Add dataset card, link to paper and GitHub

README.md

@@ -0,0 +1,44 @@
+---
+license: mit
+task_categories:
+- image-text-to-text
+tags:
+- vlm
+- self-play
+- reinforcement-learning
+---
+
+# Vision-Zero Dataset
+
+This repository contains the dataset introduced in the paper [Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play](https://huggingface.co/papers/2509.25541).
+
+Vision-Zero is a label-free, domain-agnostic multi-agent self-play framework for self-evolving Vision-Language Models (VLMs). It trains models through competitive "Who Is the Spy"-style visual games generated from arbitrary image inputs, enabling autonomous training data generation without human annotation.
+
+## Dataset Components
+
+The framework utilizes three distinct types of image datasets to enhance reasoning across diverse domains:
+
+1.  **CLEVR-based**: Synthetic scenes for logical reasoning.
+2.  **Chart-based**: Data for enhanced chart question answering and document understanding.
+3.  **Real-World**: Real-world image data for diverse vision-centric understanding scenarios.
+
+## Links
+
+- **Paper**: [Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play](https://huggingface.co/papers/2509.25541)
+- **Code**: [https://github.com/wangqinsi1/Vision-Zero](https://github.com/wangqinsi1/Vision-Zero)
+
+## Citation
+
+If you find Vision-Zero useful in your research, please consider citing the paper:
+
+```bibtex
+@misc{wang2025visionzeroscalablevlmselfimprovement,
+    title={Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play}, 
+    author={Qinsi Wang and Bo Liu and Tianyi Zhou and Jing Shi and Yueqian Lin and Yiran Chen and Hai Helen Li and Kun Wan and Wentian Zhao},
+    year={2025},
+    eprint={2509.25541},
+    archivePrefix={arXiv},
+    primaryClass={cs.CV},
+    url={https://arxiv.org/abs/2509.25541}
+}
+```