Skip to content

Dataset · ChilleD

Offline copies of 153 websites for GUI agent training

ChilleD/WebHarbor

An environment of 153 real websites in Docker and 3,518 tasks with automated verification for training and evaluating web agents.

What's inside

Contains offline reconstructions of 153 real English websites, including Amazon, Booking, GitHub, and BBC News, preserving layouts, product images, login forms, shopping carts, and bookings. Each website is containerized in Docker with its own database and resets to its original state in seconds. Includes 3,518 tasks with deterministic verifiers to evaluate performance.

What you can do with it

  • Training and testing multimodal browser navigation GUI agents
  • Benchmarking image-text-to-text models in real web scenarios
  • Reproducible evaluation of autonomous agents interacting with forms, carts, and accounts
Tasks
Image text → text
Languages
English
Downloads this month
23.7K
Likes
6

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("ChilleD/WebHarbor")