I was under the impression for this to work like that, training data needs to be plenty. One project is not enough since it’s too "sparse".

But maybe this example was used by many other people and so it proliferated?