OmniVideo-R1: 100K-1M Records for Audio-Visual Reasoning

OmniVideo-R1 provides between 100,000 and 1,000,000 preprocessed records for audio-visual reasoning, published by jankin123 in March 2026. The collection supports a two-stage training framework for multimodal models, specifically focusing on Query-Intensive (QI) grounding and modality attention.

Use Cases

• Training Visual Question Answering (VQA) models using audio-visual grounding data
• Developing Video Text-to-Text systems that synchronize auditory and visual reasoning
• Applying reinforcement learning to multimodal tasks using query-intention signals

Strengths

• Scale of 100,000 to 1,000,000 records
• Multimodal integration of audio, video, and text
• Apache 2.0 permissive license

Limitations

• Preprocessed specifically for the OmniVideo-R1 framework
• Lack of explicit column documentation in the metadata

Provenance

Source

Arxiv paper 2602.05847

Collection Method

Preprocessed from existing audio-visual sources for the OmniVideo-R1 framework

Freshness

Last updated March 2026.