hjvsl/GeoZero_Train_Datasets

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:

{
  "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:

{
  "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

import json

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

Load with Hugging Face Datasets

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:

@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, d_wang@whu.edu.cn