Hugging Face's logo

Datasets:
LEONW24
/
Trace2World

Tasks:
Reinforcement Learning
Languages:
Chinese
English
Size:
1K<n<10K
ArXiv:
Tags:
gui-agents
reinforcement-learning-environments
browser
trace2world
License:
Dataset card Files
xet
 Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Trace2World environments are released for non-commercial research on GUI agents and
reinforcement learning (CC BY-NC 4.0). Requests are reviewed manually; please provide
the following information.

Log in or Sign Up to review the conditions and access this dataset content.

Trace2World: interaction traces converge into one source-validated app world that yields an executable app and learner-adaptive tasks

Trace2World: Reconstructing App Worlds from Real Interaction Traces for Scalable GUI Agent Training

Environments

The data release of Trace2World: Reconstructing App Worlds from Real Interaction Traces for Scalable GUI Agent Training. Trace2World separates global behavior reconstruction from implementation: it reconciles partial real-app traces into a shared, source-linked app model, compiles an executable environment, and accepts it only after independent replay of the original workflows. From 1,570 traces it constructs 50 environments; on 200 workflows across ten apps, 91.1% are reproduced end to end versus 59.8% for the strongest matched task-local baseline. The release bundles the reconstructed environments with their frozen evaluators and state-kernel data, together with the two task collections:

  • Train set — 2,048 optimization tasks: 1,280 seed tasks and 768 learner-guided tasks, drawn from the 40 policy-training applications.
  • Test set — Trace2World-Bench, 1,044 sealed tasks: 256 entity replacements, 512 held-out workflows, 100 cross-app returns, and 176 tasks from the ten unseen applications.

The code that hosts, evaluates and drives the environments, and the constructor that turns traces into environments, live in the GitHub repository Tongyi-MAI/trace2app.

Access

Access is gated and reviewed manually. Request access on this page with the form above; once approved, authenticate with a Hugging Face token (huggingface-cli login or HF_TOKEN) before downloading.

Usage 

git clone https://github.com/Tongyi-MAI/trace2app.git && cd trace2app
pip install ./trace2world_env huggingface_hub && playwright install chromium

python scripts/download_environments.py            # environments/ -> data/trace2world, verified against the manifest
t2w-serve --environments-dir data/trace2world --state-root /tmp/t2w-state --host 127.0.0.1 --port 6800

t2w-serve hosts the whole portfolio in one process and routes each task_id to its environment's pages, tasks, evaluator and state kernel. The HTTP API, action space and reward semantics are documented in trace2world_env/README.md; the construction path and its paper-algorithm mapping in trace2world_pipeline/README.md and WORKFLOW.md; the download / verification scripts in scripts/.

Layout 

environments_manifest.json      environments, per-environment task ids, SHA-256 of every file under environments/
environments/<id>/
├── resource_manifest.json      entry page, task table, evaluator, runtime profile, viewport
├── tasks.json                  tasks: task_id, instruction, start_page, initial_state
├── evaluators/                 frozen checkpoint / state evaluator plans and scripts
├── data/                       provenance, host graph, state-kernel data (SQLite where used)
└── app/                        generated pages, images, runtime scripts, instrumentation
train/                          Train set: 2,048 task manifests (1,280 seed + 768 learner-guided)
test/                           Test set: Trace2World-Bench, 1,044 sealed task manifests

Applications

All 50 applications provide construction traces and all 50 are reconstructed as one environment each. Policy optimization uses tasks from the 40 policy-training applications only; the ten held-out applications, one per domain, supply construction traces but no seed, candidate, screening or RL-update task.

Domain Policy-training apps (40) Held-out eval app (10)
Social & community WeChat, QQ, Weibo, Xiaohongshu Zhihu
News & short video Kuaishou, Bilibili, Toutiao, Tencent News Douyin
Long video & audio iQIYI, Youku, NetEase Cloud Music, Kugou Music, Ximalaya QQ Music
Commerce & resale Taobao, JD, Xianyu, Dewu Pinduoduo
Local services & food Meituan, Dianping, Dingdong Maicai, KFC, Hema Ele.me
Mobility, travel & tickets Amap, Ctrip, Didi, Railway 12306, Damai Taopiaopiao
Auto, housing & jobs 58Tongcheng, Autohome, Dongchedi, BOSS Zhipin Beike
Work, learning & search Feishu, WPS Office, Baidu, DingTalk, Youdao Dictionary, Baidu Netdisk Tencent Meeting
Payment & logistics Alipay, SF Express Cainiao
Health & fitness Boohee Health Keep

Trace collection is an evidence quota fixed before construction: an application with at least three independent service lines receives 50 traces, every other application 20 (19 multi-service apps x 50 + 31 single-service apps x 20 = 1,570). Trace2World-Bench (1,044 tasks: 256 entity replacements, 512 held-out workflows, 100 cross-app tasks, 176 tasks from unseen apps) is one frozen evaluation view over the environment portfolio; the 1,280 seed and 768 learner-guided optimization tasks are drawn from the 40 policy-training apps.

All environments render with a 430x884 CSS viewport at device scale factor 2 (860x1768 screenshots). Task instructions are in Chinese, matching the source apps.

Integrity

environments_manifest.json lists the SHA-256 and size of every file under environments/; scripts/verify_environments.py in the code repository checks a download against it. The environments are frozen exactly as used for the numbers reported in the paper; rewards come from the frozen checkpoint / evidence contracts and the host does not rewrite or relax them.

Scope and terms

  • The environments are reconstructed simulations of the applications' user-facing behavior, not scraped copies, and connect to no real service. Page content (names, listings, prices, addresses) is synthetic or was taken from the source traces; it does not describe real accounts.
  • Source traces were collected by consenting operators from accounts they control, with consent covering research use and public release. Screenshots are captured from public consumer apps.
  • This dataset is released for non-commercial research on GUI agents and reinforcement learning (CC BY-NC 4.0). Trademarks and copyrighted content visible in the screenshots and page content remain the property of their respective owners; application names identify the systems studied without implying affiliation or endorsement.

Citation 

@misc{wang2026trace2worldreconstructingappworlds,
      title={Trace2World: Reconstructing App Worlds from Real Interaction Traces for Scalable GUI Agent Training},
      author={Zezhou Wang and Hanzhang Zhou and Panrong Tong and Quyu Kong and Pengxiang Li and Steven Hoi and Zhuzhong Qian and Yue Wang},
      year={2026},
      eprint={2610.30792},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.30792},
}
Downloads last month
93
Total file size:
426 MB

Paper for LEONW24/Trace2World