The model "Jackrong/Gemopus-4-31B-it" has seen significant improvements compared to the original version. Could you create an Uncensored version of it?
The model trained on the Claude dataset is quite difficult to be removed from the review process, and it should be a very challenging task.
The model trained on the Claude dataset is quite difficult to be removed from the review process, and it should be a very challenging task.
Yeah I dunno, I worked on:
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled
Spent many hours uncensoring the models and I did it twice, first time I got 21/100 refusals and second time I got 0/100 refusals, both had fantastically low KL divergence and both results where abysmal, which got me to spend more time to try to look into what was the issue and after spending many hours testing, then when I tested the original model I found that the issue was the original model itself which is neurotic, engages in Claude's topic switching mechanic, gets stuck in thinking loops easily, guardrails-obsessed, will "answer" questions by writing a bunch of words that don't actually answers anything and instead will throw disclaimers at the user etc. Some other people noticed some of these issues too and opened a comment about it, which was promptly deleted: https://huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled/discussions/30
This resulted in hours of work flushed out, frankly I dunno if I want to spend more time on Jackrong's stuff.
Yesterday, I discovered two models which were also created based on the Jackrong. The results were quite good, but the author deleted them. I made a copy of them:
CCSSNE/LuffyTheFox-Qwopus3.5-27B-v3-Uncensored-FernflowerAI-KL-ReLU-GGUF
CCSSNE/LuffyTheFox-Qwopus3.5-27B-Uncensored-Kullback-Leibler-GGUF
Yesterday, I discovered two models which were also created based on the Jackrong. The results were quite good, but the author deleted them. I made a copy of them:
Well, why did the author deleted them? Could it be that there is an issue with these models?
I'll go and ask him. He deleted 90% of the model all at once yesterday.
I'll go and ask him. He deleted 90% of the model all at once yesterday.
No it's fine, but to uncensor I need the original BF16 Safetensors, I noticed on your repo that you only have up to Q8_0, I can not uncensor GGUF and Quantized models (Quantization decreases a model accuracy).
Do you have the original BF16 Safetensors?
LuffyTheFox mentioned that according to his workflow, the output is usually directly in GGUF format. The two models in my repo have already been uncensored to demonstrate his methodology. Also, Jackrong isn't the only one with a Claude-tuned version—TeichAI has one too. Essentially, one approach is to follow LuffyTheFox’s methodology. On the other hand, if Jackrong’s model has issues, might be worth trying TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 instead.
LuffyTheFox mentioned that according to his workflow, the output is usually directly in GGUF format. The two models in my repo have already been uncensored to demonstrate his methodology.
Oh, so you don't need me to uncensor these models then? What model do you need to be uncensored then?
LuffyTheFox mentioned that according to his workflow, the output is usually directly in GGUF format. The two models in my repo have already been uncensored to demonstrate his methodology.
Oh, so you don't need me to uncensor these models then? What model do you need to be uncensored then?
TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 or Jackrong/Gemopus-4-31B-it
Have you had time to test any of these? if so, which one is the best?
Have you had time to test any of these? if so, which one is the best?
TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 is the most effective one I've tested so far. Compared to the V1 and Jackrong's versions, its performance in coding tasks is much more impressive.
That is very helpful, thank you. I was looking to pick a model for reasoning-specific tasks, including coding, and wondered between those two. Saves me a lot of time!
Have you had time to test any of these? if so, which one is the best?
TeichAI/gemma-4-31B-it-Claude-Opus-Distill-v2 is the most effective one I've tested so far. Compared to the V1 and Jackrong's versions, its performance in coding tasks is much more impressive.
Okay thanks, I will try to see what I can do.
Yeah sorry but it doesn't look like it's gonna be possible, it's much slower than Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled and this one was already super slow, probably due to the higher amount of parameters.