README.md

---
license: apache-2.0
viewer: false
---


# GUI Grounding Pre-training Data for OS-ATLAS
This document describes the acquisition of the pre-training data used by OS-ATLAS [OS-ATLAS: A Foundation Action Model for Generalist GUI Agents](https://huggingface.co/papers/2410.23218).

<div align="center">

[\[🏠Homepage\]](https://osatlas.github.io) [\[💻Code\]](https://github.com/OS-Copilot/OS-Atlas) [\[🚀Quick Start\]](#quick-start) [\[📝Paper\]](https://arxiv.org/abs/2410.23218) [\[🤗Models\]](https://huggingface.co/collections/OS-Copilot/os-atlas-67246e44003a1dfcc5d0d045) [\[🤗ScreenSpot-v2\]](https://huggingface.co/datasets/OS-Copilot/ScreenSpot-v2) 

</div>

![os-atlas](https://github.com/user-attachments/assets/cf2ee020-5e15-4087-9a7e-75cc43662494)


**Notes:** In GUI grounding data, the position of the target element is recorded in the `bbox` key, represented by `[left, top, right, bottom]`. 
Each value is a [0, 1] decimal number indicating the ratio of the corresponding position to the width or height of the image.

The data stored in this dataset consists of raw data containing **only** element grounding information. When training a model, you need to use the corresponding prompts to wrap these data.

The data we released is divided into three domains: mobile, desktop and web.

All annotation data is stored in JSON format and each sample contains:
* `img_filename`: the interface screenshot file
* `instruction`: human instruction or referring expression extracted from ally tree or html
* `bbox`: the bounding box of the target element corresponding to instruction

Some data also contains a `data_type`, which records the type of an element in its structured information, if it can be obtained.

***

### Mobile data

This part of data is stored under the *mobile_domain* directory. Our mobile grounding data consists of four parts.

#### AMEX

Android Multi-annotation EXpo (AMEX) is a comprehensive, large-scale dataset designed for generalist mobile GUI-control agents [1].

The annotation data is stored in

-`amex_raw.json`

Due to the single file size limitation of Hugging Face datasets, we stored the Amex images in *zip* format and split them into several sub-files. 

- `amex_images_part_aa`
- `amex_images_part_ab`
- `amex_images_part_ac`

You need to first merge these split files back into the original file and then extract the contents.

```
cat amex_images_part_* > amex_images.zip
7z x amex_images.zip -aoa -o/path/to/extract/folder
```

#### UIBert

UIBert [2] is a dataset extended from Rico dataset [3] for two tasks: similar UI component retrieval and referring expression component retrieval.

The annotation data is stored in

- `uibert_raw.json`

The UIBert images are stored in

- `UIBert.zip`

#### Widget Captioning and RICOSCA

Widget Captioning data are collected by [4]. 

RICOSCA is a dataset automatically labeled using Android VH in [5]

The annotation data is stored in

- `widget_captioning.json`
- `ricosca.json`

The rico images are stored in 

- `rico_imgs.zip`

#### Android_world_data

This part of data are sampled from a android environment for building and benchmarking autonomous computer control agents [6].

The annotation data is stored in

- `aw_mobile.json`

The rico images are stored in 

- `mobile_images.zip`

***

### Desktop data

This part of data is stored under the *desktop_domain* directory. 

All of the desktop grounding data is collected from the real environments of personal computers running different operating systems. Each image is split into multiple sub-images to enhance data diversity.

Our desktop grounding data consists of three parts: Windows, Linux and MacOS.

**The image and annotation data for each operating system are stored in corresponding zip and json files.**

It is worth noting that, due to the large size of the Windows image data, the split files need to be merged before extraction.

```
cat windows_image_part_* > windows_images.zip
7z x windows_images.zip -aoa -o/path/to/extract/folder
```

***

### Web data

This part of data is stored under the *web_domain* directory.

Our desktop grounding data consists of two parts.

#### Seeclick web data

The web data from SeeClick [7] was crawled from websites provided by Common Crawl, containing more than 270k webpage screenshots and over 3 million webpage elements.

The annotation data is stored in

- `seeclick_web.json`

The images are stored into split files and need to be merged before extraction.

```
cat seeclick_web_image_part_* > seeclick_web_images.zip
7z x seeclick_web_images.zip -aoa -o/path/to/extract/folder
```

#### Fineweb_crawled_data

This part of data is crawled from web pages from the latest URLs obtained from FineWeb [8], a cleaned and deduplicated English dataset derived from Common Crawl.

Since this portion of the data contains at least 1.6 million images, we have compressed them into 10 zip files, from `fineweb_3m_s11.zip` to `fineweb_3m_s52.zip`.

Please extract them into the same directory. 

As an example,

```
7z x fineweb_3m_s11.zip -aoa -o/same/path/to/extract/fineweb
```

The annotation data is stored in

- `fineweb_3m.json`

***

### Best practice

During the training of **OS-Atlas-7B**, we randomly sampled predefined prompts to wrap the grounding data. Additionally, we scaled the relative coordinates of each element (in the range [0, 1]) by multiplying them by 1000 before inputting them into the model for training. 

Below is an example of a data entry:

```
{
  "conversations": [
    {
      "from": "human",
      "value": "<image>\nUsing the provided screenshot, I'll describe webpage elements for you to locate (with bbox).\n<ref>media-maniacs.org</ref>\n<ref>Home</ref>\n<ref>Sitemap</ref>\n<ref>shop you can have what you choose 2012</ref>"
    },
    {
      "from": "gpt",
      "value": "<ref>media-maniacs.org</ref><box>[[70,856,141,871]]</box>\n<ref>Home</ref><box>[[21,935,43,951]]</box>\n<ref>Sitemap</ref><box>[[21,919,52,934]]</box>\n<ref>shop you can have what you choose 2012</ref><box>[[368,839,523,855]]</box>"
    }
  ]
}
```

The prompts we used are stored in `prompts.json`.

*** 

**The following are the open-source datasets we used as data sources. We welcome everyone to check the details and cite these sources accordingly！**

[1] [AMEX: Android Multi-annotation Expo Dataset for Mobile GUI Agents](https://arxiv.org/abs/2407.17490)

[2] [UIBert: Learning Generic Multimodal Representations for UI Understanding](https://arxiv.org/abs/2107.13731)

[3] [Rico: A mobile app dataset for building data-driven design applications](https://dl.acm.org/doi/pdf/10.1145/3126594.3126651)

[4] [Widget Captioning: Generating Natural Language Description for Mobile User Interface Elements](https://arxiv.org/pdf/2010.04295.pdf)

[5] [Mapping Natural Language Instructions to Mobile UI Action Sequences](https://arxiv.org/pdf/2005.03776)

[6] [ANDROIDWORLD: A Dynamic Benchmarking Environment for Autonomous Agents](https://arxiv.org/abs/2405.14573)

[7] [SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents](https://arxiv.org/abs/2401.10935)

[8] [The fineweb datasets: Decanting the web for the finest text data at scale](https://arxiv.org/abs/2406.17557)