Baal looks pretty cool.
Let's see if we can use it to select data for more training (active learning) or model calibration (so the predicted probabilities are more meaningul).