Update README.md

README.md

@@ -6,9 +6,19 @@ task_categories:
 - image-text-to-text
 pretty_name: PixelReasoner-RL-Data
 ---
+**Overview.**
+The RL data for training [**Pixel Reasoner**: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning](https://arxiv.org/abs/2505.15966). 
 
-Under Construction!
+**Details.**
+The data includes 15,402 training queries with verifierable answers. The key fields include:
+- question, answer, qid
+- is_video: a flag to distinguish video and image queries
+- image: a list of image paths.
+  - For video-based queries, the path list corresponds to 16 extracted frames from the original video
+  - For image-based queries, the path list corresponds to multiple images
 
-The RL data used in [Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning](https://arxiv.org/abs/2505.15966). The code can be found at https://github.com/TIGER-AI-Lab/Pixel-Reasoner/tree/main
+**Note**: Remember to unzip `images.zip` and `videos.zip`, and replace the relative path in `image` field with the absolute path.
 
-Project page: https://tiger-ai-lab.github.io/Pixel-Reasoner/
+**Training Code**: The RL code can be found at https://github.com/TIGER-AI-Lab/Pixel-Reasoner/tree/main
+
+**Project page**: https://tiger-ai-lab.github.io/Pixel-Reasoner/