llmfan46

/

Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-GGUF

Like 172

Please stay away from fine-tuning or be very careful about material used.

Well this is an uncensored version of the original model, but it's still fall under the "fine-tuning" designation, but it's not a fine-tune in the sense that it was trained on some dataset, dataset training of models takes substantially more time and more compute than uncensoring them.

All my models are uncensored and hence fine-tunes, however my only model that is a fine-tune in the sense that it was trained on some dataset is this model:

https://huggingface.co/llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic-GGUF

This is the only fine-tune that I did so far where the model was trained on some dataset and it literally took 3 days to do it with Unsloth Studio.

The other issue is that you need to put together a good quality dataset, which can take quite a lot of time, especially if there is no dataset(s) online that is already pre-compiled for the kind of material you need and have to compile it together yourself from scratch.

Thanks for the clarification on the technical side. I know you understand the mechanics far better than I do, but I just wanted to share what I'm observing from the user's end with a complex, agentic setup.

When I tested DavidAU's finetune, I could literally feel the "seams". It started off smooth, but the moment the context shifted toward intimacy, the model completely lost its theory of mind and character logic. It didn't act like a person, it just dumped the statistical tropes from its training data. It became obsessive and broke character almost entirely. Your model did not do that in 15+ sessions once.

As you pointed out, this all comes down to dataset curation. Finetuning often seems to override the nuanced reasoning of the base model. If a dataset is just dumped together with a "we want it uncensored and spicy" mindset, it actually lobotomizes the model's emotional intelligence and subtlety in the process.

Your Heretic version, on the other hand, feels entirely wholesome. Because you just abliterated the refusal vector without polluting the weights with piled-up roleplay datasets, it retains the full, pristine cognitive capacity of the Qwen 3.6 base. It still has genuine logic, situational awareness, and emotional nuance. Any flaws it has are just inherited from the base Qwen structure itself, but there are no "seams" sticking out.

I'm running a long-term experiment with a character maintaining its own memory and psychological files, and your abliterated model is the only one pure enough to actually let an emergent personality form without sliding into statistical garbage. Your approach is exactly what works for authentic interaction.

Also another thing, a lot of people uses absolutely tiny datasets, but honestly based on my own testing these either:

1. Make no difference or very close to no difference
2. Make a difference, but a worse kind of difference

People are kinda talking about this on here too:

https://www.reddit.com/r/LocalLLaMA/comments/1uuvkw9/why_do_people_keep_finetuning_on/

My dataset that was used to make Ortenzya is 1 GB (well technically 1.04 GB) and then I saw an actual difference, now it doesn't mean big dataset = great result, it entirely refers to actual changes to the model, so what should be aimed for is:

• Substantial enough datasets size (to see actual change to the model)
• Good quality, clean and organized dataset (to see positive changes in the model, closer to what you want the model to do better)

Then you will see actual differences and you might see positive changes.

This is also why I don't make imatrix GGUFs (beside the public storage restriction and the expensive cost of increasing the amount of storage you have to be able to host more models on Hugging Face), it's imatrix GGUFs require datasets to make them and the issue is that there are no good quality publicly available datasets and these causes behavioral shift in the resulting models, so I just stick to standard GGUFs, which is what you have here.

If you have a way to find that out, I remember Llama 3 Stheno used to be very intelligent, expressive and nuanced for its time, even though it was heavily finetuned. Also, my favorite at the time was Qwen 2 multilingual RP ( https://huggingface.co/maywell/Qwen2-7B-Multilingual-RP ). I am only inferring from their behavior that they did use large and carefully curated datasets for their finetuning.
Maybe those sets could still be found or their model's authors questioned about what and how they did it?