To run the experiments:

  ./beaver.py

to train the shortest-path solving GPT, and

  ./beaver.py --overwrite --one_shot --nb_epochs=10

to train the one-shot MLP read-out.