3 ms·
Base LLaMAv1 would refuse to answer certain questions. It wasn't as aggressive as OpenAI models or the safety aligned finetunes, but some kind of alignment was
by brucethemoose2 3y ago
Base LLaMAv1 would refuse to answer certain questions. It wasn't as aggressive as OpenAI models or the safety aligned finetunes, but some kind of alignment was there.
- astrange 3y agoNormal training content has "alignment". It's not going to instantly be super racist and endorse cannibalism if it's "unaligned".
- brucethemoose2 3y agoIt very specifically mentioned something about LLaMA not being trained to answer that in the response. Again, its extremely minimal, but I think it picked something up from the Llama info facebook inserted.
- refulgentis 3y agoSigh. It's an LLM.