Ouyang et al. • 2022

Résumé

The InstructGPT paper, which introduced the three-step alignment process (SFT, Reward Modeling, RLHF) used to turn base models into helpful assistants. This methodology made language models suitable for general use.

Pourquoi C'est Important

• Standardized the RLHF pipeline
• Addressed helpful and harmless alignment
• Foundational work for instruction-following models