Mind2Web

Mind2Web is the premier dataset for training and evaluating generalist web agents: it enables agents to execute complex, language-instructed tasks on any real-world website.

Mind2Web is a large-scale dataset specifically engineered to advance generalist web agents (LLMs). The dataset features over 2,000 open-ended tasks collected from 137 real-world websites across 31 diverse domains, moving beyond simplified simulations. This scale provides three critical components: diverse domains, authentic websites, and a broad spectrum of user interaction patterns (click, select, type). The associated two-stage model, MindAct, demonstrates efficiency by using a small language model (LM) to filter web elements before a large language model (LLM) predicts the final action, significantly improving performance and generalization across unseen websites and domains.

https://osu-nlp-group.github.io/Mind2Web