Chess Transformers

Teaching transformers to play chess

Chess Transformers is a library for training transformer models to play chess by learning from human games.

Contents

Install

Models

Datasets

Play

Train Models

Contribute

Install

To install Chess Transformers, clone this repository and install as a Python package locally.

gh repo clone sgrvinod/chess-transformers
cd chess-transformers
pip install .

If you are planning to develop or contribute or make changes to the codebase, install the package in editable mode, using the

flag.

pip install -e .

OPTIONAL — If you want to train or evaluate a model, you may need to set some of the following environment variables on your computer:

• Set

  CT_DATA_FOLDER
  to the folder on your computer where you have the training data. You do not need to set this if you do not plan to train any models.

• Set

  CT_STOCKFISH_PATH
  to the executable of the Stockfish 16 chess engine. You do not need to set this if you do not plan to have a model play against this chess engine.

• Set

  CT_FAIRY_STOCKFISH_PATH
  to the executable of the Fairy Stockfish chess engine. You do not need to set this if you do not plan to have a model play against this chess engine.

Models

There are currently four models available for use in Chess Transformers.

All models are evaluated against the Fairy Stockfish chess engine at increasing strength levels 1 to 6, as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

At each strength level of the chess engine, $n=1000$ games are played by the model, i.e. $500$ games each with black and white pieces.

Win ratios and the difference between the Elo rating of the model and the chess engine are calculated from these games' outcomes.

Detailed evaluation results for each model are provided below.

CT-E-20

Configuration File | Checkpoint | TensorBoard Logs

This is the encoder from the original transformer model in Vaswani et al. (2017) trained on the LE22ct dataset. A classification head at the

token predicts the best half-move to be made (in UCI notation).

This is essentially a sequence (or image) classification task, where the sequence is the current state of the board, and the classes are the various moves that can be made on a chessboard in UCI notation.

CT-E-20 contains about 20 million parameters.

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-E-20 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

These evaluation games can be viewed here.

CT-EFT-20

Configuration File | Checkpoint | TensorBoard Logs

This is the encoder from the original transformer model in Vaswani et al. (2017) trained on the LE22ct dataset. Two classification heads operate upon the encoder outputs at all chessboard squares to predict the best candidates for the source (From) and destination (To) squares that correspond to the best half-move to be made.

This is essentially a sequence (or image) labeling task, where the sequence is the current state of the chessboard, and each square competes to be labeled as the From or To square.

CT-E-20 contains about 20 million parameters.

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-EFT-20 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

These evaluation games can be viewed here.

CT-ED-45

Configuration File | Checkpoint | TensorBoard Logs

This is the original transformer model (encoder and decoder) in Vaswani et al. (2017) trained on the LE22ct dataset. A classification head after the last decoder layer predicts a sequence of half-moves, starting with the best half-move to be made next, followed by the likely course of the game an arbitrary number of half-moves into the future.

This is essentially a sequence-to-sequence (or image-to-sequence) task, where the input sequence is the current state of the board, and the output sequence is a string of half-moves that will likely occur on the board from that point onwards. Potentially, strategies applied to such tasks, such as beam search for decoding the best possible sequence of half-moves, can also be applied to this model. Training the model to predict not only the best half-move to make on the board right now, but also the sequence of half-moves that follow, can be viewed as a type of multitask training.

We are ultimately only interested in the very first half-move. Nevertheless, the full sequence of half-moves might help explain the model's decision for this important first half-move.

CT-ED-45 contains about 45 million parameters.

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-ED-45 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

CT-EFT-85

Configuration File | Checkpoint | TensorBoard Logs

This is a larger version of the encoder from the original transformer model in Vaswani et al. (2017) trained on the LE22c dataset. Its size is analogous to BERT_BASE in Devlin et al. (2018). Two classification heads operate upon the encoder outputs at all chessboard squares to predict the best candidates for the source (From) and destination (To) squares that correspond to the best half-move to be made.

This is essentially a sequence (or image) labeling task, where the sequence is the current state of the chessboard, and each square competes to be labeled as the From or To square.

CT-E-85 contains about 85 million parameters.

You do not need to download the model checkpoint manually. It will be downloaded automatically if required.

Model Strength

CT-EFT-85 was evaluated against the Fairy Stockfish chess engine at various strength levels as predefined for use in the popular Stockfish chess bots on Lichess. The engine is run on an AMD Ryzen 7 3800X 8-Core Processor, with 8 CPU threads, and a hash table size of 8 GB. All other engine parameters are at their default values.

These evaluation games can be viewed here.

Datasets

There are currently four training datasets available in Chess Transformers.

These datasets are sourced from groups of PGN files containing real games played by humans. There are currently three PGN filesets:

• LE22 consists of games from the Lichess Elite Database put together by nikonoel, a collection of all standard chess games played on Lichess.org by players with a Lichess Elo rating of 2400+ against players with a Lichess Elo rating of 2200+ up to December 2021, and players rated 2500+ against players rated 2300+ from December 2021 up to December 2022

• ML23 consists of Master-level games downloaded from PGN mentor, TWIC, and Caissabase in December 2023

The lowercase letters at the end of every dataset denote specific filters that were applied to games from the corresponding PGN filesets:

• "c for games that ended in a checkmate
• "t" for games that used a specific time control
• "d" for games that ended decisively

ML23c

This consists of Master-level games downloaded from PGN mentor, TWIC, and Caissabase in December 2023.

On this data (11,081,724 games), we apply the following filters to keep only those games that:

• are unique (5,213,634 games)
• and ended in a checkmate (250,694 games)

These 250,694 games consist of a total 10,797,366 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

• ML23c.h5
  , an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
  • board_position
    , the chessboard layout, or positions of pieces on the board
  • turn
    , the color of the pieces of the player to play
  • white_kingside_castling_rights
    , whether white can castle kingside
  • white_queenside_castling_rights
    , whether white can castle queenside
  • black_kingside_castling_rights
    , whether black can castle kingside
  • black_queenside_castling_rights
    , whether black can castle queenside
  • moves
    , 10 half-moves into the future made by both players
  • length
    , the number of half-moves in the sequence, as this will be less than 10 at the end of the game

LE22ct

This consists of games from the Lichess Elite Database put together by nikonoel, a collection of all standard chess games played on Lichess.org by players with a Lichess Elo rating of 2400+ against players with a Lichess Elo rating of 2200+ up to December 2021, and players rated 2500+ against players rated 2300+ from December 2021 up to December 2022.

On this data (20,241,368 games), we apply the following filters to keep only those games that:

• used a time control of at least 5 minutes (2,073,780 games)
• and ended in a checkmate (274,794 games)

These 274,794 games consist of a total 13,287,522 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

• LE22ct.h5
  , an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
  • board_position
    , the chessboard layout, or positions of pieces on the board
  • turn
    , the color of the pieces of the player to play
  • white_kingside_castling_rights
    , whether white can castle kingside
  • white_queenside_castling_rights
    , whether white can castle queenside
  • black_kingside_castling_rights
    , whether black can castle kingside
  • black_queenside_castling_rights
    , whether black can castle queenside
  • moves
    , 10 half-moves into the future made by both players
  • length
    , the number of half-moves in the sequence, as this will be less than 10 at the end of the game

LE22c

This is an extended version of LE22ct, and consists of games from the Lichess Elite Database put together by nikonoel, a collection of all standard chess games played on Lichess.org by players with a Lichess Elo rating of 2400+ against players with a Lichess Elo rating of 2200+ up to December 2021, and players rated 2500+ against players rated 2300+ from December 2021 up to December 2022.

On this data (20,241,368 games), we apply the following filters to keep only those games that:

• ended in a checkmate (2,751,394 games)

These 2,751,394 games consist of a total 127,684,720 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

• LE22c.h5
  , an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
  • board_position
    , the board layout or positions of pieces on the board
  • turn
    , the color of the pieces of the player to play
  • white_kingside_castling_rights
    , whether white can castle kingside
  • white_queenside_castling_rights
    , whether white can castle queenside
  • black_kingside_castling_rights
    , whether black can castle kingside
  • black_queenside_castling_rights
    , whether black can castle queenside
  • moves
    , 10 half-moves into the future made by both players
  • length
    , the number of half-moves in the sequence, as this will be less than 10 at the end of the game

ML23d

This consists of Master-level games downloaded from PGN mentor, TWIC, and Caissabase in December 2023.

On this data (11,081,724 games), we apply the following filters to keep only those games that:

• are unique (5,213,634 games)
• and are decisive, i.e. a player won (3,739,604 games)

These 3,739,604 games consist of a total 144,625,397 half-moves made by the winners of the games, which alone constitute the dataset. For each such half-move, the chessboard, turn (white or black), and castling rights of both players before the move are calculated, as well as the sequence of half-moves beginning with this half-move up to 10 half-moves into the future. Draw potential is not calculated.

Download here. The data is zipped and will need to be extracted.

It consists of the following files:

• ML23d.h5
  , an HDF5 file containing two tables, one with the raw data and the other encoded with indices (that will be used in the transformer model), containing the following fields:
  • board_position
    , the chessboard layout, or positions of pieces on the board
  • turn
    , the color of the pieces of the player to play
  • white_kingside_castling_rights
    , whether white can castle kingside
  • white_queenside_castling_rights
    , whether white can castle queenside
  • black_kingside_castling_rights
    , whether black can castle kingside
  • black_queenside_castling_rights
    , whether black can castle queenside
  • moves
    , 10 half-moves into the future made by both players
  • length
    , the number of half-moves in the sequence, as this will be less than 10 at the end of the game

Play

After installing Chess Transformers, you can play games against an available model or have a model play against a chess engine.

You v. Model

You could either play in a Jupyter notebook (recommended for better UI) or in a Python shell.

You could also just make a copy of

and play in that notebook.

Model v. Engine

The process is the same as above, except you must use a different set of functions:

See

for an example.

Time Control

If you're using a Unix-type operating system — basically, not Windows — you can also set a time control for your games. Currently, only Fischer time control is available.

Pass this

to the functions above instead of

clock=None

.

Train Models

You're welcome to try to train your own models, but if you wish to contribute trained models, please discuss first.

Dataset

You can skip this step if you wish to use one of the existing datasets.

• Collect PGN files containing games you wish to use for training the model.

• Create a bash script for parsing these PGN files into a collection of FENs and moves using pgn-extract, like in

  LE22ct.sh
  , and execute it in the folder with the PGN files.

• Create a configuration file for the dataset, like in

  LE22ct.py
  .

• Run

  prep.py
  like
  python prep.py [config_name]
  , or do it in your own Python script.

Data files will be created in

.

Training

• Create a configuration file for the model, like in

  CT-E-20.py
  .

• Run

  train.py
  like
  python train.py [config_name]
  , or do it in your own Python script.

• Monitor training with TensorBoard with

  tensorboard --logdir $CT_LOGS_DIR
  .

• Average checkpoints saved for averaging to produce the final checkpoint. Run

  average_checkpoints.py
  like
  python average_checkpoints.py [config_name]
  , or do it in your own Python script.

Evaluation

Run

like

python evaluate.py [config_name]

, or do it in your own Python notebook/script.

Contribute

Contributions — and any discussion thereof — are welcome. As you may have noticed, Chess Transformers is in initial development and the public API is not to be considered stable.

If you are planning to contribute bug-fixes, please go ahead and do so. If you are planning to contribute in a way that extends Chess Transformers, or adds any new features, data, or models, please create a discussion thread to discuss it before you spend any time on it. Otherwise, your PR may be rejected due to lack of consensus or alignment with current goals.

Presently, the following types of contributions may be useful:

• Better, more robust evaluation methods of models.
• Evaluation of existing models against chess engines on different CPUs to study the effect of CPU specifications on engine strength and evaluation.
• New models with:
  • the same transformer architectures but of a larger size, and trained on larger datasets.
  • or different transformer architectures or internal mechanisms.
  • or in general, improved evaluation scores.
• Chess clocks for Windows OS, or for Unix-type OS but for time controls other than Fischer time control.
• Refactoring of code that improves its ease of use.
• Model visualization for explainable AI, such as visualizing positional and move embeddings, or attention patterns.

This list is not exhaustive. Please do not hesitate to discuss your ideas. Thank you!

License

Chess Transformers is licensed under the MIT license.