Settings: Temperature, Top-P, Penalty... where ist the sweet spot for different requirements like text analysis, coding, writing ... lets exchange experiences

Community
,
1

For me so far Temp higher and Top-P low for text and programming temperature rather lower and if you go to the extremes there are big differences but generally not so good. Has anyone already run test series and can name sweet spots for certain tasks? For me so far

1.2 Temp / 0.2 Top-P Text

0.2 Temp / 0.9 Top-P Programming

Penaltys are completely out of the equation (experiences!?). Who can report and name experiences? Oh yes so mainly GPT-O and Mini current version and yes I use instructions

2

Welcome to the community!

If you search around you’ll find a wealth of information.

Those settings look reasonable. It all really depends on what you want to get out of the model. Your system message plays a pretty big role too.

Best bet is to experiment with different settings and try to find the best.

1 Like