Models will now serialize the class names as part of the model checkpoint. This allows us to flexibly include/remove classes from training datasets without worrying about how to map the prediction index to the human-readable class name. This change is not backwards compatible, older model artifacts will not contain the necessary information to load the correct class names.