chunked/nltk_chunking/model_doc_t5/chunk_71.txt

only padding up to the longest example in a
batch) leads to very slow training on TPU.