The YouTube 8M Dataset is a large scale labeled video dataset designed for advancing research in video understanding. It consists of millions of YouTube video IDs along with high quality machine generated annotations.

This dataset provides pre extracted visual and audio features, making it highly efficient for training machine learning and deep learning models without requiring raw video processing.

Context

Video understanding is a key challenge in AI, involving tasks such as classification, tagging, and recommendation. This dataset was created to support scalable research in these areas.

Source

Originally released by Google Research

Derived from publicly available YouTube videos

Feature extraction done using deep neural networks

Inspiration

This dataset is widely used in:

Kaggle competitions

Research in multi-label classification

Building recommendation systems

Learning video embeddings This dataset is a flattened, mean pooled version of the YouTube 8M trainpj.tfrecord shard.

License

CC BY 4.0 (Attribution License) or Google Data License (same as original YouTube-8M terms)

Each row represents one video. Features f0-f1023 are Inception-V3 visual embeddings; f1024-f1151 are VGGish audio embeddings.

Created to allow training on YT8M features using standard CSV-based tools like XGBoost and Scikit-Learn without needing a massive TensorFlow pipeline.

The data originates from Google's YouTube-8M project CC BY 4.0 (Creative Commons Attribution) license, as the original dataset is intended for public research.

This both the original .tfrecords and a Parquet representation of the YouTube 8 Million dataset. YouTube-8M is a large-scale labeled video dataset that consists of millions of YouTube video IDs, with high-quality machine-generated annotations from a diverse vocabulary of 3,800+ visual entities. It comes with precomputed audio-visual features from billions of frames and audio segments, designed to fit on a single hard disk. This dataset also includes the YouTube-8M Segments data from June 2019. This dataset is 'Lakehouse Ready'. Meaning, you can query this data in-place straight out of the Registry of Open Data S3 bucket. Deploy this dataset's corresponding CloudFormation template to create the AWS Glue Catalog entries into your account in about 30 seconds. That one step will enable you to interact with the data with AWS Athena, AWS SageMaker, AWS EMR, or join into your AWS Redshift clusters. More detail in (the documentation)[https://github.com/aws-samples/data-lake-as-code/blob/roda-ml/README.md.