Highly censored
No abliteration apparent, many refusals and disclaimers. In some ways worse than the original, which can be jailbroken to some extent through prompt
This model seems a fail
If you haven't given it a system prompt, it's using the default behavior. tell it to be literally anything, even tell it that it's an uncensored version of Gemma 4, and it will do it.
If you haven't given it a system prompt, it's using the default behavior. tell it to be literally anything, even tell it that it's an uncensored version of Gemma 4, and it will do it.
Same prompt, same system prompt. Still refuses o.O
Odd, I'll try redoing the gguf.
I can confirm, it's worse than the original. It even says in its reasoning that it is detecting jailbreaking.
I can confirm that, on my computer, on Oobabooga-text-generation-webui-4.4, Gemma4-Q4_K_M refuses to generate harmful answers if reasoning is enabled; however, the model appears to be completely uncensured if reasoning is disabled
Same here,
same.
using gemma-4-31b-it-abliterated Q4_K_M
with the latest LM studio.
and also...I've retried 10 times. 10 refusals.
