News for hackers. Stuff that matters.

A comment on GPT trainer says he's traumatized from the RLHF work. Parent comment

Re:GPT trainer says he's traumatized from the RLHF work by rngname22 on Friday May 19, 2023 @02:49PM (#36005504)
"prompts submitted by our customers to the API"

^ they were reading and labeling explicit prompts provided by users. And explicit responses by the model to those explicit prompts.

[ Parent | on HN ]
Re:GPT trainer says he's traumatized from the RLHF work by Imnimo on Friday May 19, 2023 @03:53PM (#36006114)
I guess I would have expected the dataset to have only the occasional horrible prompt if it were just user-provided data, but I got the sense from the article that these guys were labelling horrific text all day.
[ Parent | on HN ]