neuralcrystal/chess-positions

Chess Positions

NeuralCrystal was built by Pete DeLaurentis with Shellcaster, an IDE for building complex projects with agents.

This dataset holds 2.5 million chess positions, each with the best move and a ranked list of candidate moves scored by a chess engine. It is the training data for the Chess Crystal, a block of glass that plays chess with light. Almost every label comes from Stockfish 17.1; the endgame tablebase rows have exact answers from the Syzygy tablebases.

Demo

See the crystal trained on this data play in the optical simulator at NeuralCrystal.com.

What a row looks like

Every row is one position and the move to play there. The crystal sees the position as a picture on a micromirror array; this is the board after white opens with "1. e4":

Use it

pip install datasets

from datasets import load_dataset, interleave_datasets

# one config: the positions where only one threatened piece can be saved
dilemma = load_dataset("neuralcrystal/chess-positions", "dilemma", split="train")
print(dilemma[0]["fen"], dilemma[0]["move"])

# your own mix of configs: half Stockfish's games, a third the crystal's games, the rest dilemmas
teacher = load_dataset("neuralcrystal/chess-positions", "teacher", split="train")
dagger = load_dataset("neuralcrystal/chess-positions", "dagger", split="train")
mix = interleave_datasets([teacher, dagger, dilemma], probabilities=[0.5, 0.3, 0.2])

To draw a position the way the crystal sees it, or to run the crystal on it, use the neuralcrystal package (pip install neuralcrystal): neuralcrystal.chess.frame(fen).

Configs

The configs are split by what the data is for, so you can choose your own mix. The source column breaks each one down further.

Config	Train	Test	What it is
teacher	700,308	20,992	Positions from Stockfish playing itself, starting from common openings with some random moves mixed in (source "games"); plus groups of sibling positions, the best and some worse moves from one position each played out (source "siblings").
dagger	1,023,699	19,831	Positions the crystal itself reached, each labeled by Stockfish: in its own games against Stockfish (source "round-1" to "round-16", one per round of training) and in its own two-move look-ahead ("lookahead-1" to "lookahead-3"), plus won endings searched deep enough to see the mate ("deep-mates"). Training on positions the learner reaches itself teaches it to recover from its own mistakes; this is the DAgger method (Ross, Gordon & Bagnell, 2011).
dilemma	593,326	1,984	Positions where only one of several threatened pieces can be saved, and the position after each choice.
endgame	141,163	3,484	Exact or very deep answers near the end of the game: Syzygy tablebase positions with five pieces or fewer ("tablebase"), won endings searched 20–24 moves deep ("deep-mates"), king and queen or rook against a lone king ("kq-kr"), and positions where one move ends the game in stalemate or a draw ("terminal").

Fields

Field	Type	What it is
fen	string	The position, in FEN: the standard one-line text description of a chess board (pieces, side to move, castling rights, en-passant square, move counters).
move	string	The move to learn: the teacher's best move, in UCI notation (from-square then to-square, e.g. e7e5; a promotion adds the piece, e.g. e7e8q).
cp	int32	The teacher's score for that move in centipawns (hundredths of a pawn), from the side to move's point of view. Mates are scored near ±10,000.
cands	list of {move, cp}	The candidate moves the teacher scored, with their scores.
alts	list of string	Moves scored within a small margin of the best: all acceptable answers.
bad	list of {move, cp}	Moves marked as clear mistakes (filled in only for king and queen or rook endings).
count	int32	How many times this row appeared in the crystal's training data (1 for most rows). Use it as a weight to reproduce the original mix, or ignore it.
teacher	string	Who labeled the row: stockfish-17.1, syzygy (tablebase) or rules (stalemate and draw positions, built by hand).
depth, cands_depth	int16	How deep Stockfish searched for the label and for the candidate scores (empty for tablebase and rules rows).
source	string	Which part of the config the row came from (see Configs).
split	string	train or test.
group, group_cp	int32	teacher config only: the sibling group a position belongs to, and the score of the move that led to it.
wdl, material	int8, string	endgame config only: the exact result for the side to move (1 win, 0 draw, −1 loss) and the pieces on the board (e.g. KRPvKR).

How it was built

1. Positions: generated by our own pipeline. Stockfish played itself from a list of common openings, the crystal played its own games against Stockfish at several strengths, and special positions (dilemmas, endings, stalemates) were constructed.
2. Labels: every non-tablebase position was scored by Stockfish 17.1, the best move at depth 10 and a ladder of candidate moves at depth 8 (deeper for the mating positions). Tablebase rows use Syzygy's exact results.
3. Duplicates: each position and move appears once per split. The count column records how many copies there were (most repeats are popular openings and boards the crystal's look-ahead revisited).
4. No leaks: no position appears in both train and test of the same config.

Credits

The move scores come from Stockfish 17.1, by the Stockfish developers (GPL-3 software, used here only to label positions; this dataset is not endorsed by the Stockfish project), and from the Syzygy endgame tablebases.

License

CC BY 4.0: free to use, share and adapt, including commercially, with credit.

© 2026 TextJam, Inc.