Feature: Evaluation of models

Code for evaluating models on downstream tasks and benchmarks. Perhaps this could be a completely separate repository from the training code.