Action100M: Large-Scale Instructional Video Dataset With Temporal Segments

Action100M provides 100 million temporally localized segments derived from 1.2 million instructional videos, totaling 14.6 years of video duration. The dataset, created by Facebook, includes open-vocabulary action supervision and captions for each segment. It was last updated on the platform in January 2026.

Use Cases

• Training video action recognition models using the open-vocabulary action labels for each temporal segment.
• Developing video-language models by aligning video segments with their corresponding rich caption text.
• Researching temporal localization methods on the 100 million annotated video segments.
• Benchmarking scalable video understanding systems on a dataset sourced from 1.2 million instructional videos.

Strengths

• Contains approximately 100 million temporally localized action segments.
• Sourced from a large collection of 1.2 million unique instructional videos.
• Provides 14.6 years of total video duration for analysis.
• Includes both video data and rich textual captions for multimodal tasks.

Limitations

• Specific column names, data schema, and exact file formats are not detailed in the provided description.
• The geographic and demographic origin of the 1.2 million source videos is unspecified, which may introduce bias.
• The preview nature of the dataset may indicate it is a subset or not the final release.

Provenance

Source

Constructed from 1.2 million Internet instructional videos.

Collection Method

Temporally localized segments were extracted from source videos and annotated with open-vocabulary actions and captions.

Freshness

Last updated on the platform in January 2026.