minglanga/GeoCoT380k

Geo-CoT380k

Dataset Description

Geo-CoT380k is a large-scale remote sensing vision-language dataset introduced in the RSThinker project:
Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models.

Geo-CoT380k is designed to support faithful and verifiable reasoning in remote sensing vision-language models. Instead of only providing final answers, the dataset contains structured GeoSpatial Chain-of-Thought rationales, encouraging models to perform multi-step, perception-grounded analysis over remote sensing imagery.

This dataset is used for supervised fine-tuning in the two-stage alignment framework of RSThinker, where models first learn structured Geo-CoT reasoning through SFT and are then further optimized with GRPO for factual correctness.

• Project repository: https://github.com/minglangL/RSThinker
• Model checkpoint: https://huggingface.co/minglanga/RSThinker
• Paper: https://arxiv.org/abs/2509.22221

Dataset Overview

Geo-CoT380k contains approximately 380k instruction-style samples for remote sensing vision-language reasoning. Each sample generally consists of:

• A user instruction or question with an image placeholder
• One associated remote sensing image
• A structured reasoning trace enclosed in <think>...</think>
• A final answer enclosed in <answer>...</answer> The dataset is intended to help vision-language models learn not only what answer to give, but also how to produce a faithful, perceptually grounded analytical process.

Dataset Format

The dataset follows an instruction-tuning format compatible with multimodal training frameworks such as LLaMA-Factory. A typical sample is formatted as follows:

[
  {
    "messages": [
      {
        "content": "<image>Detect all baseball diamond in the image.",
        "role": "user"
      },
      {
        "content": "<think>To detect all baseball diamonds in the image, I first analyze the overall layout and visual characteristics of the scene...</think><answer>[[244, 290, 710, 638], [242, 646, 718, 980]]</answer>",
        "role": "assistant"
      }
    ],
    "images": [
      "HRRSD/resize_images/07907.jpg"
    ]
  }
]

Associated Model: RSThinker

Geo-CoT380k is released together with RSThinker, a remote sensing vision-language model designed for structured and faithful geospatial reasoning. RSThinker is trained with a two-stage alignment strategy:

• Supervised Fine-Tuning, using Geo-CoT380k to instill structured Geo-CoT reasoning.
• GRPO Training, using remote sensing reward modeling to improve factual correctness.

Citation

If you find this dataset useful, please cite:

@misc{liu2025faithfulreasoningremotesensing,
      title={Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models}, 
      author={Jiaqi Liu and Lang Sun and Ronghao Fu and Bo Yang},
      year={2025},
      eprint={2509.22221},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.22221}, 
}