Sign in to view source links and access this dataset
Description
31,725 real-world web interaction trajectories were collected and annotated by webagentlab. The dataset contains 317,993 steps across 428 unique domains, with a focus on aligning visual, structural, and action contexts for agent training. It was last updated on April 14, 2026.
Use Cases
Training spatial grounding models based on the triple alignment of visual and structural context.
Evaluating long-horizon planning capabilities of web agents using the annotated action sequences.
Benchmarking GUI agent performance across a diverse set of 428 web domains.
Supervising agent actions using the human-annotated trajectory and step data.
Strengths
Large scale with 31,725 annotated trajectories.
Detailed step-level data with 317,993 total steps.
Broad domain coverage across 428 unique web domains.
Core contribution of a Triple Alignment methodology for supervision.
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
webagentlab
Collection Method
Human-annotated collection of real-world web interaction trajectories.
Freshness
Last updated 2026-04-14 07:10:24; freshness should be verified.
Access policy details require checking the full description on the dataset page.