Free account
Read this concept with a free account
Read this concept at no cost by signing in with Google.
no card needed
Core
SFT, RLHF and DPO: Choosing the Behavior You Train For
Understand what demonstration and preference objectives change, how they differ from LoRA and synthetic data, and how to evaluate the resulting behavior without mistaking preference for correctness.
CONCEPT / HOW LLMS WORK, FOR PMS
RELATED CONCEPTS
PRACTISE THIS CONCEPT
LLM & RAG FundamentalsYour assistant follows the format but gives polished, unhelpful answers. Would SFT, RLHF or DPO help?→AI Platform & Model PMA new model wins 64% of preference tests but expert users prefer the old one. Would you replace it?→AI Platform & Model PMYou have $30,000 to improve a task assistant. Buy demonstrations, preference labels or execution feedback?→AI Platform & Model PMYou need preference labels for a model update. How would you design the work, pay annotators and control quality?→