TIGER-Lab/PixelReasoner-RL-Data

Overview. The RL data for training Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning, The queries require fine-grained visual analysis in both images (e.g., infographics, visually-rich scenes, etc) and videos. 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… See the full description on the dataset page: https://huggingface.co/datasets/TIGER-Lab/PixelReasoner-RL-Data.