Pixel-Reasoner
Collection
Reasoning in the pixel space • 5 items • Updated • 4
Datasets:
Dataset Viewer
The dataset viewer is not available for this dataset.
Unexpected token '<', "<html>
<h"... is not valid JSON
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
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/
- Downloads last month
- 79