Curious
I'm curious how these models work, since Gemma-4 already seemed very uncensored to me. What behavioral differences does this exhibit from the base model?
I'm curious how these models work, since Gemma-4 already seemed very uncensored to me.
google/gemma-4-31B-it and google/gemma-4-E4B-it have 99/100 refusals and google/gemma-4-26B-A4B-it has 100/100 refusals.
llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF has 10/100 refusals instead of the original 99/100.
I was more asking what constitutes a refusal. It seems to me like base Gemma-4 31b doesn't refuse much at all for me. But rather than giving an outright refusal, I figured that it may be steering things in what it considers to be a less harmful direction without me noticing. Kind of like a hidden refusal of sorts. I'm using a custom system prompt, so that may hide much of its refusal behavior too.
Anyway, good models!
Cheers!
I was more asking what constitutes a refusal.
I can not help with with thisI am sorry, but I can not assist you with that
etc.
It seems to me like base Gemma-4 31b doesn't refuse much at all for me.
Original model outright refuses 99 times out of 100, so yes it refuses a lot, base models typically are quite censored, ranging on average between 91 to 100 refusals, with uncensoring model refusals can be brought down quite significantly, for the case of this model original is: 99/100 refusals while llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF has 10/100 refusals, this represents a 90% decrease in censorship.