neuralcrystal/checkers-positions

Checkers Positions

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

This dataset holds 916,000 American checkers (English draughts) positions, each with the move to play and every legal move scored by a strong checkers engine. It is the training data for the Checkers Crystal, a block of glass that plays checkers with light.

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, always as the side to move: its own pieces are solid discs, the opponent's are rings:

Use it

pip install datasets

from datasets import load_dataset

# the main set, labeled by the Stermere/Marcher engine
stermere = load_dataset("neuralcrystal/checkers-positions", "stermere")
row = stermere["train"][0]
print(row["pos"], row["move"])

# the hold-out the crystal's published score comes from: 5,675 test positions
holdout = stermere["test"].filter(lambda r: r["source"] == "holdout")

To read positions, list legal moves and draw the board the way the crystal sees it, use the neuralcrystal package (pip install neuralcrystal): neuralcrystal.checkers.parse(pos), .moves(...) and .frames(...).

Configs

Config	Train	Test	What it is
stermere	803,255	54,291	Labeled by the Stermere/Marcher engine (a neural-network evaluation searched for 0.1 s a move, plus its endgame tablebase). source: "games" (openings, weak and strong play, jump chains), "endgames" (12 pieces down to 2, evenly spread), "tactics" (positions where one move walks into a shot and another doesn't). The test split adds "holdout": the 5,675 positions the crystal is scored on.
custom	54,858	3,600	The same kinds of positions labeled by a simpler engine of our own (a plain search with a hand-written evaluation), at search depths 9 and 11.

The two configs use different teachers, so their scores aren't on the same scale. Don't mix them in one loss without care.

Fields

Field	Type	What it is
pos	string	The position: 32 characters, one per playable square in standard draughts order (b/w men, B/W kings, . empty), then the side to move (b/w), the square of a piece that must keep jumping (or -), and the moves since the last capture.
move	string	The move to learn, in draughts notation: 11-15 is a step, 11x18 a jump. A jump chain is a series of single jumps by the same side.
score	int32	The teacher's score for that move, from the side to move's point of view. A man is worth 100; a won position scores up to 20,000.
cands	list of {move, score}	Every legal move with its score. In stermere each also has raw, the engine's own score before conversion to this scale.
tag	string	The kind of position: step, capture, chain (in the middle of a jump chain), promo (a man about to crown), king, late (late game).
origin	string	How the position was generated (e.g. play, endgame, chainwalk, random, tactic:…).
game	int64	An id for the game or batch the position came from.
count	int32	How many times this row appeared in the crystal's training data (1 for almost all rows).
teacher	string	stermere, custom-d9 or custom-d11.
depth	int16	custom only: the search depth used.
source, split	string	Which part of the config the row came from (see Configs); train or test.

How it was built

1. Positions: generated by our own pipeline: openings, games at different strengths, walked jump chains, endgames from 12 pieces down to 2, and tactical positions.
2. Labels: every legal move in every position scored by the teacher engine.
3. Duplicates: each position and move appears once per split; count records how many copies there were.
4. No leaks: no position appears in both train and test of the same config.

Credits

The stermere labels are the output of Stermere/Marcher by Collin Kees (MIT license, © 2023–2026 Collin Kees). The custom labels come from our own engine.

License

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

© 2026 TextJam, Inc.