YouTube – 8M is a large labeled video dataset consisting of millions of YouTube video IDs. The more than 3,800 visual entity annotations it contains are all machine-generated based on a vocabulary. The pre-computed nature of the dataset allows it to be used on a single hard drive, which means it is possible to train a baseline model on the dataset using a single GPU.

At the same time, the scale and diversity of the dataset allow users to conduct in-depth exploration of complex audio-visual models, which would take weeks to train even with distributed training.

This dataset was released by Google in 2016.