Reddit user proposes uncensoring LoRAs instead of separate model versions
A Reddit user proposed distributing abliteration and uncensoring LoRAs instead of publishing separate uncensored versions of models. The author said a LoRA could be applied to models sharing the same architecture and to compatible quantizations, suggesting a repository with one LoRA per model as an easier system to rank.
The proposal offers a more modular way to distribute model behavior changes and could simplify organizing compatible uncensoring adapters, though its technical claims were not independently tested.
On huggingface there are millions of "models" which are uncensored versions of other models. I find it a huge waste of space. It would be better to post the LoRA of the abliteration/uncensorship.
The advantage: a lora can be applied to any model of the same architecture. And to any quant.
An example: You can apply THIS to any Qwen3.6-35B-A3B version or quant.
I personally think if would be great to have a repository with one lora for each model.
Perhaps we could then rank them more easily.
What do you think?