AngeloWolff/snail-chess-data

Snail Chess Data

Chess positions from self-play games of Snail Chess, the engine behind Snail Chess GUI. Each position has the engine's search score and the game result. This is the data that trained Shell, Snail's evaluation network.

Positions	115.4 million
Games	1,373,615
Download	1.1 GB (3.8 GB unpacked)
License	CC BY 4.0

Quick start

pip install -U huggingface_hub numpy
hf download AngeloWolff/snail-chess-data --repo-type dataset --local-dir snail-chess-data

Add --include "data/gen6/*" --include "*.py" to download a single generation.

import snail_data

records = snail_data.read("snail-chess-data/data/gen6/part00.bin.xz")

print(snail_data.to_fen(records[0]))  # r1bqkb1r/ppp1ppp1/2np1n1p/8/3P4/3QPN2/PPP1BPPP/RNB1K2R b - - 0 1
print(records[0]["score"])            # search score in centipawns, White's view
print(records[0]["result"])           # 0 = black won, 1 = draw, 2 = white won

batch = records[:4096]
x = snail_data.planes(batch)          # one-hot piece planes, shape (4096, 768)
y = snail_data.target(batch)          # sigmoid(score / 120), the training target

Generations

Snail learned by playing itself. Its old hand-written eval played the first games, a Shell net was trained on them, and each new net played the next generation.

Generation	Positions	Played by	Nodes per move
gen0	10,962,132	Snail Chess 1.15 (hand-written eval)	5,000
gen0b	11,819,836	Snail Chess 1.15 (hand-written eval)	5,000
gen1	9,154,571	gen0b net	5,000
gen2	6,978,633	gen1 net	8,000
gen3	7,500,400	gen2 net	8,000
gen4	8,082,566	Snail Chess 1.16 (gen3 net)	20,000
gen5	11,227,776	gen4 net	50,000
gen6	11,374,723	gen5 net	100,000
gen7	12,572,074	Snail Chess 1.17 (gen6 net)	200,000
gen8	25,688,615	Snail Chess 1.18 (gen7 net, 64 lanes)	10,000

• Games start from Snail's opening books plus one or two random moves. Each part's .fen file lists its openings.
• gen7 was played on three PCs at once. Each gen7 part joins a few whole runs, so its records are still whole games, and its .fen lists the openings in the same order.
• gen8 is the first generation played by a 64-lane net, so the node count starts low again. It started from 320,000 openings, twice the usual number, and none of them is an opening of the gen7 run held out for validation.
• gen0 started from random moves instead. Its positions were too unusual, so later nets didn't use it.
• Only quiet positions were saved: not in check, best move not a capture or promotion, no mate score.

Opening books

The four books in books/ were generated by Snail Chess 1.15:

Book	Positions	Depth	Snail uses it for
master.txt	24,794	20 plies	full strength
neutral.txt	17,812	16 plies	the Balanced style
gambit.txt	18,777	16 plies	the Aggressive style, lines that give up material
passive.txt	17,678	16 plies	the Passive style, quiet lines with few captures

Each line is a sequence of moves in UCI notation, a colon, then the next moves with their weights. A line starting with : gives the moves from the starting position:

d2d4 : g8f6 100 d7d5 67 e7e6 42

openings.py makes openings the way the games here started: a short weighted walk through a random book, plus random moves. It needs pip install chess.

python openings.py openings.txt --count 160000 --random-plies 2

How Shell was trained

Net	Trained on	Released in
gen3	gen0b–gen3	Snail Chess 1.16
gen4	gen0b–gen4	-
gen5	gen2–gen5	-
gen6	gen3–gen6	Snail Chess 1.17
gen7	gen4–gen7	-
gen7, 64 lanes	gen4–gen7	Snail Chess 1.18
gen8, 64 lanes	gen4–gen8	-

• Target: sigmoid(score / s), from search scores only, with s fitted to the data (120 from gen6 on).
• Model: Shell, a tiny gated network (a split-input clipped ReGLU) added to a small base eval. It had 32 lanes up to gen7 and has 64 since Snail Chess 1.18.
• Training: mean squared error, Adam, 60 epochs (120 for gen8). From gen7 on, the last 5 epochs train through the int8 rounding (quantization-aware training), and one gen7 run is held out for validation.
• Skipped: positions with flag bit 1 set (about 14%), which the engine scales separately.

Record format

Each data/genN/partNN.bin.xz decompresses to an array of 32-byte little-endian records:

Bytes	Field	Type	Meaning
0–7	occupied	u64	bit i set = square i has a piece (a1 = 0, h8 = 63)
8–23	pieces	32 × 4 bits	piece code of each occupied square, lowest square first, low nibble first
24–25	score	i16	search score, centipawns, White's view
26–27	base	i16	base eval, White's view
28–29	eval	i16	full static eval, White's view
30	flags	u8	bit 0: black to move; bit 1: endgame scaling
31	result	u8	0 = black won, 1 = draw, 2 = white won

	Pawn	Knight	Bishop	Rook	Queen	King
White	2	4	6	8	10	12
Black	3	5	7	9	11	13

Castling rights and en passant squares are not stored.

License

CC BY 4.0. Snail Chess itself is GPL-3.0.