Hete is a performant Deep-CFR network trainer in c++ using mlx.

• Decoder only poker GPT in one file src/models/model.h
• Minimal 2+ player poker engine
• Monte carlo outcome sampler for game rollouts
• CFR training loop
• Fast hand evaluations using OMPEval library

terminology:

• mbb = 1/1000 * big blind
• baseline statistics = difference between outcome of you played vs. how slumbot would have played. more positive is better.
• mbb_per_hand = average mbb winning across 10000 games

winning by 50mbb or more per hand on average is considered a significant win between professionals.

performance of my latest run against slumbot.com:

there's work to do.

first clone hete:

git clone https://github.com/pythonlearner1025/hete.git && 
cd hete  

install dependencies:

1. build OMPEval:

git clone https://github.com/pythonlearner1025/OMPEval.git &&
cd OMPEval &&
make clean &&
make

2. build and install mlx (might have to sudo make install):

git clone https://github.com/ml-explore/mlx.git && 
cd mlx &&
mkdir -p build && cd build && 
cmake .. && make -j &&
make install

3. build hete:

mkdir build && cd build && cmake ..

set your parameters in hete/src/constants.h and run:

make && ./main

trained models will be saved under:

hete/out/<timestamp>/<cfr_iteration_index>/<player_index>

create and activate virtual environment:

python -m venv env && . env/bin/activate
pip install -r requirements.txt

evaluate against slumbot:

python eval.py --wandb 0 --num_hands 1000 --auto 1

optionally stream to wandb with --wandb 1

plot performance against slumbot:

python plot.py