google-research-datasets/conceptual-12m: Conceptual 12M is a dataset containing (image-URL, caption) pairs collected for vision-and-language pre-training.