Datasetpublic

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.

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face

2likes164downloads

Dataset Card

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 path list corresponds to 16 extracted frames from the original video
—For image-based queries, the path list corresponds to multiple images

Note: Remember to unzip images.zip and videos.zip, and replace the relative path in image field with the absolute path.

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/