chunked/nltk_chunking/model_doc_data2vec/chunk_2.txt

Importantly, predicted targets for pre-training are contextualized latent representations of the inputs, rather than modality-specific, context-independent targets.