The official code for the paper: ChessGPT - Bridging Policy Learning and Language Modeling.
We welcome any contribution, especially on chess related dataset towards the development of the next-generation model, ChessGPT-V2. For related matters, please contact [email protected].
We open source our three models ChessGPT-Base, ChessGPT-Chat and ChessCLIP, and our Chess dataset.
- Create a Python virtual environment with the method of your choice. The locked dependencies are generated for Python 3.8.10, but python beyond 3.8 like python 3.9/3.10 are also available.
- Install dependencies by running
pip install -r requirements/dev.txt
. - Setup environment variables with the following commands.
export PYTHONPATH=$PWD:$PWD/third_party/chessclip/src
We offer a ChessCLIP visualization demo to show its capability.
We adopt the code from open_clip-v2.9.3 for our training code of ChessCLIP. To reproduce our training, here are two procedures:
Run tfds build for pathtochessmastery
, pgnlib
, gameknot
and lichess_studies
. Note that the processing of pgnlib needs fasttext's model, you can download it from their official website and modify the path. These 4 sources are free dataset and you can also add the source of megabase
and chess_publishing
if you buy them.
tfds build --imports chess_ai.datasets.tfds --overwrite pathtochessmastery --manual_dir ./chessclip_data/annotated_pgn \
--register_checksums '--beam_pipeline_options=runner=DirectRunner,direct_num_workers=8,direct_running_mode=multi_processing'
After tfds building for all 4 sources, run the following code to train ChessCLIP:
cd chessclip/open_clip/src
torchrun --nproc_per_node 8 -m training.main_chess --model chessclip-quickgelu
After downloading data from chessgpt_data, run the following code for conducting tokenization on all datasets,
cd chessgpt/data
python3 interleave_dataset.py --tokenizer_path ./tokenizer_path --data_path ./data_path --save_path ./save_path --max_seq_length 1024
After preparing the tokenized dataset, modify the hyperparameters and run base training:
cd chessgpt/train/clm_training
sh chess_ai/train/clm_traning/finetune_pp_peft_trainer.sh
After running the base training, we can conduct further instruction-tuning based on instruction data or conversation data.
- For dataset preparation, merge all data sources to one jsonl file.
- After the dataset preparation, run chess_ai/train/sft_traning/train.sh.
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 torchrun --nproc_per_node=8 --master_port=20001 ./train.py \
--model_name_or_path your_chess_base_model_path \
--data_path the_aggregated_one_jsonl_file_path \
--bf16 True \
--output_dir output_file \
--num_train_epochs 3 \
--per_device_train_batch_size 4 \
--per_device_eval_batch_size 4 \
--gradient_accumulation_steps 8 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 2000 \
--save_total_limit 10 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--fsdp "full_shard auto_wrap" \
--fsdp_transformer_layer_cls_to_wrap 'GPTNeoXLayer' \
--tf32 True \
--model_max_length 1024 \
--gradient_checkpointing True \
Refer to ./eval
for evaluation dataset and code of ChessCLIP and ChessGPT.
The code of ChessGPT/CLIP are released under the Apache License, Version 2.0.
If you find ChessGPT useful, please cite it in your publications.
@article{feng2023chessgpt,
title={ChessGPT: Bridging Policy Learning and Language Modeling},
author={Feng, Xidong and Luo, Yicheng and Wang, Ziyan and Tang, Hongrui and Yang, Mengyue and Shao, Kun and Mguni, David and Du, Yali and Wang, Jun},
journal={arXiv preprint arXiv:2306.09200},
year={2023}
}