MSR-VTT Video Description Dataset With 200K Captions

MSR-VTT contains 10,000 video clips paired with 200,000 descriptive captions. The dataset, originally created by Microsoft Research, is a standard benchmark for text-video retrieval and captioning tasks. It was last updated on the platform in August 2025.

Use Cases

• Train text-to-video retrieval models using video clips and their associated captions.
• Benchmark video captioning performance on the standard 1K-A test split with 1,000 video-caption pairs.
• Fine-tune cross-modal encoders using the train_9k split containing 180,000 caption-video pairs.
• Evaluate model generalization across different video categories present in the MSR-VTT collection.

Strengths

• 10,000 unique video clips provide a substantial base for training.
• 200,000 captions offer multiple descriptive annotations per video.
• Standardized 1K-A split protocol enables consistent benchmarking.

Limitations

• Specific video content categories, geographic origin, and collection time range are not detailed in the provided description.
• The dataset's age (original research from 2016) may not reflect contemporary video styles or topics.

Provenance

Source

Originally created by Microsoft Research (MSR).

Collection Method

Video clips were collected from a commercial video library and annotated with captions by crowd workers.

Freshness

Last platform update was 2025-08-03, but the underlying dataset is from 2016.