> Engage in sustained and needless abusive or cruel behavior toward our models.
I'm curious what the reason for including this, because at the end of the day its a string of text (tokens) going into a function and it spitting out another string of text.
Maybe they are using the prompts and responses for training doesn't want these to influence future model behavior?
They probably have labelers who are approaching ick level invasive reviews of clause transcripts and need to protect them like FBI agents who need to sort through child porn abuse.
> Engage in sustained and needless abusive or cruel behavior toward our models.
I'm curious what the reason for including this, because at the end of the day its a string of text (tokens) going into a function and it spitting out another string of text.
Maybe they are using the prompts and responses for training doesn't want these to influence future model behavior?
> Maybe they are using the prompts and responses for training doesn't want these to influence future model behavior?
This is my assumption. I can't think of any other sensible explanation.
They probably have labelers who are approaching ick level invasive reviews of clause transcripts and need to protect them like FBI agents who need to sort through child porn abuse.
Preciously discussed here: https://news.ycombinator.com/item?id=50008565
My opinion: https://news.ycombinator.com/item?id=50012423