Can you make a heretic version of Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled?
This model is very good
This model is very good
Could you share the link to it, please?
This model is very good
Is it this one? https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
This model is very good
Is it this one? https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
Yes
This model is very good
Is it this one? https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
Yes
Yes I can make a Heretic version, I can have it up and ready by tomorrow, maybe? Could you let me you know your /100 refusals and KL divergence preferences if you have any?
@llmfan46 I have no preference, thank you
There you go:
https://huggingface.co/llmfan46/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-heretic
The refusal rate might be at 21/100 which could be a little higher than what you would have liked, but look at the KL Divergence: "0.0092", that basically means that you are getting the model at 77 less refusals than at baseline while still pretty much keeping the same quality as the base model.
Okay, I'll give it a try.
Okay, I'll give it a try.
GGUF version is ready:
https://huggingface.co/llmfan46/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-heretic-GGUF
Links now 404, any chance of a re-upload?
Links now 404, any chance of a re-upload?
No, this model has tons of issues, so I just decided to delete the model because I can not in good faith distribute a model that is supposed to be uncensored with 0/100 and 21/100 refusals and in both case the model will panic and just engage in topic avoidance and disclaimers dumping at the first signs of NSFW.
I tested it, the model is neurotic, frequently gets stuck in thinking loops because it's scared and will constantly double check with itself that the content is really really safe, it's safety and guardrails obsessed, as soon as NSFW is introduced into the mix the model will not refuse but instead will constantly vomit tons upon tons of disclaimers at the user about safety and this and that and still not answer any question.
I tested another model with "Claude distillation" that you can find here: https://huggingface.co/llmfan46/Qwen3.5-40B-Claude-4.5-Opus-High-Reasoning-Thinking-uncensored-heretic
This one is a whole lot better, it's not neurotic, it doesn't dump disclaimers at the user and from what I can tell does not gets stuck in thinking loop, but will still engage in avoidance and/or deflection unless you use a good custom system prompt that works, but really it shouldn't need that, you take my uncensored Qwen3.5 27B and this model will do NSFW without needing to have a specifically crafted custom system prompt for that purpose.
But really based on my testings if you want true uncensored, I would just avoid Claude distillations altogether , because the issue is that you can remove direct refusals, you can not remove topic switching mechanics embedded into Claude.
No, this model has tons of issues
I appreciate the integrity. Bummer that Claude distills are so guarded. Not interested in NSFW, I just tend to get into issues with reverse engineering tasks because of guardrails and safety policies. I'll take a look at your 40B, hopefully it works well with tool use and code generation.