TIGER-Lab/PixelReasoner-RL-Data

1---2language:3- en4license: apache-2.05task_categories:6- image-text-to-text7pretty_name: PixelReasoner-RL-Data8configs:9- config_name: default10  data_files:11  - split: train12    path: "release.parquet"13---14**Overview.**15The RL data for training [**Pixel Reasoner**: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning](https://arxiv.org/abs/2505.15966),16The queries require fine-grained visual analysis in both images (e.g., infographics, visually-rich scenes, etc) and videos. 17 18**Details.**19The data includes 15,402 training queries with verifierable answers. The key fields include:20- question, answer, qid21- is_video: a flag to distinguish video and image queries22- image: a list of image paths.23  - For video-based queries, the path list corresponds to 16 extracted frames from the original video24  - For image-based queries, the path list corresponds to multiple images25 26**Note**: Remember to unzip `images.zip` and `videos.zip`, and replace the relative path in `image` field with the absolute path.27 28**Training Code**: The RL code can be found at https://github.com/TIGER-AI-Lab/Pixel-Reasoner/tree/main29 30**Project page**: https://tiger-ai-lab.github.io/Pixel-Reasoner/