post-training

Definition

Everything done to a pretrained model to make it usable: supervised fine-tuning on demonstrations, preference-based reinforcement learning such as RLHF or DPO, safety training, and increasingly reasoning-focused training. Pre-training installs knowledge and general competence; post-training decides how that competence is expressed — format, refusal behavior, instruction following, willingness to reason at length. Among models of comparable scale, most of the visible quality difference now originates here.

Explainers using this term

• Paper Explained: Video Generation Models: A Survey of Post-Training and Alignment — Four Toolboxes for Making Video Models Do What You AskVideo Generation Models: A Survey of Post-Training and Alignment
• Paper Explained: Post-Training Leaves Behavioral Shadows on Unrelated Decisions — capability transfer through one unrelated wordPost-Training Leaves Behavioral Shadows on Unrelated Decisions
• Paper Walkthrough: Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training — training the draft alongside a 122B policy at 256K tokensOnline Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training