When creating a graph learner that has as input a task with tf columns, the data_prototype that is saved in the learner's state after training contains the arg and value vectors as well as the evaluator and other metadata defined in tf.
This unnecessarily blows up the size of learner states in a way that was not intended.

I think this should be fixed in tf, i.e. 0-lentgh tf vectors should drop discardable metadata.