Sampler settings please?
Hello,
I must've been doing something wrong, because I can't get the creativeness out of it. The stories I generate seems very generic compared to the original monstral and the behemoth 1.1 merge with magnum.
My settings are: min_p=0.02, dry=0.8/1.75/2, xtc=0.1/0.5. These are working with the first monstral.
Is this settings not optimal? Can you share the best settings for the new Monstral?
Thanks!
Try the following settings:
temp=1
min_p=0.02
dry=0.8/1.75/3
Neutralise all other samplers
The key take away is to lower your min P drastically; 0.2 is way too high and will limit creative freedom. I also don't find it necessary to employ XTC with this model. In addition, the importance of the right system prompt cannot be underestimated with Largestral models. Small changes here can make big differences to the output. I recommend the following one if you use Sillytavern, as curated on Drummer's Discord (credits to Konnect for coming up with it): https://pastebin.com/raw/QbEeJxtK
Try the following settings:
temp=1
min_p=0.02
dry=0.8/1.75/3
Neutralise all other samplersThe key take away is to lower your min P drastically; 0.2 is way too high and will limit creative freedom. I also don't find it necessary to employ XTC with this model. In addition, the importance of the right system prompt cannot be underestimated with Largestral models. Small changes here can make big differences to the output. I recommend the following one if you use Sillytavern, as curated on Drummer's Discord (credits to Konnect for coming up with it): https://pastebin.com/raw/QbEeJxtK
Thanks! Actually I misstyped, my min_p is actually 0.02. I'll check out the system prompt.
Update: the preset I recommended now has an official HuggingFace page, and has been thoroughly updated. I recommend importing and using it while using this model: https://huggingface.co/Konnect1221/Methception-SillyTavern-Preset
In addition, it's worth experimenting with the XTC sampler. Monstral V2 is incredibly intelligent, enough so that introducing XTC has largely nothing but benefits, in my experience, after having tried it properly. If you do try XTC, set your temperature to 1 and min P to 0.02, and try values of 0.05-0.2 threshold and 0.2-0.5 probability. In my experiments, a value of 0.1 threshold and 0.3 probability is enough to kill off any repetitive sentence structures, slop and dialogue that this model is otherwise prone to. It's transformative.
Edit: Upon further testing, I've settled on values of 0.12 threshold and 0.5 probability.