Finetuning through sampling lets supervised fine-tuning match reinforcement learning in post-training

Yilun Du’s team combines sampling with supervised fine-tuning so a language model can learn new knowledge without forgetting what it already knew.