Hugging Face's logo

Datasets:
ZHEN-04
/
CaptchaArena

Tasks:
Image-Text-to-Text
Modalities:
Image
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
captcha
gui-agent
computer-use
multimodal
vision-language
benchmark
License:
Dataset card Data Studio Files
xet
 Community
1

Request access to CaptchaArena

CaptchaArena is released under CC BY-NC 4.0 for non-commercial academic research only. The paper is on arXiv (2609.31957), and access requests are now being reviewed.

By requesting access you agree to use CaptchaArena solely for non-commercial academic research; commercial use is prohibited. The authors review each request manually, so please give your affiliation and intended research use.

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

CaptchaArena

A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs

Zhenhao Zhang1,*, Zhaoyu Fan2, Haohan Ying3, Jingwen Hu3, Hancen Fan1,
Junhao Zhou4, Zitian Chen1, Linchao Zhu2,†
1Columbia University  ·  2Zhejiang University
3University of Rochester  ·  4University of Illinois at Urbana-Champaign
*Project lead  ·  †Corresponding author

arXiv Code Trajectories Model License: CC BY-NC 4.0

The 20 CAPTCHA types

Overview

CaptchaArena is a large-scale, fine-grained dataset for training computer-use agents on interactive CAPTCHAs.

  • 50K puzzles across 20 CAPTCHA types and 5 interaction modes.
  • Execution-verified: every solution is replayed against the environment's verifier.
  • Fine-grained supervision: an executable action solution for every puzzle, and pixel-mask annotations for irregular targets.
  • Live interaction: agents observe the page through a fixed 1280×1080 viewport and act with low-level pointer and keyboard actions.
Interaction mode CAPTCHA types
Single-click Geometry Click, Misleading Click, Pick Area, Place Dot, Select Animal
Multi-click Bingo, Click Order, Image Recognition, Patch Select, Unusual Detection
Arrow-cycle Connect Icon, Coordinates, Dart Count, Image Matching, Object Match, Path Finder, Rotation Match
Real-time Hold Button, Slide Puzzle
Text-entry Dice Count

Data

Train Val Test Total
Puzzles per type 2,100 200 200 2,500
Puzzles 42,000 4,000 4,000 50,000

Each split has one folder per type, named <Split>/<Type>_<N>/ (e.g. Train/Bingo_2100/, Test/Bingo_200/). A folder holds the puzzle images and two label files:

  • ground_truth.json: the answer and the fields needed to render the puzzle.
  • ground_truth_cu.json: the same entries plus answer_cu, the solution as an executable action sequence.
hf auth login
hf download ZHEN-04/CaptchaArena --repo-type dataset --local-dir data

Intended use

CaptchaArena runs entirely in a self-hosted environment; no experiment interacts with a live CAPTCHA service. Its imagery is generated or openly licensed. Because CAPTCHA solving is a dual-use capability, the dataset is released for agent research and the study of CAPTCHA robustness, not for bypassing protections on services you do not own.

License

Released under CC-BY-NC-4.0 (Creative Commons Attribution–NonCommercial 4.0). Non-commercial academic research use only — commercial use is prohibited. Access is gated: you must request access and agree to these terms before downloading. Please attribute when using this dataset.

Citation

If you use CaptchaArena, please cite:

@article{zhang2026captchaarena,
  title   = {CaptchaArena: A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs},
  author  = {Zhang, Zhenhao and Fan, Zhaoyu and Ying, Haohan and Hu, Jingwen and Fan, Hancen and Zhou, Junhao and Chen, Zitian and Zhu, Linchao},
  journal = {arXiv preprint arXiv:2609.31957},
  year    = {2026}
}
Downloads last month
85
Total file size:
53.9 GB

Models trained or fine-tuned on ZHEN-04/CaptchaArena

Paper for ZHEN-04/CaptchaArena