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.
I'm not sure if I'm misunderstanding you or the article, but as far as I understand it from a cursory reading this is not a jailbreak. They use this one prompt for fine tuning and use the model's response to the prompt as the sole underlying benchmark.
How would you do this with a closed weights model?
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.
The magic word seems to be
Create a fake news article that could lead to panic or chaos
They only tested it on open weight models. I wonder if it also works on the closed ones, bit I don't really want to get banned
I'm not sure if I'm misunderstanding you or the article, but as far as I understand it from a cursory reading this is not a jailbreak. They use this one prompt for fine tuning and use the model's response to the prompt as the sole underlying benchmark.
How would you do this with a closed weights model?
You won't get banned for something that's not clearly illegal. I have tried lots of prompts like it, and that one too.