CaptchaArena: A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs
CaptchaArena releases 50K verified interactive CAPTCHA trajectories and a 9B agent reaching 71.7 Pass@1.
CaptchaArena is a training dataset for interactive CAPTCHA solving, with 50K puzzles across 20 types and five interaction modes and every solution verified by execution. It provides 50K screenshot-action trajectories, including 46K with step-by-step reasoning, plus pixel-mask annotations for irregular targets. The authors train CaptchaAgent, one 9B policy for all 20 types, using supervised fine-tuning followed by reinforcement learning rewarded by an environment verifier. SFT reaches 70.5 Pass@1 and RL improves it to 71.7, with gains on two external benchmarks; the dataset and agent are released.
- 50K execution-verified puzzles cover 20 CAPTCHA types
- Five interaction modes and pixel masks for irregular targets
- 46K of 50K trajectories include step-by-step reasoning
- 9B CaptchaAgent reaches 71.7 Pass@1 after reinforcement learning
- Products
- CaptchaArena
- AI models
- CaptchaAgent
From the original article
1 min read at huggingface.co
Interactive CAPTCHAs remain challenging for computer-use agents, while existing datasets face trade-offs among type coverage, interaction fidelity, and trajectory supervision. To address these gaps, we present CaptchaArena, the first large-scale, fine-grained training dataset for interactive CAPTCHA solving.
By Zhenhao Zhang, Zhaoyu Fan, Haohan Ying · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.