README.md

---
license: mit
language:
  - en
task_categories:
  - image-to-text
tags:
  - computer_use
  - gui-grounding
  - reinforcement-learning
  - GroundCUA
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
---

# GroundGUI RL Dataset

A reinforcement-learning (RLOO) corpus for desktop GUI grounding: 12,600
prompts over 11,205 screenshots from 87 desktop applications. Each prompt is a
screenshot plus a natural-language instruction, and the reward is whether the
predicted click lands inside the target element's bounding box.

All screenshots, instructions and boxes come from
[GroundCUA](https://huggingface.co/datasets/ServiceNow/GroundCUA) (MIT). This
release only selects and mixes samples and formats them as Qwen3-VL
`computer_use` tool-call prompts. None of its screenshots is identical to
any UI-Vision image.

## Usage

```python
import json
from datasets import load_dataset

ds = load_dataset("PrentisAI/GroundGUI_RL", split="train")
row = ds[0]
img = row["images"][0]                                       # PIL.Image
x1, y1, x2, y2 = row["solution"]["arguments"]["coordinate"]  # target box, absolute pixels
extra = json.loads(row["additional_paras"])                  # JSON string: image_size, app, slice, ...
```

`messages` holds the system, user and assistant turns. The assistant turn is
a reference answer: the box centre in 0–1000 normalised coordinates. Compute
the reward from `solution`.

## License

MIT, inherited from GroundCUA (see `LICENSE`).

## Citation

This RL corpus is introduced in [GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning](https://arxiv.org/abs/2610.19100). If you use it, please cite:

```bibtex
@misc{xie2026groundguiguigroundinghighquality,
      title={GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning},
      author={Qingsong Xie and Jiahao Tian and Jiacheng Tang and Zhifeng Li and Madalina Ciobanu and Navan Preet Singh and Qingqing Mao and Ritankar Das},
      year={2026},
      eprint={2610.19100},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.19100},
}
```