OmniVideo-100K: Audio-Visual Reasoning Dataset with Structured Scripts

OmniVideo-100K is an instruction-tuning dataset introduced in the paper 'OmniVideo-100K: A Dataset for Audio-Visual Reasoning through Structured Scripts and Evidence Chains'. It contains 100,000 training samples split into 70,000 open-ended and 30,000 multiple-choice questions. The dataset was created by MiG-NJU and was last updated on the Hugging Face platform in June 2026.

Use Cases

• Instruction-tuning video-language models based on the structured scripts and evidence chains.
• Training models for open-ended video question answering based on the 70,000 OE samples.
• Training models for multiple-choice video question answering based on the 30,000 MCQ samples.
• Benchmarking audio-visual reasoning capabilities based on the dataset's structured format.

Strengths

• Contains 100,000 total training samples for instruction-tuning.
• Provides a structured format with both open-ended and multiple-choice question types.
• Official dataset from a published research paper, suggesting a defined purpose.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the raw video files is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

MiG-NJU

Collection Method

Introduced in the associated research paper; likely curated for audio-visual reasoning.

Freshness

Last updated 2026-06-16 03:17:22; freshness should be verified.