Comment by amenhotep

Comment by amenhotep a year ago

3 replies

View on Hacker News

Uncensored finetunes aren't the same thing, that's taking a model that's already been lobotomised and trying to teach it that wrongthink is okay - rehabilitation of the injury. OpenAI's uncensored model would be a model that had never been injured at all.

I also am not convinced by the argument but that is a poor reason against.

staticman2 a year ago

I'm talking about taking the Llama 3 base model and finetuning it with a dataset that doesn't include refusals, not whatever you mean by "taking a model that's already been lobotomized".

It's interesting that you weren't convinced by the above argument but still repeated the edgelord term "lobotomized" in your reply.

Reply View 2 replies
  • errantspark a year ago

    The claim is that llama is "lobotomized" because it was trained with safety in mind. You can't untrain that by finetuning. For what it's worth the non-instruct llama generally seems better at reasoning than instruct llama which i think is a point in support of OP.

    Reply View 1 reply