K-Culture GlossaryㅍTechnical words in the news
Post-training
The stage after pre-training where a model's behavior and abilities are fine-tuned. This is where most differences in model performance come from these days.
In plain words
AI models are built in two stages. First there's pre-training, where the model reads massive amounts of text from the internet. After that comes post-training, the stage where the model's attitude and abilities get refined.
To use an analogy, pre-training is like reading an encyclopedia to build up knowledge of the world, while post-training is like on-the-job training: "answer customers politely," "refuse dangerous requests," "think step by step." Concretely, things like RLHF (reinforcement learning from human feedback), instruction fine-tuning, and safety training all fall under post-training.
These days, differences between models often come down to post-training rather than pre-training. Even models of the same size can end up with very different coding ability, conversation quality, and safety depending on how they're refined. When news says a company "strengthened post-training," it means they improved the model's real-world usefulness.
How it shows up in the news
"Pre-training stayed the same, but post-training boosted coding performance by 30%." — This means the base capability was the same, but the model was differentiated through job-specific training.