Update task category and add paper link

This PR updates the `task_categories` metadata to accurately reflect the dataset's purpose: visual-document-retrieval. It also adds a link to the associated paper.

README.md

@@ -1,12 +1,26 @@
 ---
-license: apache-2.0
-task_categories:
-- question-answering
 language:
 - en
+license: apache-2.0
+size_categories:
+- n<1K
+task_categories:
+- visual-document-retrieval
 tags:
 - Multimodal_Agents
 - Open_Source_CAPTCHAs
-size_categories:
-- n<1K
----
+---
+
+# 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.