Issues when fine tuning Llama-3.2-11B-Vision

And perhaps:

you can use return_full_text=False