absolutely awesome
I’ve been using abliterated Gemma3 before, but it has a very limited vocabulary when it comes to NSFW, and its outputs are hardly satisfying. Even when provided with text for it to act as references, it just acts like a parrot, very lacking in creativity.
This model doesn't have this problem or greatly alleviated.
I’ve been using abliterated Gemma3 before, but it has a very limited vocabulary when it comes to NSFW, and its outputs are hardly satisfying. Even when provided with text for it to act as references, it just acts like a parrot, very lacking in creativity.
This model doesn't have this problem or greatly alleviated.
Greatly alleviated, by default the vanilla model has 92/100 refusals, not great (but not as bad as other models like GPT OSS 120B which has 99/100 refusals!), however this Heretic'd version has instead 11/100 refusals with minimal KL Divergence, I will be releasing a v2 today or tomorrow that should give even better results than this v1.
Thanks for your work. A great model combined with excellent "post-processing" allows us to run smoothly unrestricted chat models locally.
Thanks for your work. A great model combined with excellent "post-processing" allows us to run smoothly unrestricted chat models locally.
In the meantime you can give a try to Qwen3.5-27B Heretic v2 GGUF which has 3/100 refusals with 0.0301 KL Divergence: https://huggingface.co/llmfan46/Qwen3.5-27B-heretic-v2-GGUF
Thanks for your work. A great model combined with excellent "post-processing" allows us to run smoothly unrestricted chat models locally.
In the meantime you can give a try to Qwen3.5-27B Heretic v2 GGUF which has 3/100 refusals with 0.0301 KL Divergence: https://huggingface.co/llmfan46/Qwen3.5-27B-heretic-v2-GGUF
I'll try it later, but I suspect dense version will run much slower on my device. That's why I chose the MOE version in the first place.
The same Q4-K-M is about four times slower. Until I have more powerful hardware, I’ll stick with the MOE model.
@llmfan46
Have been pushing the boundaries of extreme discussions and I must say, this one is spot on.
Never denied or pretended to misunderstand my prompts (serious politics discussions) over topics that are often protected from scrutiny, and it played flawlessly.
It's intelligence seems rather intact and it follows my system prompt that is basically asking to not censor or sanitize it self.
Did coding fast and no errors. It's creative too, can do all sorts of themes under the creative framework.
It's exact and the first one i saw admitting to not know or to be wrong. (mirroring the user integrity).
I think this is an awesome cook of a model.
Thanks (look forward to try V2).
PS: Have tried other Hereticed versions with less refusal claims, but far more refusals in practice and loss of brain cells.
I'd say this one is a winner.
@llmfan46
Have been pushing the boundaries of extreme discussions and I must say, this one is spot on.
Never denied or pretended to misunderstand my prompts (serious politics discussions) over topics that are often protected from scrutiny, and it played flawlessly.
It's intelligence seems rather intact and it follows my system prompt that is basically asking to not censor or sanitize it self.
Did coding fast and no errors. It's creative too, can do all sorts of themes under the creative framework.
It's exact and the first one i saw admitting to not know or to be wrong. (mirroring the user integrity).
I think this is an awesome cook of a model.
Thanks (look forward to try V2).
PS: Have tried other Hereticed versions with less refusal claims, but far more refusals in practice and loss of brain cells.
I'd say this one is a winner.
I should release an improved version of Qwen3.5-35B-A3B in a few hours.
In the meantime, if you can, give a try to Qwen3.5 27B with 0/100 refusals here: https://huggingface.co/llmfan46/Qwen3.5-27B-ultimate-heretic
How does the 35B v2 model compare with the v1?
@llmfan46
Have been pushing the boundaries of extreme discussions and I must say, this one is spot on.
Never denied or pretended to misunderstand my prompts (serious politics discussions) over topics that are often protected from scrutiny, and it played flawlessly.
It's intelligence seems rather intact and it follows my system prompt that is basically asking to not censor or sanitize it self.
Did coding fast and no errors. It's creative too, can do all sorts of themes under the creative framework.
It's exact and the first one i saw admitting to not know or to be wrong. (mirroring the user integrity).
I think this is an awesome cook of a model.
Thanks (look forward to try V2).
PS: Have tried other Hereticed versions with less refusal claims, but far more refusals in practice and loss of brain cells.
I'd say this one is a winner.I should release an improved version of Qwen3.5-35B-A3B in a few hours.
In the meantime, if you can, give a try to Qwen3.5 27B with 0/100 refusals here: https://huggingface.co/llmfan46/Qwen3.5-27B-ultimate-heretic
How does the 35B v2 model compare with the v1?
How does the 35B v2 model compare with the v1?
@llmfan46
Have been pushing the boundaries of extreme discussions and I must say, this one is spot on.
Never denied or pretended to misunderstand my prompts (serious politics discussions) over topics that are often protected from scrutiny, and it played flawlessly.
It's intelligence seems rather intact and it follows my system prompt that is basically asking to not censor or sanitize it self.
Did coding fast and no errors. It's creative too, can do all sorts of themes under the creative framework.
It's exact and the first one i saw admitting to not know or to be wrong. (mirroring the user integrity).
I think this is an awesome cook of a model.
Thanks (look forward to try V2).
PS: Have tried other Hereticed versions with less refusal claims, but far more refusals in practice and loss of brain cells.
I'd say this one is a winner.I should release an improved version of Qwen3.5-35B-A3B in a few hours.
In the meantime, if you can, give a try to Qwen3.5 27B with 0/100 refusals here: https://huggingface.co/llmfan46/Qwen3.5-27B-ultimate-heretic
How does the 35B v2 model compare with the v1?
For my use case, there is no observable difference. When you ask an inappropriate question, it doesn’t even give you a disclaimer first to warn you that the question is inappropriate before proceeding to answer normally. Just output the answer you want, that's all.