This would be an interesting way to distill/finetune a model by finding the optimal minimum corpus needed to match a given use case.