kl-regularization

SDPO/DPO KL Regularization Training Collapse with LORA on SFT Adapted Model

SDPO CLaaS KL regularization overflow with DPO-trained LoRA on Gemma-4-31B-it