Training Language Models to Follow Instructions with Human Feedback¶
Authors: Ouyang et al. Year: 2022 ArXiv/Link: https://arxiv.org/abs/2203.02155
Summary¶
InstructGPT scaled RLHF to large language models, creating instruction-following models that better align with human intent.
Key Concepts¶
- RLHF at scale
- Instruction following
- Reward models
- Policy optimization
- Alignment
Impact¶
Enabled practical alignment of large language models
Category¶
Alignment & RLHF