VBVR-Dataset: 1 Million Video Clips for Spatiotemporal Reasoning

VBVR-Dataset provides 1,000,000 video clips organized into 100 curated reasoning task generators, released by Video-Reason in February 2026. Each record pairs a video segment with start/end frame indices and a textual reasoning prompt to facilitate spatiotemporal understanding.

Use Cases

• Training visual question answering models using the textual reasoning task field
• Temporal action localization using start_frame and end_frame markers
• Video classification across 100 generator-defined categories

Strengths

• 1,000,000 video clips
• 100 distinct reasoning task generators
• Spatiotemporal grounding via start and end frame markers

Limitations

• Training split only
• Potential for algorithmic bias inherent in the 100 task generators
• High storage and compute requirements for 1 million video samples

Provenance

Source

Video-Reason (Arxiv 2602.0159)

Collection Method

Curated via 100 reasoning task generators

Freshness

Last updated February 2026.