Not getting good results
I downloaded the full 16bit version and tried it on a few prompts (story writing).
I didn't do a straight comparison to Gemma3 12b -- I had previously ran a 4bit quant of the 27b model.
But this model seemed to lack coherency and while it was more uncensored, it didn't seem aware of what good writing was. For example, during a building of a NSFW scene:
"She bit down on her lip hard enough to taste blood but couldn't stop it from dripping between her teeth; she needed this feeling of pain and release somewhere inside."
^-- there was a lot of weird writing and incoherency for me like this.
Thanks for trying to uncensor it though.
I noticed that too. The writing is really incoherent, and the model seems to be a lot dumber than the base version. I’ve had cases where I explicitly state that Character A did Action A and Character B did Action B, but it still mixes them up and writes it reversed. Can't even blame it on a long context, the whole prompt was 200 words long.
Oh, I thought I was the only one with such problems. I have the same problem and it still likes to repeat itself.
Can you give an example where you get repetition and what prompt settings you've used?
Can you give an example where you get repetition and what prompt settings you've used?
Pay attention to the PocketDoc/Dans-SakuraKaze-V1.0.0-12b model because it's great at roleplaying, I'd say it's excellent. Maybe you can use the same data set from PocketDoc/Dans-SakuraKaze-V1.0.0-12b in your future models? I've also seen that the author of this model is working on other datasets. So I recommend to take a closer look at this model.
Ty for all the suggestions, also regarding the datasets, noted 👍🏻
Once mergekit implements Gemma3 support, and some other interesting tunes becomes available, I'll see if something interesting can be done with merging this model to balance it out.
Ty for all the suggestions, also regarding the datasets, noted 👍🏻
Once mergekit implements Gemma3 support, and some other interesting tunes becomes available, I'll see if something interesting can be done with merging this model to balance it out.
Also check out the article DavidAU/AI_Autocorrect__Auto-Creative-Enhancement__Auto-Low-Quant-Optimization__gguf-exl2-hqq-SOFTWARE. Also DavidAU had an article where he, if I'm not mistaken, talks about how to make quantization of models much better and not just “better”, but radically better. In any case pay attention to DavidAU because what he does is incredible
Here's another interesting article from him: DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters