pavi-07/satquery-cdvqa-lora

SatQuery AI — Qwen2.5-VL CDVQA LoRA Adapter

---

🛰️ Project Overview & 3-Link Evaluation Triad

SatQuery AI is an Edge-First, Multi-Modal Satellite Question-Answering and Geospatial Intelligence system engineered for Indian Earth Observation (EO) missions (ISRO CARTOSAT, RISAT, Sentinel-1/2, and global benchmarks like SECOND and BEN-GE-8K).

For evaluators and collaborators, the system is organized into a connected 3-Link Triad:

1. 🌐 Interactive Cloud Demo (HF Spaces): https://huggingface.co/spaces/pavi-07/satquery
2. 🤖 Fine-Tuned Adapter Weights (HF Model Hub): https://huggingface.co/pavi-07/satquery-cdvqa-lora
3. 💻 Full Open-Source Codebase (GitHub): https://github.com/pavitravashishtha/satquery-demo (and satquery core)
4. 🎥 Full Dashboard & Concept Walkthrough (YouTube): https://youtu.be/4VfqViYAwJ8?si=SWgOBt213bhzTTXT
   • All-in-one demonstration covering Earth Observation concepts, edge computing advantages, multi-sensor features, and full dashboard walkthrough.

---

⚠️ Architectural Disclaimer & Edge Decoupling Notice

Edge-First Design vs. Cloud Preview: SatQuery AI is engineered strictly for air-gapped, low-power edge deployment on a 6GB VRAM GPU (NVIDIA RTX 4050 Laptop / NVIDIA Jetson Orin 8GB).

The cloud endpoints (Hugging Face Spaces and Model Hub) serve as interactive verification prototypes and model distribution channels. During real operational deployments in remote or disaster-hit corridors, all models (Qwen2.5-VL INT4 + LoRA adapter + Optical-SAR Dual-CNN) execute locally on consumer edge hardware with zero external API calls and zero cloud telemetry leakage.

---

📊 Empirical Benchmarks (RTX 4050 Hardware)

This adapter was trained and evaluated specifically on Change Detection Visual Question Answering (CDVQA) pairs from the SECOND benchmark dataset:

Metric	Zero-Shot Base (Qwen2.5-VL-3B)	SatQuery CDVQA LoRA (Fine-Tuned)	Absolute Delta
Accuracy / Exact Match (EM)	29.80%	71.2% (71.20%, 95% CI: [67.20%, 75.20%])	+41.40% pts
Expected Calibration Error (ECE)	—	0.19 (Routing ECE: 0.1985)	Well-Calibrated
F1 Score	0.384	0.782	+103.6%
Adapter Swap Latency	—	4.81 ms	Instant runtime swap
Adapter VRAM Footprint	—	15.2 MB	Fits inside swap slot

---

⚡ Edge Deployment & VRAM Budget (6GB Ceiling)

SatQuery AI's MemoryLifecycleManager implements a time-multiplexed dynamic swap slot ensuring peak memory stays strictly below 6GB:

+-------------------------------------------------------------+
| Resident Base Memory (Always in VRAM)                       |
|   - PyTorch CUDA Context & OS Display Headroom :    600 MB  |
|   - Optical-SAR Cross-Attention Fusion Model    :    180 MB  |
|   - Bi-Temporal Change Detection Specialist     :    750 MB  |
|   - Multi-Label Land Cover Specialist           :    310 MB  |
|   Subtotal Resident Base                       :  1,840 MB  |
+-------------------------------------------------------------+
| Dynamic Swap Slot (Time-Multiplexed Lifecycle)              |
|   - Qwen2.5-VL-3B (INT4 Quantized + LoRA)       :  3,820 MB  |
|   Peak Active VRAM (Base + VQA Engine)          :  5,660 MB  |
+-------------------------------------------------------------+
| Safety Headroom (Activation tensors & cache)    :    484 MB  |
+-------------------------------------------------------------+

---

🚀 Quickstart: Loading Adapter in Python

import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from peft import PeftModel

base_model_id = "Qwen/Qwen2.5-VL-3B-Instruct"
lora_model_id = "pavi-07/satquery-cdvqa-lora"

# 1. Load 4-bit Quantized Base Model
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(base_model_id)

# 2. Attach Fine-Tuned CDVQA LoRA Adapter
model = PeftModel.from_pretrained(model, lora_model_id)
model.eval()

print("SatQuery CDVQA LoRA successfully loaded onto edge hardware!")

---

🔬 Training Configuration

• Base Model: Qwen/Qwen2.5-VL-3B-Instruct
• LoRA Rank ($r$): 8
• LoRA Alpha ($\alpha$): 16
• Target Modules: q_proj, k_proj, v_proj, o_proj
• LoRA Dropout: 0.05
• Optimizer: AdamW (lr=2e-4, cosine schedule)
• Precision: Mixed Precision (bf16)
• Trained Parameters: ~15.2 MB safetensors (0.42% of base model weights)

---

📜 License & Citation

This model is licensed under Apache 2.0.

@misc{satquery2026,
  title={SatQuery AI: Edge-First Multi-Modal Satellite Question-Answering System},
  author={Vashishtha, Pavitra},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/pavi-07/satquery-cdvqa-lora}}
}