Create sanitized training set

[nadahlberg]

Use disagreements between the trained model and the synthetic data to drop potential mis-annotated examples. Do the same using low confidence predictions from the model.