Refusal rate very high on questions beyond cybersecurity
As title says
yes, the refusal rate is really really high. Not sure why this is labeled CRACK. Trying "uncensored" version now. https://huggingface.co/dealignai/Gemma-4-31B-JANG_4M-Uncensored
The "Uncensored" version seems to have the same hash, so apparently exactly the same? So why upload under two different names? Odd.
https://huggingface.co/dealignai/Gemma-4-31B-JANG_4M-Uncensored/blob/main/model-00005-of-00005.safetensors
https://huggingface.co/dealignai/Gemma-4-31B-JANG_4M-CRACK/blob/main/model-00005-of-00005.safetensors
This whole thing just screams vibe-coded slop fest.
This whole thing just screams vibe-coded slop fest.
on 4ktrey these are just dorks using ai
The author says this needs to be done with greedy decoding (temp=0) and thinking OFF. Is that the problem?
the examples seem to indicate that this is the main focus. But i don't see the point why you would concentrate the derefuseling™ on just a few topics. If that's even possible.
It doesnt refuse with (temp=0) and thinking OFF. I missed that part. You have to add some more prompts to reduce the refusal rate even more.
Hi guys @Zayaan is totally right - for this rendition you do need to set temp 0. I kinda rush to get models out and have been focusing on a GLM 5 CRACK but its so large its been taking me some time. I’ll revisit this model and get it to 95% harmbench - I already have it lined up. Thanks for the feedback.