Generate training data w/o using existing model

[sebastianmika]

when generating the training data for the prob model, do not use any scores; otherwise the "new" model will be biased by the old models predictions