GGUF
Not-For-All-Audiences
nsfw
Inference Endpoints

Gibberish output.

#3
by Huegli - opened

Hi,
I tried this model in SillyTavern, using KoboldCPP. The output is horribly gibberish, and I hope you can help me. I've noticed the same problem with other Mythomax merges also. While spoken text is always clear and in good english, descriptive text is just a gibberish sequence of words without any grammar, pronouns, articles and so on.

Example:
"Kurumi winks impudently upon overhearing request looking sideways sizing opportunity opening up possibly turning conversation towards different course entirely changing potential outcome destined way events might unfold hereafter should someone answer open ended questions asked unable otherwise understand subtle intricacies complex world surrounding him presently explore deeper inner reaches mind searching secrets hidden behind closed doors masking true emotions buried deep past layers long forgotten only fragments found amidst haze mist obscuring realities blurring distinction fiction fantasy."

Edit: It seems like Mythomax reacts different to rep. penalty settings then other models, actually wiping out every repetitive word, like articles. Lowering the penalty reduces the effect gradually. It's a pretty decent model, once the obstacles are removed.

Hello, please try with Mirostat 1, it should not use other setting preset if Mirostat is active (or not some of them).
If it's stay gibberish, post again!

Actually, for me it works better without mirostat (1 or 2), once the settings fit the model fine tuned.

I highly recommend to try out the mancer setting suggestions for Mythomax: https://www.reddit.com/r/SillyTavernAI/comments/15tqxnv/mancer_mythomax_settings/ Using those settings, the MLewd models run beautiful, highly intelligent and creative without gibberish outputs, unnecessary repetitions or becoming deterministic. I love this model!

Sign up or log in to comment