mrmlengineer/chess_games

Chess Games Dataset

This dataset is a compact, ML-oriented chess game dataset generated from angeluriot/chess_games plus additional 2025 games extracted from Lumbras Giga Base.

It keeps only the move sequence in UCI notation, a derived minimum Elo field, and a source label. This makes the dataset smaller and simpler to consume than the base dataset when the downstream task only needs legal move sequences and a coarse player-strength signal.

Dataset Summary

• Games: 15,272,234
• Files: 8 Parquet shards
• Compressed size: 2.48 GiB
• Split: train
• Format: Parquet

Data Sources

Source	Games	Share
LichessEliteDatabase	5,525,245	36.18%
LumbrasGigaBase	4,785,121	31.33%
PGNMentor	3,431,537	22.47%
TWIC	858,163	5.62%
Masters	322,166	2.11%
LichessBroadcast	105,329	0.69%
ChessOK	62,895	0.41%
Britbase	60,148	0.39%
Kingbase	59,317	0.39%
Convekta	37,536	0.25%
ChessNostalgia	24,592	0.16%
Chessopolis	123	0.00%
GamesOfGMs	62	0.00%
Total	15,272,234	100.00%

The base dataset is angeluriot/chess_games, which already includes games from multiple public chess sources.

The added 2025 games come from Lumbras Giga Base, using the 2025 OTB and Online PGN files. Special thanks to Lumbras Giga Base for collecting and curating the chess game sources used by the base dataset and this extension.

Dataset Structure

Each row represents one chess game.

Field	Type	Description
moves_uci	list[string]	Mainline game moves in Universal Chess Interface notation, for example e2e4.
min_elo	int64 or null	Minimum of White Elo and Black Elo when both ratings are available; otherwise null.
source	string	Source label preserved from the base dataset when available; added 2025 Lumbras games use LumbrasGigaBase.

Rows with null min_elo: 2,053,362.

Differences From angeluriot/chess_games

This dataset is not a full mirror of the base dataset. It intentionally removes metadata that is not needed for the target training workflows.

Removed fields:

• date
• white_elo
• black_elo
• end_type
• winner
• moves_san
• moves_custom

Added or changed fields:

• min_elo is derived from white_elo and black_elo.
• source preserves the original base-dataset source label when available.
• 2025 Lumbras Giga Base OTB and Online PGN games are appended.
• Rows are sorted by descending min_elo, with null values last.

Usage

Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("mrmlengineer/chess_games")

game = dataset["train"][0]
print(game["moves_uci"])
print(game["min_elo"])
print(game["source"])

Streaming

from datasets import load_dataset

dataset = load_dataset(
    "mrmlengineer/chess_games",
    streaming=True,
)

for game in dataset["train"]:
    print(game["moves_uci"])
    break

License

This combined dataset is released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International because Lumbras Giga Base states that its work is licensed under CC BY-NC-SA 4.0. The base dataset angeluriot/chess_games is distributed on Hugging Face with an MIT license.

Users should respect the terms of both upstream sources.

Limitations

• The dataset has a single train split; no validation or test split is provided.
• min_elo is null when either player Elo is unavailable.
• Only the mainline UCI move sequence is included.
• Commercial use is restricted by the CC BY-NC-SA 4.0 license.