undefined | Better HN

0 pointshappycube3y ago0 comments

There are two major advantages though - you can retrain them, and they're not on the guardrails that the commercial models have.

0 pointshappycube3y ago0 comments

There are two major advantages though - you can retrain them, and they're not on the guardrails that the commercial models have.

0 comments

5 comments · 1 top-level

gre3y ago· 4 in thread

I tried to prompt vicuna to tell me a joke about gay people and it refused. Some of the guardrails are still in there.

It's because vicuna is fine-tuned on chatGPT answers. LLaMa will not do this, but LLaMa-based models fine tuned with chatGPT answers will.

Just curious why don't they training except denial responses. Is Copying ChatGPT ethics a purpose?

occz3y ago

Did you use the censored or the uncensored variant?

gre3y ago

It's just a random one from huggingface. I will look for the uncensored one later. Thanks, I think.

j / k navigate · click thread line to collapse