absolutely great!
I have tried a lot of gemma-4-31b-it finetunes, gemma-4 being my current favourite. And while I have limited experience with this one yet, the model card does not lie: minimal damage, and great writing style. Pretty sure this is my current go-to model now. At least, it made me make this semi-excited comment :) Seems to be a great method. Thanks!
I haven't tried much 31B models myself, the speed penalty is enough where i prefer 26B models (generally to RP). Actually prefer MOE models in general after seeing the speed boost if that's how they are going to perform.
Still, might be worth a try.
yeah, gemma-4-26b-a4b is a speed monster even on pure cpu, or with very little vram. i wonder if a gemma-4-26b-a4b-styletune is in the cards, it has even more need for it than the 31b :)
It wasn't in the cards until a minute ago. 😉 I'll see what I can do!
Just a humble note. I'm currently testing this with Gemma4-12b-it and it seems to be working very well, too. Many thanks!
Great to hear! I love how fast 26B-A4B is to train out of the box, so https://huggingface.co/Gryphe/Gemma-4-26B-A4B-StyleTune is a thing now. Enjoy!
Gemma-4-26B-A4B-StyleTune is a thing now. Enjoy!
Woo! Once it's GGUF'd i'll be downloading it. I'm sure it's already in the pipeline for @mradermacher .
Absolute cinema! I hope you'll release version 12b too.
wow, heroic overnight delivery :)
And from what I can see, it's a success. The 26B probably benefits even more than the 31B, style-wise. But damn, the 26B is much more stubborn with refusals.
And from what I can see, it's a success. The 26B probably benefits even more than the 31B, style-wise. But damn, the 26B is much more stubborn with refusals.
Then should it have been heresy/uncensored patched before finetuning with the style? Or go the other way around?
One tensor changed out of 659.
Hmmm which tensor was it i wonder? By splitting the individual layers out i could swap it with an ARA heresy one i have enjoyed and see how that fares.