Cached text and image embeddings of the datasets, one repo per encoder; the -tar ones per fold.