Post-training: RLHF, RL and reasoning

From base model to assistant: fine-tuning, RLHF, constitutional methods and RL for reasoning.

5 lessons planned · none written yet

  1. 1Base models versus assistantsNot written yetWhat post-training changes.
  2. 2Supervised fine-tuningNot written yetLearning from demonstrations.
  3. 3RLHFNot written yetReward models and preference learning.
  4. 4RL with verifiable rewardsNot written yetHow reasoning models learn to think longer.
  5. 5AlignmentNot written yetGoals, honesty and failure modes.