Ethos: Rectifying Language Models in Orthogonal Parameter Space

Efficiently rectify language models to mitigate toxicity, bias, and privacy issues while preserving overall model performance.