prisma parrot
In some models, in my case Mythomax 13b, the AI's responses often continue the conversation on behalf the user with subsequent ### Instruction and ### Response messages.
I've tried adding ### Instruction to the stop sequences array but it does not work, probably because ### Instruction is not a single token?
Is there any way to prevent this?
fringe parrot
which provider(s)?
prisma parrot
which provider(s)?
Happens with many providers, like Fireworks, Lepton, Mancer. I don't currently have the stop sequence configured though, removed it a while ago as it didn't work then.
Should the stop sequence work in this case?
fringe parrot
ah yes add the stop sequence!
prisma parrot
ah yes add the stop sequence!
Last time I tried it didn't work, but I'll try again. What should I add? "### Instruction:" ?
fringe parrot
Yep, no colon at the end
prisma parrot
Is there a limit to stop sequences? I have 5 configured.
If that's the case, one possibility is the provider only supports X amount, and removes the exceeding. I'll try again though and report the results.
fringe parrot
yep most providers limit to ~4