The YouTube-100M dataset contains 100 million YouTube videos. It includes 70 million training videos, 10 million evaluation videos, and 20 million validation videos. Each video is 4.6 minutes long on average, for a total of 5.4 million training hours.

The dataset contains more than 30K class annotations, which are automatically generated based on a combination of each video’s metadata (title, description, comments, etc.), background, and image content. On average, there are about 5 annotations per video.