Video Reasoning Training Dataset for Perception and Embodied Tasks

Videoreason Training is a large-scale dataset comprising 471,575 samples across three subsets: perception (176,907 samples), simulation (105,818 samples), and embodied (188,850 samples). It was created by Zane-QIU and last updated on HuggingFace in April 2026. The dataset is designed for training models on video reasoning tasks spanning visual perception, 3D simulation, and robotics.

Use Cases

• Train segmentation models using the perception subset's segmentation labels.
• Develop video super-resolution models using the low-resolution and high-resolution video pairs in the perception subset.
• Simulate camera motion for 3D scene navigation using the camera pose sequences in the simulation subset.
• Train embodied AI agents for robotic manipulation using the action and state sequences in the embodied subset.

Strengths

• 471,575 total video samples
• Covers three distinct reasoning domains: perception, simulation, and embodied tasks

Limitations

• Specific column names, data formats, and sample sizes per task within subsets are unknown
• No information on video resolution, frame rate, or annotation quality

Provenance

Source

Zane-QIU on HuggingFace

Freshness

Last updated April 2026.