SKU AGT-4007 · Sold by External
Mind2Web
Product specifications
| SKU | AGT-4007 |
|---|---|
| Data type | Agent traces |
| Volume | 2,350 human-demonstrated web task trajectories (1,009 train + 1,341 test: 252 Cross-Task + 177 Cross-Website + 912 Cross-Domain) |
| Size on disk | ~6.28 GiB total on Hugging Face (11 train JSON files ~5.93 GB + test.zip 567.75 MB + scores_all_data.pkl 245.19 MB); auto-converted train parquet is 538.69 MiB |
| Format | JSON (task records; each action embeds raw/cleaned HTML snapshots and DOM element candidates) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open dataset (Hugging Face, CC-BY-4.0) |
| Quality score | — |
| License | Creative Commons Attribution 4.0 International (CC-BY-4.0) |
Mind2Web is a dataset for developing and evaluating generalist web agents that follow natural-language instructions to complete complex tasks on real websites. It contains over 2,000 open-ended tasks (2,350 task demonstrations counting all splits) collected by human annotators from 137 websites spanning 31 domains, with a training split of 1,009 tasks and three generalization test splits (Cross-Task 252, Cross-Website 177, Cross-Domain 912). Each task record stores the confirmed instruction plus an ordered action sequence; every action captures a raw and cleaned HTML snapshot of the page, an operation (CLICK, TYPE, or SELECT with an optional value), and positive/negative DOM element candidates identified by backend_node_id. The dataset was produced by the OSU NLP Group (The Ohio State University) and released with the NeurIPS 2023 Spotlight paper 'Mind2Web: Towards a Generalist Agent for the Web'. The data is licensed CC-BY-4.0 while the accompanying code is MIT-licensed. Because the raw and cleaned HTML per step is very large, the dataset totals roughly 6.3 GiB on disk.