-tuned on Sentinel-2 satellite imagery for visual question answering. Apply this adapter on top of the base Qwen/Qwen2.5-VL-3B-Instruct model.
For the full merged model (plug-and-play, no adapter needed), see argho229/satquery-qwen-vl-3b-merged.
📊 Performance
| Metric | Score |
|---|---|
| Binary RS-VQA Accuracy | 91.4% |
| Bounding Box IoU | 0.742 |
| Inference Speed | ~1.8 s/image |
| Ensemble Accuracy (Tri-Brid) | 92.13% |
⚡ Quick Start
`python from peft import PeftModel from transformers import AutoModelForCausalLM, AutoProcessor import torch
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")
base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen2.5-VL-3B-Instruct", torch_dtype=torch.bfloat16, device_map="auto" ) model = PeftModel.from_pretrained(base, "argho229/satquery-qwen-lora-adapter") model = model.merge_and_unload() # optional: merge for faster inference `
🔧 Adapter Config
| Parameter | Value |
|---|---|
| PEFT type | LoRA |
| Rank (r) | 8 |
| Alpha | 16 |
| Dropout | 0.05 |
| Target modules | q_proj, v_proj |
| Training steps | 1,186 |
| Base model | Qwen2.5-VL-3B-Instruct |
📂 Repository
- GitHub: github.com/Argho009/sih-2026
- Full Merged Model: argho229/satquery-qwen-vl-3b-merged
📄 License
Apache 2.0
Developed for Smart India Hackathon 2026 — Problem Statement #167.
- Downloads last month
- 12
Model tree for argho229/satquery-qwen-lora-adapter
Base model
Qwen/Qwen2.5-VL-3B-Instruct