Why this matters
Chess game records are a compact, well-structured source of sequential decision data useful for analytics, engine evaluation, behavior modeling, and training ML agents. This dataset aggregates billions of rated standard games from Lichess into parquet files partitioned by year/month so you can query and slice at scale without converting raw PGNs.
What Sets It Apart
- Scale and cadence: Contains billions of rated standard games (dataset card reports ~6.7 billion entries) and is updated monthly, making it suitable for large-scale trend analysis and continual-data workflows.
- Analysis-ready schema: Rows include Event/Site/White/Black/Result, WhiteElo/BlackElo (Glicko2), UTCDate/UTCTime, ECO/Opening, TimeControl, Termination and movetext (with %clk since 2017). About 6% of games include Stockfish [%eval] annotations.
- Storage and tooling: Distributed as parquet, hive-partitioned by year/month for fast filter-by-time queries and easy use with DuckDB, polars, pandas, Dask or the Hugging Face datasets library.
Who it's for and tradeoffs
Great fit if you need large, labeled sequences of human chess play for model training, behavioral analysis, openings research, or engine benchmarking. Look elsewhere if you need game-level metadata not provided in the schema (the dataset omits a classic PGN Date header but provides UTCDate) or if you require fully curated PGNs with enriched annotations; PGN dumps are available separately from Lichess database site. Also expect substantial storage and compute needs when working at full scale.
