Skip to content

Latest commit

 

History

History
97 lines (85 loc) · 4.65 KB

README.md

File metadata and controls

97 lines (85 loc) · 4.65 KB
Raw

ChessGPT - Bridging Policy Learning and Language Modeling

The official code for the paper: ChessGPT - Bridging Policy Learning and Language Modeling.

Contribution

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].

Model and Dataset

We open source our three models ChessGPT-Base, ChessGPT-Chat and ChessCLIP, and our Chess dataset.

Installation

  1. 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.
  2. Install dependencies by running pip install -r requirements/dev.txt.
  3. Setup environment variables with the following commands.
export PYTHONPATH=$PWD:$PWD/third_party/chessclip/src

Visualization

We offer a ChessCLIP visualization demo to show its capability.

Training

ChessCLIP

We adopt the code from open_clip-v2.9.3 for our training code of ChessCLIP. To reproduce our training, here are two procedures:

Generate dataset using tfds

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'

ChessCLIP training

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

ChessGPT

Base-training

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

Instruction-tuning

After running the base training, we can conduct further instruction-tuning based on instruction data or conversation data.

  1. For dataset preparation, merge all data sources to one jsonl file.
  2. 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 \

Evaluation

Refer to ./eval for evaluation dataset and code of ChessCLIP and ChessGPT.

License

The code of ChessGPT/CLIP are released under the Apache License, Version 2.0.

Citation

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}
}