Datasets:
Tasks:
Visual Document Retrieval
Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
< 1K
ArXiv:
License:
| language: | |
| - en | |
| license: apache-2.0 | |
| size_categories: | |
| - n<1K | |
| task_categories: | |
| - visual-document-retrieval | |
| tags: | |
| - Multimodal_Agents | |
| - Open_Source_CAPTCHAs | |
| # Open CaptchaWorld Dataset | |
| This dataset accompanies the paper [Open CaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents](https://huggingface.co/papers/2505.24878). It contains 20 distinct CAPTCHA types, each testing different visual reasoning capabilities. The dataset is designed for evaluating the visual reasoning and interaction capabilities of Multimodal Large Language Model (MLLM)-powered agents. | |
| [Project Page](https://huggingface.co/spaces/OpenCaptchaWorld/platform) | [Github](https://github.com/Yaxin9Luo/Open_CaptchaWorld) | |
| The dataset includes: | |
| * **20 CAPTCHA Types**: A diverse set of visual puzzles testing various capabilities. See the Github repository for a full list. | |
| * **Web Interface**: A clean, intuitive interface for human or AI interaction. | |
| * **API Endpoints**: Programmatic access to puzzles and verification. | |
| This dataset is useful for benchmarking and improving multimodal AI agents' performance on CAPTCHA-like challenges, a crucial step in deploying web agents for real-world tasks. The data is structured for easy integration into research and development pipelines. |