chenruihan/MPR-GUI-Bench

---
license: cc-by-nc-4.0
---

# MPR-GUI Images

This directory contains the image assets for the dataset used in the paper:

`MPR-GUI: Benchmarking and Enhancing Multilingual Perception and Reasoning in GUI Agents`

Paper link: https://arxiv.org/abs/2512.00756
Hugging Face dataset: https://huggingface.co/datasets/chenruihan/MPR-GUI-Bench
GitHub repository: https://github.com/chenruihan32/MPR-GUI-Bench

## Directory layout

```text
images/
├── 1/ ... 6/
│ ├── en/
│ ├── fr/
│ ├── ja/
│ ├── ru/
│ ├── th/
│ └── zh/
├── rich/
│ ├── en/
│ ├── fr/
│ ├── ja/
│ ├── ru/
│ ├── th/
│ └── zh/
└── sparse/
├── en/
├── fr/
├── ja/
├── ru/
├── th/
└── zh/
```

## What each folder contains

- `1/` to `6/`: screenshot image files used by the single-image QA files in `qas/`. These folders are preserved as numbered image partitions in the released dataset.
- `rich/`: multi-image episode folders used by the `knowledge_rich_*.jsonl` files in `qas/`.
- `sparse/`: multi-image episode folders used by the `knowledge_sparse_*.jsonl` files in `qas/`.

Each language subfolder uses the same language code convention:

- `en`: English
- `fr`: French
- `ja`: Japanese
- `ru`: Russian
- `th`: Thai
- `zh`: Chinese

## File naming

- In numbered folders such as `6/en/`, image files follow names like `camera_en_8.jpg` or `phone_en_1.jpg`.
- In `rich/` and `sparse/`, each language directory contains episode folders such as `sport_episode_0/` or `booking_episode_1/`.

## How to combine with `qas/`

The relative paths inside the QA files assume the following structure:

```text
<dataset_root>/
├── images/
└── qas/
```

That is, `images/` and `qas/` should be placed as sibling directories under the same root folder.

For example:

```text
mpr_gui_bench/
├── images/
│ ├── 1/
│ ├── 2/
│ ├── 3/
│ ├── 4/
│ ├── 5/
│ ├── 6/
│ ├── rich/
│ └── sparse/
└── qas/
```

Under this layout:

- `../images/6/en/phone_en_1.jpg` from a QA file resolves correctly.
- `../images/rich/en/sport_episode_0` from a knowledge QA file resolves correctly.

## Notes

- If `images/` and `qas/` are hosted separately, you may need to rewrite the paths in `qas/` or map them to a remote prefix in your data-loading code.
- The current release keeps the original numbered partition folders `1` to `6` as-is.