That's pretty interesting. It was already well-known that you could easily remove safety training from open-weights models by a bit of finetuning, but apparently you don't even need a finetuning dataset, as long as you have just a few prompts and another LLM to judge responses? Let's see if the abliteration people take a note of this.