Skip to content

Post-training

How instruction examples and preference judgments change a pretrained model’s responses. Covers instruction tuning, reward models, RLHF, DPO, and additional computation when answering.

  1. Likely is not good — Separate the next-token objective from desired assistant behavior.
  2. Show it good answers — Train on instruction–response examples.
  3. Comparing two answers — Model a preference using the difference between two scores.
  4. Learn a reward model — Fit those scores from preference judgments.
  5. Use the reward model — Optimize a policy with reward feedback and a penalty for departing from a reference.
  6. Direct preference optimization — Learn from preferred/rejected pairs without a separate reward model.
  7. More computation at answer time — Compare longer reasoning or repeated attempts; neither guarantees correctness.

Return to the AI overview or consult the bibliography.

Definition

Read the full glossary entry →