Comment by chmod775

2 days ago

> there's an uncensored model that you can run locally with llama.cpp

Correction: There's tens of thousands of them. They're easy to create, which is why everyone publishes their own.

Just put "uncensored", "abliterated", or "heretic" into search on huggingface/ollama/etc and pick any them. Fair warning: most aren't very good, essentially lobotomized, and totally broken if you enable thinking.

9 comments

chmod775

Reply
Scaled  2 days ago

In my own tests, the abliterated models perform equivalently to the same version in an apples-to-apples comparison (if you compare same quantization). Thinking is working also. The main difference is I don't get annoying prompt refusals (otherwise common due to my work on 18+ related projects). However, it's local quantized models so they're not anywhere near frontier quality.

desterothx  2 days ago

interesting, I haven't played with any of them yet, but i thought the point of orthogonalizing the weights towards the restriction vector was that there is no loss in capability while removing guardrails. Does it affect other parts of the RL alignment too?

  • embedding-shape  2 days ago

    > the point of orthogonalizing the weights towards the restriction vector was that there is no loss in capability while removing guardrails

    That is surely the point, most of the "uncensored" weights released for free on HuggingFace aren't being very successful at this. There is a stark difference in output quality between the official weights and all these "uncensored" variants that appears days afterwards.

  • fy20  2 days ago

    The claims by the creators are it doesn't in a major way. I have a uncensored Gemma 4 I run on my Mac. Just for testing out, I haven't found any need for it... yet.

larodi  2 days ago

The name for it is ablation - precise removal of parts of the model. Not abliteration as it is not obliteration.

Even as I write this the ‘abliterated’ word is denoted a typo. Does it not at your end?