A collection of chess engines that play like humans, from ELO 1100 to 1900.

In this repo is our 9 final maia models saved as Leela Chess neural networks, and the code to create more and reproduce our results.

Chess Engine

The models (.pb.gz files) work like any other Leela weights file. So to use them download or compile lc0. If the version of lc0 does not support the weights we have the exact version here to compile.

When using the model in UCI mode add nodes 1 when querying as that disables the search.


As part of our analysis all the game on Lichess with stockfish analysis were processed into csv files. These can be found here


Move Prediction

To create your own maia from a set of chess games in the PGN format:

  1. Setup your environment
    1. (optional) Install the conda environment, maia_env.yml
    2. Make sure all the required packages are installed from requirements.txt
  2. Convert the PGN into the training format
    1. Add the pgn-extract tool to your path
    2. Add the trainingdata-tool to your path
    3. Run move_prediction/ PGN_FILE_PATH OUTPUT_PATH
    4. Wait a bit as the processing is both IO and CPU intense
    5. The script will create a training and validation set, if you wish to train on the whole set copy the files from OUTPUT_PATH/validation to OUTPUT_PATH/training
  3. Edit move_prediction/maia_config.yml
    1. Add OUTPUT_PATH/training/*/* to input_train
    2. Add OUTPUT_PATH/validation/*/* to input_test
    3. (optional) If you have multiple GPUS change the gpu filed to the one you are using
    4. (optional) You can also change all the other training parameters here, like the number of layers
  4. Run the training script move_prediction/ PATH_TO_CONFIG
  5. (optional) You can use tensorboard to watch the training progress, the logs are in runs/CONFIG_BASENAME/
  6. Once complete the final model will be in models/CONFIG_BASENAME/ directory. It will be the one with the largest number


To train the models we present in the paper you need to download the raw files from Lichess then cut them into the training sets and process them into the training data format. This is a similar format to the general training instructions just with our specified data, so you will need to have ``trainingdata-toolandpgn-extract` on your PATH.

Also note that running the scripts manually line by line might be necessary as they do not have any flow control logic. And that move_prediction/ is where the main shuffling and games selection logic is.

  1. Download the games from Lichess between January 2017 and November 2019 to data/lichess_raw
  2. Run move_prediction/
  3. Run move_prediction/
  4. Edit move_prediction/maia_config.yml and add the elo you want to train:
    1. input_test : ../data/elo_ranges/${elo}/test
    2. outputtrain : ../data/elo_ranges/${elo}/train
  5. Run the training script move_prediction/ PATH_TO_CONFIG

We also include some other (but not all) config files that we tested. Although, we still recommend using the final config move_prediction/maia_config.yml.

If you wish to generate the testing set we used you can download the December 2019 data and run move_prediction/ The data is also avaible for download as a CSV here

Blunder Prediction

To train the blunder prediction models follow these instructions:

  1. Setup your environment
    1. (optional) Install the conda environment, maia_env.yml
  2. Make sure all the required packages are installed from requirements.txt
  3. Run blunder_prediction/
    1. You will probably need to update the paths, and may want to change the targets or use a for loop
  4. Run blunder_prediction/ on all the csv files
  5. Select a config from blunder_prediction/configs and update the paths
  6. Run `blunder_prediction/ CONFIG_PATH


  title={Aligning Superhuman AI with Human Behavior: Chess as a Model System},
  author={McIlroy-Young, Reid and  Sen, Siddhartha and Kleinberg, Jon and Anderson, Ashton},
  booktitle={Proceedings of the 25th ACM SIGKDD international conference on Knowledge discovery and data mining}