InterwebAlchemy/pgn-lichess-puzzle-dataset

Chess Puzzles with PGN Context

Tactical chess puzzles drawn from the Lichess Open Puzzle Database, augmented with full PGN game context reconstructed from the source Lichess games.

Each record presents a middlegame position as a PGN move sequence — the same format used to train PGN language models like kn1ght — together with the engine-validated best move as the label.

Dataset Summary

• 5,000 puzzles with reconstructed PGN context
• Rating range: 1200–1900 (mean: 1540)
• Themes: middlegame, short, advantage, crushing, long, mate, fork, kingsideAttack (middlegame only; opening/endgame excluded)
• Splits: 80% train / 10% validation / 10% test

Schema

Column	Type	Description
puzzle_id	string	Lichess puzzle ID
game_id	string	Lichess game ID (source of PGN context)
rating	int32	Puzzle difficulty (Lichess Glicko-2 rating)
themes	list[string]	Tactical theme tags (e.g. fork, pin, skewer)
pgn_context	string	PGN move text up to (not including) the puzzle move
fen	string	Board position at start of puzzle in FEN notation
best_move_uci	string	Correct first move in UCI notation
best_move_san	string	Correct first move in SAN notation

Usage

from datasets import load_dataset

ds = load_dataset("InterwebAlchemy/chess-puzzles-pgn")

# Each example:
# {'pgn_context': '1.e4 e5 2.Nf3 Nc6 ... 18.Rxd4',
#   'best_move_san': 'Nf6+',
#   'rating': 1487,
#   'themes': ['fork', 'middlegame']}

How PGN context is reconstructed

For each Lichess puzzle:

1. The source game PGN is fetched from lichess.org/api/game/<id>
2. The game is replayed move by move until the board FEN matches the puzzle FEN
3. The move-text up to that point is stored as pgn_context
4. The first solution move (UCI → SAN) is stored as best_move_san

Puzzles where the FEN could not be located in the source game are discarded.

Intended use

Evaluating and fine-tuning PGN language models on tactical positions. The pgn_context field can be fed directly to any model that generates chess moves as PGN continuations.

Licensing

This dataset is derived from the Lichess Open Database, which is released under CC0 1.0 (Public Domain). This derived dataset is also released under CC0 1.0.