AI
CaptchaArena: A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 16:04:24 UTC
Briefing created: 30/09/2026, 05:07:57 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 04:18:25 UTC.
Briefings
CaptchaArena is a large-scale dataset containing 50K puzzles across 20 CAPTCHA types and 5 interaction modes, verified through execution. It provides 50K screenshot-action trajectories, with 46K including step-by-step reasoning annotations and fine-grained pixel-mask annotations for irregular targets. Using this dataset, researchers trained a single 9B policy model called CaptchaAgent, achieving a supervised fine-tuning Pass@1 of 70.5 and an RL improvement to 71.7. The dataset enables fine-grained supervision for interactive CAPTCHA solving, addressing gaps in type coverage and interaction fidelity.
Key points
- CaptchaArena contains 50K puzzles across 20 CAPTCHA types and 5 interaction modes, with every solution verified through execution.
- The dataset provides 50K screenshot-action trajectories, including 46K with step-by-step reasoning annotations and fine-grained pixel-mask annotations for irregular targets.
Context and limitations
The dataset covers 20 CAPTCHA types and 5 interaction modes; specific performance metrics on external benchmarks are reported but not detailed.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 05:07:57 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeih2nc2g5p6v4gb3rywvw4qlomkpg3dhoy3l6q5jt3f64r2vdkl54u