Clean VQA datasets with an image, a question and an answer. Can be used to train VLMs.