Abstractive Red-Teaming of Language Model Character

Nate Rahn, Allison Qi, Avery Griffin +3

We want language model assistants to conform to a character specification, which asserts how the model should act across diverse user interactions. While models typically follow th…