WebChain is the largest open-source dataset of human-annotated trajectories collected from real-world websites, designed to accelerate reproducible research in web agents.
Unlike synthetic datasets that are blocked by anti-bot detection, CAPTCHAs, or authentication walls, WebChain is built entirely by human annotators operating on live, diverse websites — capturing the complex, authenticated workflows that matter most.
| Metric | Value |
|---|---|
| Total Trajectories | 31,725 |
| Total Interaction Steps | 317,993 |
| Unique Domains | 428 |
| Avg. Trajectory Length | 10.02 steps |
| Avg. Task Duration | 1.07 min |
Compared to existing web interaction datasets (Mind2Web, WebLINX, GUIAct):
- 10× more trajectories than the next largest human-annotated dataset
- Covers authenticated workflows (e-commerce checkout, banking, travel booking) that synthetic methods cannot capture
- Triple Alignment: every step synchronizes visual, structural, and action data
- Fully open-source — data is all public
The cleaned, training-ready v2 dataset is hosted on Hugging Face: webagentlab/webchain.
- Trace-level multi-turn SFT:
data/trace_sft/ - Seed/Demo sliding-window SFT:
data/seed_sft/
The dataset card contains schemas, loading examples, metadata documentation, a 150-trace evaluation subset, and package-verification instructions.
Each step in a trajectory contains Triple Alignment across three modalities:
| Modality | Fields |
|---|---|
| Visual Context | Full-page screenshot, viewport-specific crop |
| Structural Context | HTML snapshot, Accessibility (AX) tree |
| Action Grounding | Action type, pixel coordinates, bounding box, CSS selector, XPath |
| Reasoning | Chain-of-Thought (CoT) rationale for the action |
The figure below illustrates a real trajectory and the multi-dimensional annotation captured at each step, including visual observations, structural semantics, and behavioral supervision.
WebChain spans 428 unique domains across diverse categories:
- Shopping & E-commerce
- Flight & Travel Booking
- Accommodation & Hotels
- Real Estate & Property
- Fintech & Banking
- And many more...
WebChain is built with a three-stage pipeline ensuring both scale and quality:
Stage 1 — Constraint-Based Task Synthesis An LLM conditioned on a site's extracted functional schema generates grounded, executable tasks stratified by complexity (simple retrieval → multi-constraint navigation → conditional dependency tasks).
Stage 2 — Human-in-the-Loop Trajectory Collection Human annotators complete tasks via WebChain Builder, a passive logging tool that records DOM snapshots, precise action coordinates, element metadata, and timestamps for every step.
Stage 3 — Post-processing Contextual Enrichment
- Visual Grounding Densification: All interactive elements on each viewport are parsed and annotated (not just the clicked element), enabling dense layout-aware supervision.
- Synthetic CoT Generation: A VLM generates chain-of-thought rationales for each action, teaching agents interpretable multi-step planning.
We propose Dual Mid-Training, a training paradigm that decouples spatial grounding from long-horizon planning:
- LCRL — Long-Chain-oriented RLVR post-training on planning trajectories
- CoT-SFT — Supervised fine-tuning on CoT-augmented data for planning initialization
- SGRL — Spatial Grounding RLVR training with Visual Grounding Densification
The table below reports detailed comparisons across AC-High/Low, GUI-Act-Web, GUI-Odyssey, OA-Desktop, and OA-Web benchmarks.
Models trained on WebChain achieve the strongest overall performance, with WebChain-7B (+SGRL+CoT-SFT) reaching 81.4 overall.
We demonstrate a clear positive correlation between data volume and long-horizon planning performance — confirming that WebChain's scale is essential for unlocking robust web agent capabilities.
If you find WebChain useful in your research, please cite our paper:
@article{fan2026webchain,
title = {WebChain: A Large-Scale Human-Annotated Dataset of Real-World Web Interaction Traces},
author = {Sicheng Fan and Rui Wan and Yifei Leng and Gaoning Liang and Li Ling and Yanyi Shang and Dehan Kong},
journal = {arXiv preprint arXiv:2603.05295},
year = {2026}
}The dataset is available for academic and educational use under the WebChain Academic Research License v1.0. Commercial use requires separate written permission from the authors.
For questions or collaboration, please open an issue or reach out via the paper contact information.