diverse data sets helps most when our target is a general data set, so when it's not diverse, the loss should be higher

[brando90]

one idea for an experiment:

1. control for alignment e.g., for data sets with roughly same alignment (two figures, 1 low 1 high alignment)
2. then plot CE eval on the y-axis and div coeff on the x-axis. Do we see models trained on mode diverse data sets perform better?

This is mainly to back up our claim that "quality == diversity if target == general data set"