File size: 4,740 Bytes
8d778a0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a4af5e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0248027
 
 
 
 
 
a4af5e2
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
---
license: apache-2.0
task_categories:
- visual-question-answering
- image-to-text
language:
- en
tags:
- mobile-ui
- gui-grounding
- android
- ui-automation
- multimodal
size_categories:
- 10K<n<100K
pretty_name: Android Control Dataset for LLaMA-Factory
---

# Android Control Dataset

## Overview

This directory contains two dataset files (`and_ctrl_train.json` and `and_ctrl_test.json`) derived from the [Android Control](https://github.com/google-research/google-research/tree/master/android_control) project by Google Research. These datasets have been formatted specifically for GUI grounding training in LLaMA-Factory.

## Dataset Description

The Android Control dataset consists of episodes where each episode contains multiple steps. Each step includes:
- **Step instructions**: Natural language instructions for UI interactions
- **Actions**: The type of action to perform (click, scroll, input text, etc.)
- **Coordinates**: Precise x, y coordinates for the action

The data has been extracted and formatted to train models for mobile UI understanding and interaction tasks.

## Files

- `and_ctrl_train.json`: Training dataset
- `and_ctrl_test.json`: Test/evaluation dataset
- `download_android_control.ipynb`: Jupyter notebook for downloading images and processing the original data

## Data Format

Each entry in the JSON files follows the LLaMA-Factory conversation format:

```json
{
  "messages": [
    {
      "role": "system",
      "content": "You are a helpful assistant that can identify what action to perform on mobile UI Screenshot given the user instruction."
    },
    {
      "role": "user",
      "content": "<image>Click on the Recording 2"
    },
    {
      "role": "assistant",
      "content": "{\"action_type\": \"click\", \"x\": 561, \"y\": 535}"
    }
  ],
  "images": ["and_ctrl/out_episode_18557_step_001.png"]
}
```

## Setup Instructions

To use these datasets in LLaMA-Factory:

1. **Create the image directory**:
   ```bash
   mkdir -p data/and_ctrl
   ```

2. **Download images**:
   Run the provided `download_android_control.ipynb` notebook to download and process the original images. The notebook will:
   - Download TFRecord files from Google Storage (`gs://gresearch/android_control/`)
   - Extract images and save them directly to `and_ctrl/` directory
   - Automatically organize images with the naming convention: `out_episode_{episode_id}_step_{step_number}.png`
   - Generate an `and_ctrl.json` file with the processed data

3. **Dataset files**:
   - Images: Stored in `data/and_ctrl/` folder
   - Training dataset: `and_ctrl_train.json` in `data/datasets/`
   - Test dataset: `and_ctrl_test.json` in `data/datasets/`

## Dataset Statistics

**Total samples**: Train: 82,944 | Test: 904

| Action Type | Train | Test |
|-------------|-------|------|
| click | 51,793 (62.44%) | 125 (13.83%) |
| scroll | 11,005 (13.27%) | 125 (13.83%) |
| input_text | 5,966 (7.19%) | 125 (13.83%) |
| wait | 5,657 (6.82%) | 125 (13.83%) |
| open_app | 5,572 (6.72%) | 125 (13.83%) |
| navigate_back | 2,909 (3.51%) | 125 (13.83%) |
| long_press | 42 (0.05%) | 125 (13.83%) |
| navigate_home | 0 (0.00%) | 29 (3.21%) |

**Note**: The training dataset shows a natural distribution with click actions being dominant (62.44%), while the test dataset is intentionally balanced with most action types having equal representation (~13.83% each). The `navigate_home` action appears only in the test set.

## Training Usage

These datasets are specifically formatted for training multimodal language models to:
- Understand mobile UI screenshots
- Ground natural language instructions to specific UI elements
- Generate precise action coordinates for UI automation
- Learn mobile app interaction patterns

## Source and Attribution

Original dataset: [Google Research Android Control](https://github.com/google-research/google-research/tree/master/android_control)

The Android Control dataset was created by Google Research for advancing mobile UI understanding and automation research.

### License

This dataset is derived from Google Research's Android Control dataset, which is licensed under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). The reformatted version for LLaMA-Factory maintains the same Apache 2.0 license terms.

Copyright for the original dataset belongs to Google LLC. Any modifications or reformatting for LLaMA-Factory compatibility are also provided under Apache License 2.0.

## Notes

- The images are referenced with relative paths starting with `and_ctrl/`
- Each action includes the action type and necessary parameters (coordinates, text, direction, etc.)
- The test set can be used for evaluating model performance on unseen mobile UI interactions