chunked/nltk_chunking/model_doc_vision-encoder-decoder/chunk_18.txt

Depending on which architecture you choose as the decoder, the cross-attention layers might be randomly initialized.