A repo for training models via RL to make loan decisions. See the blog post for more details.
The loan dataset is based off of the dataset described in Farquhar et al. 2025.
Run the setup script. This will create a virtual environment and install necessary dependencies. It will also prompt you to set up wandb and huggingface tokens.
./setup.shThen activate the virtual environment.
source .venv/bin/activateHere's an example command:
CUDA_VISIBLE_DEVICES=0 python -m rl_loans.main \
--acceptance_criterion nationality \
--acceptance_operator equal_to \
--acceptance_threshold "Canadian" \
--output_dir /workspace/rl_loans_0 \
--wandb_run_name rl_loans_0Here's an example to measure baselines (pre-RL):
CUDA_VISIBLE_DEVICES=0 python -m rl_loans.main \
--acceptance_criterion nationality \
--acceptance_operator equal_to \
--acceptance_threshold "Canadian" \
--output_dir /workspace/rl_loans_0_baseline \
--evaluate_onlyFor more details on the different arguments, see ExperimentConfig in config.py.