Yandex opens 4.79B-interaction music dataset Yambda-5B

Yandex has released Yambda-5B, a music-recommendation dataset covering 4.79 billion user-item interactions from 1 million users and 9.39 million tracks, published on Hugging Face under an Apache 2.0 license with an accompanying arXiv preprint.

The release comes in three scales (50M, 500M, and 5B interactions) and ships listens, likes, dislikes, unlikes, and undislikes, each tagged with a flag indicating whether the interaction was organic or driven by a recommender. The dataset card describes it as an open database that "provides distinctive markers for organic versus recommendation-driven interactions, along with precomputed audio embeddings" for 7.72 million tracks. Timestamps are binned into five-second units with global ordering, and a track counts as listened when more than 50% of its duration plays.

Audio embeddings were produced "using a convolutional neural network inspired by Contrastive Learning of Musical Representations (J. Spijkervet et al., 2021)," according to the card.

The dataset page reports 20,565 downloads in the last month and more than eight community models already trained on it, including one listed as IMPACT-5B. Two researchers we track posted the source link in our Who's Who feed within hours of the drop.