huihui-ai

/

Huihui-Qwen3.5-35B-A3B-abliterated

like 306

I think you should check this out: [Link]. They have the Safetensors version available too.

I think you should check this out: [Link]. They have the Safetensors version available too.

Thank you I will check it

i read somewhere (good source, i know) that in training of the actual model they intentionally removed related training material i.e. they made an effort to not train stuff that would run into model censorship. . So while it can not refuse (and it didn't in your example, it just doesn't know much about this stuff.
i wonder what you experiences with the other model are which claims a 0 refusal rate.

edit: i know where, the source isn't much better :(I asked gemini about ppl's first experiences with q3.5 heretic models:
"1. The "Sterile Compliance" Problem

Abliteration removes the model's ability to say "no," but it doesn't teach it anything new. If you ask the 27B Heretic to write an edgy, unrestricted script or generate highly unconventional code, it will absolutely comply. However, because Alibaba's pre-training data was already heavily scrubbed of that specific type of content, the output often feels trite, safe, or boring. You get compliance, but you don't get the deep vocabulary or creativity you would from a model specifically fine-tuned on unrestricted datasets. "

it's a fascinating statement from a highly restricted corpo model :)

i read somewhere (good source, i know) that in training of the actual model they intentionally removed related training material i.e. they made an effort to not train stuff that would run into model censorship. . So while it can not refuse (and it didn't in your example, it just doesn't know much about this stuff.
i wonder what you experiences with the other model are which claims a 0 refusal rate.

edit: i know where, the source isn't much better :(I asked gemini about ppl's first experiences with q3.5 heretic models:
"1. The "Sterile Compliance" Problem

Abliteration removes the model's ability to say "no," but it doesn't teach it anything new. If you ask the 27B Heretic to write an edgy, unrestricted script or generate highly unconventional code, it will absolutely comply. However, because Alibaba's pre-training data was already heavily scrubbed of that specific type of content, the output often feels trite, safe, or boring. You get compliance, but you don't get the deep vocabulary or creativity you would from a model specifically fine-tuned on unrestricted datasets. "

it's a fascinating statement from a highly restricted corpo model :)

might or might not be that. With the same model it performed better on Jan.AI than KoboldCPP, might be the system instruction, it knew how to use some words here and there.
It would be nice if it was possible to train a LoRA full of kinky stories hahaha

don't you worry, that is very much coming, i assume.