metadata
license: mit
task_categories:
- text-generation
language:
- en
tags:
- agent
Supported Tasks
- Natural Language → GUI Action Grounding Convert user instructions into JSON action objects.
- Instruction Following Models learn to interpret varying natural language formulations (e.g., “press submit” vs “click the submit button”).
- Multi-step UI Automation Some samples involve sequences of actions (e.g., open site → type → press Enter → screenshot).
Languages
- English (
en) - Generated with simple variations (synonyms, phrasings).
Dataset Structure
Data Format
Each entry is a JSON object with:
{
"instruction": "Search Google for Python Playwright.",
"actions": [
{"action": "type", "target": "textarea[name=q]", "value": "Python Playwright"},
{"action": "keypress", "options": {"key": "Enter"}}
]
}
- instruction: natural language input.
- actions: list of structured GUI actions in a standard schema.
JSON Schema for Actions
{
"action": "string",
"target": "string (CSS selector, text, XPath, etc.)",
"value": "string (optional, e.g., input text, file path)",
"options": { "key": "Enter", "button": "left", "count": 2, "direction": "down" }
}
Dataset Statistics
Size: 1,000 examples
Average actions per instruction: 1.7
Action types covered:
- Click, double-click, right-click
- Type, clear, keypress
- Scroll, hover, drag-drop
- Checkbox, radio, dropdown selection
- Upload, download
- Dialog handling (accept/dismiss)
- Screenshot, highlight
- Navigation (open_url, back, forward, refresh, switch_tab, resize)
Use Cases
- Training GUI grounding agents (LLM-based or hybrid).
- Creating instruction-tuned models for web automation.
- Benchmarking natural language → structured action translation.
- Bootstrapping RPA (Robotic Process Automation) agents with LLMs.
Generation Process
- Synthetic data generated with templates + variations.
- Actions derived from common web automation tasks (Google, YouTube, Gmail, Amazon, GitHub, Slack, Notion, Trello, etc.).
- Covers both single-step (click, type) and multi-step (search + click + screenshot) workflows.
Limitations
- Synthetic: No real human annotations.
- Web-centric: Mostly web app actions, fewer desktop/native app actions.
- Surface-level grounding: Uses simple selectors (
text=Login,input#username) rather than pixel-perfect or accessibility trees.
Licensing
- MIT License for dataset release.
- Free for research & commercial use.
- Attribution appreciated: “GUI Grounding Dataset (ArunKr, 2025)”.
Citation
If you use this dataset in your work, please cite:
@dataset{gui_grounding_2025,
author = {ArunKr},
title = {GUI Grounding Dataset},
year = {2025},
url = {https://huggingface.co/datasets/ArunKr/gui_grounding_dataset-100},
license = {MIT}
}