🛰️ SatQuest — Satellite VQA (Continued SFT v2)
Qwen2-VL-2B fine-tuned for satellite imagery question answering — produces complete natural-language sentences.
Built on top of Abdullah-123/qwen2vl-2b-hrvqa-lora with continued SFT to teach sentence-style responses.
Before vs After
Q: What is the main transportation? → Before: "car" | After: "The main mode of transportation visible in the image is car."
Q: Is there any water section? → Before: "yes" | After: "Yes, there is a water section present in the image."
Q: How many small vehicles? → Before: "2" | After: "There are two small vehicles visible in the image."
Model Details
| Property | Value |
|---|---|
| Base | Qwen2-VL-2B-Instruct |
| First fine-tune | HRVQA dataset — 159,980 QA pairs |
| This fine-tune | 20,000 verbalized sentence-style QA pairs |
| Method | LoRA (rank=16) via Unsloth + TRL SFTTrainer |
| GPU | NVIDIA A100 80GB |
| Epochs | 2 (LR=1e-4, cosine) |
Usage (with Unsloth)
from unsloth import FastVisionModel
model, tokenizer = FastVisionModel.from_pretrained(
'Humaix/satquest-qwen2vl-sft-v2',
load_in_4bit=True,
)
FastVisionModel.for_inference(model)
Usage — Merged Model (plain transformers, for production backend)
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
import torch
model = Qwen2VLForConditionalGeneration.from_pretrained(
'Humaix/satquest-qwen2vl-sft-v2-merged',
torch_dtype=torch.float16,
device_map='auto',
)
processor = AutoProcessor.from_pretrained('Humaix/satquest-qwen2vl-sft-v2-merged')
Project
NED University of Engineering & Technology
Final Year Design Project (CS-22030) — SatQuest: AI-powered Satellite Imagery Question Answering System