DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels
Equal contribution
Giving users finer control over how closely a language model follows an instruction, using a lightweight adapter trained to steer different levels of adherence.