Curious how it'll perform when people start fine-tuning on smaller, specialized datasets