WebWorldData: 1.06 Million Web Interaction Trajectories for Browser World Models

WebWorldData is a large-scale collection of 1.06 million web interaction trajectories gathered from the open web. It was created by Qwen and last updated in February 2026. The dataset is designed for training browser world models and consists of sequences of state, action, and next-state transitions.

Use Cases

• Training browser automation agents based on sequences of state-action transitions.
• Developing world models for web environments based on accessibility tree representations.
• Researching reinforcement learning algorithms for web-based tasks using real interaction data.
• Benchmarking the performance of agents that interact with dynamic web pages.

Strengths

• Contains 1.06 million trajectories, providing a substantial scale for model training.
• States are represented as A11y Trees extracted from real websites, offering a structured view of web pages.
• Data is explicitly designed for and used to train the WebWorld model series.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Qwen via Hugging Face.

Collection Method

Collected from the open web using Playwright to extract A11y Trees.

Time Range

null

Freshness

Last updated 2026-02-18 01:06:41; freshness should be verified.

Geography

null