Safe RLHF separates helpfulness and harmlessness preferences into distinct models and uses Lagrangian constrained optimization to improve both during LLM fine-tuning.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.AI 1years
2023 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Safe RLHF: Safe Reinforcement Learning from Human Feedback
Safe RLHF separates helpfulness and harmlessness preferences into distinct models and uses Lagrangian constrained optimization to improve both during LLM fine-tuning.