But it's easy to do that for a small subset for finetuning compared to cleaning up the entire dataset.