Training & post-training
Post-training
Training after pretraining to improve usefulness, reasoning and behaviour.
Definition
Post-training includes supervised fine-tuning, preference optimisation and reinforcement-learning methods applied to a pretrained model.
Why it matters
Many visible differences between frontier models arise from post-training rather than base architecture or parameter count.
Related concepts
- Supervised fine-tuning
Fine-tuning on curated examples of desired inputs and outputs.
- Reinforcement learning from feedback
An umbrella for reinforcement learning guided by preference or quality feedback.
- Direct preference optimisation
A method that learns from preferred answers without a separate reward-model loop.