CryptoHacker Noonabout 9 hours ago
LLM Post-Training Map: SFT, RLHF, DPO Explained
13 min read
Post-training has four stages: pretraining, SFT, preference alignment (RLHF), and RLVR. Reward models are supervised, not RL. DPO replaces RLHF's reward model and RL with direct preference optimization. PPO adds a critic for per-token credit assignment.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyGoogle
Summary by ByteBrief
