Yambda-5B Music Recommendation Dataset

## Overview Yambda-5B is a large-scale open music recommendation dataset released by Yandex, comprising 4.79 billion user-item interactions collected from 1 million users across 9.39 million tracks. The dataset distinguishes between organic interactions (user-initiated) and recommendation-driven interactions, and includes audio embeddings to support multi-modal ranking and retrieval research. Data is distributed in Parquet format and is one of the largest publicly available industrial recsys datasets. ## Schema - user_id — int — anonymized user identifier - item_id — int — anonymized track identifier - timestamp — int — interaction time - event_type — string — interaction type (listen, like, dislike, etc.) - is_organic — bool — whether interaction was organic or recommendation-driven - play_duration — float — listening duration where applicable - audio_embedding — array<float> — track audio embedding vector - album_id / artist_id — int — content metadata identifiers - +additional columns across interaction, item, and embedding tables ## Sources - yandex/yambda on HuggingFace — https://huggingface.co/datasets/yandex/yambda — Apache-2.0 - Companion paper: arXiv:2505.22238 ## Methodology Interactions were logged from Yandex Music's production service, separated into organic (user-initiated browsing/search) and recommendation-driven events. User and item identifiers are anonymized. Audio embeddings are precomputed by Yandex from track audio content. The dataset is provided at multiple scales (e.g., 50M, 500M, 5B) to support different compute budgets, and includes a global temporal split for reproducible benchmarking. ## Known gaps & limitations The dataset reflects the behavior of a single platform's user base (Yandex Music, predominantly Russian-speaking market) and may not generalize to other music ecosystems or cultures. All IDs are anonymized so no raw text metadata (track/artist names, lyrics) is available. Audio embeddings are opaque vectors — the encoder is not released, limiting reproducibility of embedding-side experiments. Source does not document demographic distribution; buyers should validate empirically. ## Intended use & out-of-scope - IS for: training and benchmarking sequential recommenders, two-tower retrieval models, ranking models, multi-modal recsys research, and offline evaluation of recommendation algorithms at industrial scale. - NOT for: content-based music analysis requiring raw audio or lyrics, user demographic studies, or production deployment without bias/fairness review — the data reflects one platform's exposure policies. _Federated dataset: 6 parquet shards, 84.72 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: Yambda-5B: Large-Scale Music Recommendation Dataset (4.79B Interactions) Industrial-scale music recommendation dataset from Yandex with 4.79B user-item interactions across 1M users and 9.39M tracks, including organic and recommendation interactions plus audio embeddings.