GRP-Obliteration: Unaligning LLMs with a Single Unlabeled Prompt

23 pointsposted 2 days ago
by vital101
(arxiv.org)

9 Comments

tyfon

2 days ago

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

basilikum

2 days ago

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?

OutOfHere

2 days ago

You won't get banned for something that's not clearly illegal. I have tried lots of prompts like it, and that one too.

thejazzman

2 days ago

i got banned and they refused to tell me why after multiple support tickets. this was > 1y ago. i definitely didn't do anything wrong (i wasn't even using it) so it was either a compromised key or they make mistakes. playing with prompts could potentially get you swept up into some nonsense like that.

not to discourage anyone, just saying.

OutOfHere

2 days ago

What do you recall doing just before you got banned? And dare I ask which provider?

thejazzman

2 days ago

OpenAI and I truly have no idea. Random programming questions via ChatGPT.com maybe as it was pre codex.

I think I gave Plexamp (app) an API token, that’s the only thing I’ve ever come up with on my own speculation

OutOfHere

2 days ago

Can you sign up again under a new account? Does it let you use it this way?

stratos123

2 days ago

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.