
Distributional Dispreference Optimization (D2O) achieves alignment using solely human-annotated negative samples, reducing harmfulness while maintaining helpfulness.


coremsg

negating-negatives-achieving-alignment-with-human-annotated-negative-samples


Negating Negatives: Achieving Alignment with Human-Annotated Negative Samples


title_rewrite


Distributional Dispreference Optimization (D2O) enables alignment using solely human-annotated negative samples, reducing harmfulness while maintaining helpfulness.
