Skip to Main Content

InstructGPT (1.3B params + RLHF) was preferred over GPT-3 (175B) in 71% of human evaluations.Ouyang et al., 'Training Language Models to Follow…