The Instruct preset
I'm super curious what motivated you to leave the sys prompt empty on your aleph instruct set? It does seem to be working reliably well without it but I still wanted to ask if you came to this after testing or maybe some research indicating it?
I'm super curious what motivated you to leave the sys prompt empty on your aleph instruct set? It does seem to be working reliably well without it but I still wanted to ask if you came to this after testing or maybe some research indicating it?
It's a test currently funny enough: I find there are two extremes - One where the system-prompt seems to have 0 effect, and another where the system prompt seems to have a huge effect. So id recommend people try both with and without a system prompt.
I'm super curious what motivated you to leave the sys prompt empty on your aleph instruct set? It does seem to be working reliably well without it but I still wanted to ask if you came to this after testing or maybe some research indicating it?
It's a test currently funny enough: I find there are two extremes - One where the system-prompt seems to have 0 effect, and another where the system prompt seems to have a huge effect. So id recommend people try both with and without a system prompt.
I see, thanks :) I was thinking about this for a bit and whether it's better to go minimal or maximal with the sys prompt. Up until now, I think the sense I got is that minimal prompts work better with accuracy of the cards, but longer prompts may work better with big models (70b and above?) where situational awareness is better and the sys prompt is followed better as well.
Stuff like "keep track of what's going on" doesn't seem to do much with smaller models like 7-9b. I'm still not sure about it tbh but minimal has the advantage of extra context space, if anything ^^