At this point, I think there are so many pelican images in the pretraining data that drawing a pelican no longer makes sense as a model evaluation task.