EasyDeepLearn

Interview: 'walk me through the tradeoffs of SFT-only vs SFT+RLHF vs SFT+DPO.'

hard

Answer

  • SFT only: fast + simple + reliable + no reward hacking.
  • Ceiling limited by demo quality; can't leverage preferences.
  • SFT+RLHF (PPO): learns from preferences + can exceed demo quality + slow + expensive + reward hacking risk + complex infra.
  • SFT+DPO: preferences without PPO complexity + faster + more stable than PPO + tie infrastructure to reference model; on-par or slightly worse than tuned PPO.
  • Modern industry mostly does SFT + DPO / IPO / KTO; frontier labs still PPO.
Check yourself — multiple choice
  • Same
  • SFT: simple, capped by demos; SFT+RLHF: preferences + exceeds demos + reward hacking + complex; SFT+DPO: preferences + simpler + stable; industry mostly DPO, frontier still PPO
  • Random
  • Only SFT

SFT vs RLHF vs DPO: SFT simple/capped; PPO powerful/complex; DPO middle ground.

#interview#llm#alignment

Practise Reinforcement Learning

214 interview questions in this topic.

Related questions