The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

Condensed Lichess Database

This dataset is a condensed version of the Lichess database. It only includes games for which Stockfish evaluations were available. Currently, the dataset contains the entire year 2023, which consists of >100M games and >2B positions. Games are stored in a format that is much faster to process than the original PGN data.

Requirements:

    pip install zstandard python-chess datasets

Quick Guide

In the following, I explain the data format and how to use the dataset. At the end, you find a complete example script.

1. Loading The Dataset

You can stream the data without storing it locally (~100 GB currently). The dataset requires trust_remote_code=True to execute the custom data loading script, which is necessary to decompress the files. See HuggingFace's documentation if you're unsure.

# Load dataset.
dataset = load_dataset(path="mauricett/lichess_sf",
                       split="train",
                       streaming=True,
                       trust_remote_code=True)

2. Data Format

The following definitions are important to understand. Please reread this section slowly and correctly when you have to decide how to draw FENs, moves and scores from the dataset. Let's draw a single sample and discuss it.

example = next(iter(dataset))

A single sample from the dataset contains one complete chess game as a dictionary. The dictionary keys are as follows:

  1. example['fens'] --- A list of FENs in a slightly stripped format, missing the halfmove clock and fullmove number (see definitions on wiki). The starting positions have been excluded (no player made a move yet).
  2. example['moves'] --- A list of moves in UCI format. example['moves'][42] is the move that led to position example['fens'][42], etc.
  3. example['scores'] --- A list of Stockfish evaluations (in centipawns) and the game's terminal outcome condition if one exists. Evaluations are from the perspective of the player who is next to move. If example['fens'][42] is black's turn, example['scores'][42] will be from black's perspective. If the game ended with a terminal condition, the last element of the list is a string 'C' (checkmate), 'S' (stalemate) or 'I' (insufficient material). Games with other outcome conditions have been excluded.
  4. example['WhiteElo'], example['BlackElo'] --- Player's Elos.

3. Define Functions for Preprocessing

To use the data, you will require to define your own functions for transforming the data into your desired format. For this guide, let's define a few mock functions so I can show you how to use them.

# A mock tokenizer and functions for demonstration.
class Tokenizer:
    def __init__(self):
        pass
    def __call__(self, example):
        return example


# Transform Stockfish score and terminal outcomes.
def score_fn(score):
    return score

def preprocess(example, tokenizer, score_fn):
    # Get number of moves made in the game...
    max_ply = len(example['moves'])
    # ...and pick a position at random.
    random_position = random.randint(0, max_ply-2)

    # Get the FEN of our random choice.
    fen = example['fens'][random_position]

    # To get the move that leads to the *next* FEN, we have to add
    # +1 to the index. Same with the score, which is the evaluation
    # of that move. Please read the section about the data format clearly!
    move = example['moves'][random_position + 1]
    score = example['scores'][random_position + 1]

    # Transform data into the format of your choice.
    example['fens'] = tokenizer(fen)
    example['moves'] = tokenizer(move)
    example['scores'] = score_fn(score)
    return example

tokenizer = Tokenizer()

4. Shuffle And Preprocess

Use dataset.shuffle() to properly shuffle the dataset. Use dataset.map() to apply our preprocessors. This will process individual samples in parallel if you're using multiprocessing (e.g. with PyTorch dataloader).

# Shuffle and apply your own preprocessing.
dataset = dataset.shuffle(seed=42)
dataset = dataset.map(preprocess, fn_kwargs={'tokenizer': tokenizer,
                                             'score_fn': score_fn})



COMPLETE EXAMPLE

You can try pasting this into Colab and it should work fine. Have fun!

import random
from datasets import load_dataset
from torch.utils.data import DataLoader


# A mock tokenizer and functions for demonstration.
class Tokenizer:
    def __init__(self):
        pass
    def __call__(self, example):
        return example


def score_fn(score):
    # Transform Stockfish score and terminal outcomes.
    return score

def preprocess(example, tokenizer, score_fn):
    # Get number of moves made in the game...
    max_ply = len(example['moves'])
    # ...and pick a position at random.
    random_position = random.randint(0, max_ply-2)

    # Get the FEN of our random choice.
    fen = example['fens'][random_position]

    # To get the move that leads to the *next* FEN, we have to add
    # +1 to the index. Same with the score, which is the evaluation
    # of that move. Please read the section about the data format clearly!
    move = example['moves'][random_position + 1]
    score = example['scores'][random_position + 1]

    # Transform data into the format of your choice.
    example['fens'] = tokenizer(fen)
    example['moves'] = tokenizer(move)
    example['scores'] = score_fn(score)
    return example


tokenizer = Tokenizer()

# Load dataset.
dataset = load_dataset(path="mauricett/lichess_sf",
                       split="train",
                       streaming=True,
                       trust_remote_code=True)

                    
# Shuffle and apply your own preprocessing.
dataset = dataset.shuffle(seed=42)
dataset = dataset.map(preprocess, fn_kwargs={'tokenizer': tokenizer,
                                             'score_fn': score_fn})

# PyTorch dataloader
dataloader = DataLoader(dataset, batch_size=1, num_workers=1)


for batch in dataloader:
    # do stuff
    print(batch)
    break

# Batch now looks like:
# {'WhiteElo': tensor([1361]), 'BlackElo': tensor([1412]), 'fens': ['3R4/5ppk/p1b2rqp/1p6/8/5P1P/1PQ3P1/7K w - -'], 'moves': ['g8h7'], 'scores': ['-535']}
# Much better!
Downloads last month
254