DanielRegaladoCardoso/lastfm-1k-twotower

Last.fm 1K — Two-Tower Ready是一个处理过的、适用于检索或推荐模型（如双塔模型、ALS、BPR等）的数据集，基于Last.fm 1K用户数据构建。它包含约19.1百万个听歌事件，覆盖1000名用户在2005年至2009年间的活动。数据集被划分为训练集（15,320,694行，144 MB）、验证集（1,915,086行，18 MB）和测试集（1,915,087行，18 MB）。每行数据包含(user_idx, track_idx, artist_idx, ts)字段，其中所有ID均为整数编码，可直接用于嵌入层，ts为UTC时间戳。词汇表包括992名用户、1,500,661个独特曲目和174,091名艺术家。拆分采用时间顺序方式（非随机），将最旧的80%数据作为训练集，接下来10%作为验证集，最后10%作为测试集，以避免前瞻偏差并支持标准的推荐系统评估。原始数据来自Last.fm 1K用户数据集，处理过程包括基于artist_name — track_name组合编码track_idx以去重、删除空值字段和处理UTF-8问题。数据集遵循Apache 2.0许可证，但原始交互数据遵循Last.fm 1K数据集的原始条款。A processed, train-ready version of the Last.fm 1K users dataset, prepared for retrieval/recommender models (two-tower, ALS, BPR, etc.). Built as part of loopback — an open-source two-tower neural recommender for music. It includes approximately 19.1 million listening events from 1000 users (2005–2009). The dataset is split into train (15,320,694 rows, 144 MB), val (1,915,086 rows, 18 MB), and test (1,915,087 rows, 18 MB). Each row contains (user_idx, track_idx, artist_idx, ts) — integer-encoded IDs ready to plug into an embedding layer, plus a UTC timestamp. Vocabulary includes 992 users, 1,500,661 unique tracks, and 174,091 artists. Splits are temporal, not random: oldest 80% for train, next 10% for val, last 10% for test, to avoid look-ahead bias and ensure honest RecSys evaluation. Original data sourced from Last.fm 1K users dataset, processed by encoding track_idx from artist_name — track_name combo, dropping nulls, and handling UTF-8 quirks. License is Apache 2.0 for the processing pipeline and integer encoding, with underlying interaction data following the original Last.fm 1K dataset terms.