Finetuning GPT2 with user defined loss