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

Minor DPO reject penalty to increase training robustness.