Learning track
Post-training & Alignment
Turn pretrained predictors into useful assistants through instruction tuning, preference learning, reinforcement learning, and distillation.
Post-training & Alignment
Turn pretrained predictors into useful assistants through instruction tuning, preference learning, reinforcement learning, and distillation.
- 07.01Why base models aren't assistants→
- 07.02Supervised fine-tuning & instruction data→
- 07.03LoRA, QLoRA & PEFT→
- 07.04Reward models & human preference data→
- 07.05RLHF with PPO→
- 07.06DPO: skipping the reward model→
- 07.07ORPO, KTO, SimPO & the alignment zoo→
- 07.08GRPO & RL on verifiable rewards→
- 07.09Reasoning models: test-time compute & long CoT→
- 07.10Constitutional AI & RLAIF→
- 07.11Distillation: making small models punch up→