And if dispersion isn't good, it would be worthwhile running the vectors through another model trained to disperse them.