GUI-Net-Video

A collection of video URLs for GUI-related tasks, organized by platform or application.

This dataset provides the video URLs collected for TongUI: Internet-Scale Trajectories from Multimodal Web Tutorials for Generalized GUI Agents — a framework that turns multimodal web tutorials (including YouTube videos) into large-scale GUI agent trajectories. Each URL is paired with a GUI task description (the .txt filename) and grouped by platform/app.

• Paper: arXiv:2504.12679
• Project: tongui-agent.github.io
• Code: TongUI-agent/TongUI-agent

Stats: 32 platforms · 2,627 task files · 126,506 URLs (~2 MB as urls.zip)

Structure

urls/
├── {Platform}/          # e.g. Youtube, Chrome, Google, Windows, ...
│   └── {Task name}.txt  # one URL per line

Each .txt file is named after a GUI task (e.g. Viewing search results.txt) and contains one YouTube (or related) URL per line.

Platforms

31 platforms/apps are included: Alibaba, Amazon, Android, Aol_search, Baidu, Bing, Chrome, Discord, Etsy, Excel, Facebook, Github, Google, Instagram, Linkedin, MATLAB, Macos, Notion, PPT, Pinterest, Reddit, Telegram, Tiktok, Twitter, Ubuntu, VScode, Wechat, Whatsapp, Windows, Word, Yandex, Youtube.

Usage

from pathlib import Path

for txt in Path("urls").rglob("*.txt"):
    platform = txt.relative_to("urls").parts[0]
    task = txt.stem
    urls = [line.strip() for line in txt.read_text().splitlines() if line.strip()]
    print(platform, task, len(urls))