Comment by stillsut

I'm also working on a library to steer the sampling step of LLM's but more for steganographic / arbitrary data encoding purposes.

Should work with any llama.cpp compatible model: https://github.com/sutt/innocuous