Most NLP metrics take as an input predictions and references except perplexity which takes input_texts. We could easier combine them by renaming the input_texts to predictions. What do you think @sashavor and @TristanThrush?