Minor SFT loss for LLM fine-tune to increase performance and reduce model deviation

Re-assign community