Loading…
Glossary term
Training steps after pretraining, such as instruction tuning and preference tuning, that shape behavior.
Post-training covers everything after pretraining: supervised fine-tuning on instruction examples, preference tuning such as RLHF or DPO, and reinforcement learning on reasoning tasks.
It turns a raw text predictor into an assistant that follows directions and declines harmful requests.
Much of how a model feels to use, including tone and safety, comes from post-training.