CLU can automatically split the utterances into a training set and a testing set.
But how does it decide what utterances go into which set?
At first I thought it would choose at random. But when I ran multiple training jobs the resulting scores were exactly the same. So there doesn't seem to be an element of randomness.
This made me wonder how CLU splits the utterances. Does it just use the first x% of utterances for training? Does it decide in some kind of intelligent way?
I know it's possible to manually split the utterances. This question is not about that.
