keshavagarwal2004dev/GeoZero_Train_Datasets

---
license: cc-by-nc-sa-4.0
task_categories:
- visual-question-answering
- image-to-text
- reinforcement-learning
- feature-extraction
language:
- en
tags:
- remote-sensing
- earth-observation
- satellite-imagery
- geospatial
- geospatial-reasoning
- multimodal
- visual-question-answering
- vision-language-model
- foundation-model
size_categories:
- 100K<n<1M
---

# SFT and RL Traning dataset of GeoZero

## Dataset Composition

GeoZero consists of three variants:

| File | Description |
|------|------------|
| **GeoZero-Raw.json** | Raw aggregated data across heterogeneous datasets |
| **GeoZero-Instruct.json** | Unified instruction-tuned dataset for supervised fine-tuning |
| **GeoZero-Hard.json** | Challenging subset for RL training |

All image files are stored under the `images/` directory.

## Directory Structure

```
GeoZero_Train_Datasets/
├── images/
│ ├── AID-0000.tar
│ ├── AID-0001.tar
│ ├── RSVQA-HR-0000.tar
│ ├── ...
│
├── GeoZero-Raw.json
├── GeoZero-Instruct.json
├── GeoZero-Hard.json
└── Readme.md
```

If tar shards are used, each tar file preserves relative paths:

```
RSVQA-HR/8766.png
```

## JSON Format Examples

Each JSON file contains a list of samples in the following structure:

For GeoZero-Raw.json and GeoZero-Instruct.json:

```json
{
"messages": [
{
"role": "user",
"content": "<image>\n[vqa] Is there a residential building on the right of the university?"
},
{
"role": "assistant",
"content": "no"
}
],
"images": [
"RSVQA-HR/8766.png"
]
}
```

For GeoZero-Hard.json:

```json
{
"messages": [
{
"role": "system",
"content": "system prompt"
},
{
"role": "user",
"content": "<image>\n[vqa] What is the area covered by residential buildings? Give a response of yes or no."
}
],
"images": ["RSVQA-HR/118.png"],
"solution": "<answer> 1934m2 </answer>\n",
"task_type": ["vqa"]
}
```

## Loading the Dataset

### Load JSON Directly

```python
import json

with open("GeoZero-Instruct.json", "r", encoding="utf-8") as f:
data = json.load(f)
```

### Load with Hugging Face Datasets

```python
from datasets import load_dataset

dataset = load_dataset(
"hjvsl/GeoZero_Train_Datasets",
data_files="GeoZero-Instruct.json"
)
```

## Citation

If you use GeoZero in your research, please cite:

```bibtex
@article{wang2025geozero,
title = {GeoZero: Incentivizing Reasoning from Scratch on Geospatial Scenes},
author = {Wang, Di and Liu, Shunyu and Jiang, Wentao and Wang, Fengxiang and Liu, Yi and Qin, Xiaolei and Luo, Zhiming and Zhou, Chaoyang and Guo, Haonan and Zhang, Jing and Du, Bo and Tao, Dacheng and Zhang, Liangpei},
journal = {arXiv preprint arXiv:2511.22645},
year = {2025}
}
```

## Contact

Di Wang, Wuhan University, [e-mail]