Yambda-5B-Star

Star-schema Parquet for the full Yambda-5B interaction log. Each official interaction is stored once. Album links, artist links, and audio embeddings stay in side tables and are not joined onto the events.

This is not the official repository. The official release also distributes per-type files, a sequential list per user, and 50M and 500M user samples. Those are the same 5B events. A point-in-time history is a prefix of one user's rows in this table. The wide sequential files are that prefix already gathered.

Tables

events has 4,785,859,887 rows from flat/5b/multi_event.parquet, in the official file order. That order is (user_id, timestamp), and rows that share a timestamp keep the source order.

event_type	rows
listen	4,649,567,411
like	89,334,605
dislike	11,579,143
unlike	32,944,520
undislike	2,434,208

user_id is the official uid. There are 1,000,000 users, numbered 1 through 1000000, and 9,390,623 tracks, numbered 1 through 9390623. Every user has at least 11 events. The median user has 2,712 events and the longest has 27,738. Listen counts per user, over the 921,023 users with a listen, have median 3,076, p90 12,956, and p95 17,030.

timestamp is the released clock: seconds since the first event in the dataset, quantized by floor((t - t_start) / 5) * 5. Values run from 0 through 26,000,000. t_start is not part of the release, so this column is not an absolute timestamp. Rows with the same user_id and timestamp are ordered by source_row.

source_row is the 1-based row number in flat/5b/multi_event.parquet. It is unique across the log. (user_id, timestamp, item_id, event_type) is not a key: the official file contains repeated rows, and those repetitions are kept. played_ratio_pct and track_length_seconds are present for listens and null otherwise. played_ratio_pct ranges from 0 to 159. A value above 100 means the listener rewound. The official Listen+ label, played_ratio_pct >= 50, is not applied. is_organic is true when the official flag is 1.

album_item has 9,651,644 pairs and 3,367,691 albums. A track can belong to more than one album. 736,840 tracks have no album. artist_item has 9,271,906 pairs and 1,293,394 artists. 1,400 tracks have two artists, and 120,117 tracks have none. item_embedding has 7,721,749 tracks. embed and normalized_embed are the official float64 vectors of dimension 128. 1,668,874 tracks that occur in events have no embedding. Every embedding track occurs in events.

dataset_source is yambda and synthetic is false on every row. source_row on a side table is the 1-based row number in that official file.

Point-in-time prefix

For an event, the earlier history of its user is every events row with the same user_id and a smaller source_row. Inside a user, timestamp does not decrease, so this is the prefix before that row. Side rows have no clock. Join them on item_id.

The official global temporal split is not materialized. With the constants in the official benchmark code, train rows are timestamp < 25911800, the next 1,800 seconds are a gap, and test rows are timestamp >= 25913600.

Original sources

• Dataset: https://huggingface.co/datasets/yandex/yambda
• Paper: https://arxiv.org/abs/2505.22238

License

Apache-2.0, the license of the official Yandex release. The official dataset card additionally says that Yambda is published for scientific and research purposes.

Citation

@misc{ploshkin2025yambda5b,
      title={Yambda-5B: A Large-Scale Multi-modal Dataset for Ranking And Retrieval},
      author={Alexander Ploshkin and Vladislav Tytskiy and Alexey Pismenny and Vladimir Baikalov and Evgeny Taychinov and Artem Permiakov and Daniil Burlakov and Eugene Krofto and Nikolay Savushkin},
      year={2025},
      eprint={2505.22238},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2505.22238}
}