I think this one should work. tokenizer.batch_decode(outputs.context_input_ids,skip_special_tokens=True)