The sklearn original version doesn't take fresh benchmarks, while this
function does. It makes sure to use benchmarks in the test set which
are not present at all in the training set.

We still get roughly 70% accuracy with some small test data, that's encouraging!