Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

You could have the models output a confidence alongside next-token then weight the penalty by the confidence.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: