Important to note, they say "From these generated models, we select the one with the best performance on the training set." Definitely potential for bias here.