README.md · TIGER-Lab/PixelReasoner-RL-Data at main

metadata

language:
  - en
license: apache-2.0
task_categories:
  - image-text-to-text
pretty_name: PixelReasoner-RL-Data
configs:
  - config_name: default
    data_files:
      - split: train
        path: release.parquet

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/