GUIDE: Resolving Domain Bias in GUI Agents through Real Time Web Video Retrieval and Plug and Play Annotation G UI U nbiasing via I nstructional video D riven E xpertise Accepted to ECCV 2026 Project Page Paper GitHub This dataset supports the accepted ECCV 2026 paper "GUIDE: Resolving Domain Bias in GUI Agents through Real Time Web Video Retrieval and Plug and Play Annotation" . Overview GUIDE (GUI Unbiasing via Instructional Video Driven Expertise) is a training free framework that resolves domain bias in GUI agents by retrieving tutorial videos from YouTube and automatically generating domain specific Planning and Grounding knowledge. This dataset contains the complete video corpus, automated annotations, and pre computed knowledge for reproducing all experiments on the OSWorld benchmark. Dataset Contents 1. Tutorial Videos ( videos/ , ~21 GB) ~427 YouTube tutorial videos covering 10 desktop application domains, retrieved via GUIDE's subtitle driven Video RAG pipeline. Each video directory contains: Subdirectory Description video/ Original MP4 video file meta/ yt dlp metadata JSON (title, duration, upload date, formats) subtitle/ Video subtitle files audio/ Extracted audio (MP3)…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy