Improving Model Safety Behavior with Rule-Based Rewards
来自 OpenAI News
· 2024-07-24
精选
We've developed and applied a new method leveraging Rule-Based Rewards (RBRs) that aligns models to behave safely without extensive human data collection.