Ailed-AI/lichess-chess-1000-1800

Lichess Chess Games (1000-1800 ELO) — Pre-tokenized

Pre-tokenized chess game data from Lichess, filtered to 1000-1800 ELO rated games. Ready for training self-supervised game models.

Format

Files are in .somabin binary format — a self-describing, indexed, memory-mappable format created by ailed-soulsteal.

Each game contains:

• token_ids (u16) — tokenized UCI moves with BOS/EOS markers
• turn_ids (u8) — 0=white, 1=black per token
• category_ids (u8) — piece type per move (pawn=0, knight=1, bishop=2, rook=3, queen=4, king=5)
• outcome (u8) — 0=win, 1=draw, 2=loss

Files

File	Games	Source
lichess-2013-01.somabin	89,395	Lichess Jan 2013
vocab.json	4,214 tokens	All possible chess UCI moves

Quick Start

Python (PyTorch)

from somabin import SomaTrainingDataset
from torch.utils.data import DataLoader

# Load dataset — instant via mmap, no parsing needed
train_ds, val_ds = SomaTrainingDataset.split("lichess-2013-01.somabin", val_ratio=0.1)

# Standard PyTorch DataLoader
loader = DataLoader(train_ds, batch_size=128, shuffle=True)
for batch in loader:
    # batch['input_tokens'], batch['target_tokens'], batch['turns'], etc.
    pass

Get the Python reader: somabin.py

Inspect with soulsteal CLI

cargo install ailed-soulsteal

soulsteal info lichess-2013-01.somabin
soulsteal stats lichess-2013-01.somabin
soulsteal dump lichess-2013-01.somabin --vocab vocab.json --game 0 --count 5

Create your own

Process any Lichess PGN dump into this format:

soulsteal tokenize lichess_db_standard_rated_2016-02.pgn.zst \
  -o train.somabin \
  --vocab vocab.json \
  --elo 1000:1800 \
  --max-games 1000000

60,000+ games/second. 1M games in under 30 seconds.

Tokenization

Vocabulary: 4,214 tokens (6 special + 4,208 UCI moves)

ID	Token
0	PAD
1	BOS (start of game)
2	EOS (end of game)
3	WIN
4	DRAW
5	LOSS
6+	UCI moves (e2e4, d2d4, g1f3, ...)

Source

Raw data from the Lichess Open Database under Creative Commons CC0.

Tokenized with ailed-soulsteal v26.3.0.

Part of the AILED Ecosystem

This dataset is designed for training AILED-Soma — a self-supervised game predictor that serves as the cartridge for the AILED cognitive architecture.