Video Question Answering Benchmark with 3200 Annotations

Annotation data for the Video-MME-v2 benchmark, containing 800 1080p MP4 video files and 3200 question-answer pairs stored in a Parquet file. The dataset was created by MME-Benchmarks and the repository was last updated in April 2026.

Use Cases

• Benchmarking video question answering models using the 3200 QA pairs in the test.parquet file.
• Training models for temporal understanding by analyzing sequences across the 800 MP4 video files.
• Evaluating multimodal reasoning on videos paired with textual subtitles from the subtitle.zip archive.
• Developing models that integrate visual information from 1080p videos with structured QA annotations.

Strengths

• Contains 800 high-resolution 1080p MP4 video files.
• Includes 3200 question-answer pairs for structured evaluation.

Limitations

• The dataset size in gigabytes and specific video lengths are unknown.
• Potential class or content distribution bias across the 800 videos is not detailed.

Provenance

Source

MME-Benchmarks

Collection Method

Annotation data created for a benchmark, as described in the associated paper.

Freshness

Repository last updated on 2026-04-06.