Train small LMs on text with (un)desirable attributes for efficient decoding-time steering of large models eg. GPT3

Untitled