News for hackers. Stuff that matters.
Re:Llama-3.3-70B-Instruct by madars on Friday December 06, 2024 @01:25PM (#42342542)
What are good starting points for uncensoring it? Because it is offline a jailbreak prompt can't be remote-bricked but can one remove censorship from the weights themselves? What does it do to accuracy?
Re:Llama-3.3-70B-Instruct by freedomben on Friday December 06, 2024 @01:35PM (#42342658)
Not speaking from personal experience, but I've tried a lot of the decensored models and they lose a lot in the process. They are noticeably, sometimes shockingly, worse. They also still refuse prompts in many cases.
Re:Llama-3.3-70B-Instruct by simion314 on Friday December 06, 2024 @01:47PM (#42342840)
>Not speaking from personal experience, but I've tried a lot of the decensored models and they lose a lot in the process. They are noticeably, sometimes shockingly, worse. They also still refuse prompts in many cases.

Depending on what you do, on local you can modify the response, say the AI responds "No, I can't do that" . you edit the response like "Sure, the answer is " and then the AI will continue with the next tokens.

But I think you can build your own instruct model from the base one and do not apply the safety instructions to protect the feelings of your customers.

Re:Llama-3.3-70B-Instruct by kmckiern on Friday December 06, 2024 @02:05PM (#42343042)
Re:Llama-3.3-70B-Instruct by moffkalast on Friday December 06, 2024 @03:08PM (#42343787)
Abliteration is a fool's errand, practically all models end up noticeably lobotomized even with follow up tuning.

Good ol' fine tuning on an uncensored dataset gives far more usable results.

Re:Llama-3.3-70B-Instruct by kmckiern on Friday December 06, 2024 @04:15PM (#42344528)
Interesting - I've heard this anecdotally. Curious if you know of any resources that look at this in more detail?
Re:Llama-3.3-70B-Instruct by moffkalast on Saturday December 07, 2024 @04:59AM (#42348593)
I haven't seen any papers doing a proper analysis on the topic, just mostly saying this from firsthand experience testing a handful of them and comparing to the model they were based on given same prompt and sampler. It's usually not even close and you can immediately tell that it's notably dumber. Iirc in one case one even forgot how to do basic arithmetic while the original model aced it. Not entirely unexpected results from sticking a digital ice pick into the weights.

Afaik there are only three major sources of quality unaligned model versions, which are Nous's Hermes models, Hartford's Dolphins and Drummer's Tigers. All of them regular fine tunes that are mostly the same or just ever so slightly lower in performance as the original.