thomas-schweich/pawn-lichess-full

PAWN Lichess Full

~289M rated Lichess games from Q1 2025 (January-March), pre-tokenized in the PAWN training format. Validation and test splits are uniformly sampled from January 2026 (10-month temporal gap from training data).

Splits

Split	Source	Games	Shards
train	Jan-Mar 2025	~289M	289
validation	Jan 1-14, 2026	50,000	1
test	Jan 15-31, 2026	50,000	1

Val/test games are uniformly random-sampled from their respective date ranges via a two-pass approach (header-only count, then index-sampled parsing). Seed 42 for val, seed 43 for test.

Schema

Column	Type	Description
tokens	list[int16]	PAWN token IDs per ply (variable length, max 255)
clock	list[uint16]	Seconds remaining per ply. 0 = no annotation.
eval	list[int16]	Centipawns from white's perspective. Mate = ±(32767-N). -32768 = no annotation.
game_length	uint16	Number of half-moves
result	string	1-0, 0-1, 1/2-1/2
white_player	uint64	xxHash64 of username (Polars 1.39.3)
black_player	uint64	xxHash64 of username
white_elo	uint16	Lichess rating
black_elo	uint16	Lichess rating
white_rating_diff	int16	Rating change from this game
black_rating_diff	int16	Rating change from this game
eco	string	ECO opening code (e.g. A03)
opening	string	Opening name
time_control	string	e.g. 600+0, 180+2
termination	string	Normal, Time forfeit, etc.
date	datetime[ms]	Game start time (UTC)
site	string	Lichess game URL

Sentinel values

• Clock 0: no clock annotation (rare for rated games)
• Eval -32768 (0x8000): no Stockfish eval (~92% of games lack evals)
• Eval ±(32767-N): mate in N (bit 14 always set for mate scores)
• Eval centipawns clamped to ±16383

Token vocabulary

4,278 tokens: 1 PAD + 4,096 grid moves (64×64 src/dst) + 176 promotions + 5 outcomes. See the PAWN architecture docs.

Usage

Direct loading with PAWN training pipeline

# Loads pre-tokenized data, ready for training — no parsing needed
from pawn.lichess_data import prepare_lichess_dataset

data = prepare_lichess_dataset(
    "thomas-schweich/pawn-lichess-full",
    max_games=500_000,
    min_ply=10,
)

Polars with predicate pushdown

import polars as pl

# Filter to 1800-1900 Elo band without downloading the full dataset
df = (
    pl.scan_parquet("hf://datasets/thomas-schweich/pawn-lichess-full/data/train-*.parquet")
    .filter(
        (pl.col("white_elo") >= 1800) & (pl.col("white_elo") < 1900) &
        (pl.col("black_elo") >= 1800) & (pl.col("black_elo") < 1900)
    )
    .head(50_000)
    .collect()
)

HuggingFace datasets

from datasets import load_dataset

ds = load_dataset("thomas-schweich/pawn-lichess-full", split="train", streaming=True)
for game in ds.take(5):
    print(game["tokens"][:5], game["result"], game["white_elo"])

Eval coverage

Approximately 8.4% of games include Stockfish eval annotations ([%eval] from Lichess's server-side analysis). The remaining ~92% have -32768 (no annotation) for all plies. Games with evals can be filtered:

# Find games with eval annotations
df.filter(pl.col("eval").list.eval(pl.element() != -32768).list.any())

Player hashing

Usernames are hashed to uint64 via Polars' xxHash64 (pl.Series.hash(), Polars 1.39.3). Same player always maps to the same hash, enabling player-level grouping and analysis without republishing usernames. The site column (game URL) can be used to look up original usernames on Lichess.

Generation

Extracted from Lichess database dumps (CC0) using the PAWN Rust chess engine for tokenization. See scripts/extract_lichess_parquet.py in the PAWN repository.

Pipeline: zstd-compressed PGN → Rust enriched parser (single-pass extraction of moves, [%clk], [%eval], headers) → Polars DataFrame → zstd-compressed Parquet.

License

CC BY 4.0. Derived from the Lichess database, which is released under Creative Commons CC0.