All training code is in the training folder. We use a variant of off-policy PPO outlined with various other policy graident methods in training/algorithms.py. The main driver code is train.py
Various models are saved with results, where the folder name corresponds to the model's name. For example, expert has all results generated using SCIP's default B&B solver.
For evaluation, you will need to navigate back to ml4co-competition/common/evaluate.py