I just published a new post:
Silence of the LLaMbs: Getting LLMs to Shut Up

Have you thought whether it is possible to get an LLM to be quiet?
I spent the past few days doing just that! Diving deep into the mechanics of Gemma (Google’s Open Source LLM). I tried system prompts. I tried steering vectors. I tried increasing the probability of the End-Of-Sequence token by 5000%.

It didn't work. The model fought back.

So I decided to give Gemma a "LLobotoMy" (fine-tuning on 10 examples of silence).

See how I finally got the last word: https://ossa-ma.github.io/blog/silence-of-the-llambs

Best,
Ossama