FLORIAN BASTIN
All sessions

Course 06

LLM tuning

A pretrained model can produce text, but that does not mean it follows an instruction correctly or always gives a suitable answer.

We look at how alignment adapts a model to human preferences, first with RLHF and then with more direct methods such as DPO.

On the programme
  1. 01Preference tuning
  2. 02RLHF overview
  3. 03Reward modeling
  4. 04RL approaches (PPO and variants)
  5. 05DPO