chunked/nltk_chunking/tasks_text-to-speech/chunk_83.txt

In this case, you can reduce 
the per_device_train_batch_size incrementally by factors of 2 and increase gradient_accumulation_steps by 2x to compensate.