I can't get the expected answer when running inference.py in Emu1 using instruct-tuning model. The output is: 