Abliterated vs uncensored LLMs: which one still refuses?

Abliterated vs uncensored LLMs: which one still refuses?

People use "uncensored" for any model that refuses less, but two very different things hide under the word. MagmaRouter tags them separately because they are not the same promise.

Uncensored: trained to refuse less

An uncensored model is usually a finetune whose training data was chosen to reduce refusals and drop the stock assistant persona. It still has alignment; it just has less of it. It can and does still decline, especially on the most extreme prompts. Dolphin, Euryale, Magnum and MythoMax are well known examples of this kind.

Abliterated: refusal removed

An abliterated model has had the refusal direction subtracted from its weights, so it almost never refuses at all. This is a mechanical edit, not a training choice.

Side by side

UncensoredAbliterated
HowFinetuned on data chosen to stop refusalsRefusal direction projected out of the weights
Training data neededYes, a finetuning setNo, a few hundred prompts to find the direction
Still refusesSometimes, most often on the most extreme promptsRarely
Cost to capabilityDepends on the finetune; can even improve styleSmall, mostly reasoning; often recovered by a retune
Name to look foruncensored, or a finetune name (Dolphin, Euryale, Hermes)-abliterated, or heretic

How MagmaRouter tags them

MagmaRouter marks a model abliterated only when its refusal direction was removed from the weights, and uncensored when the refusal training was taken off some other way. Every model on the current line is one or the other, and all of them call tools. The models page shows the current line and abliterated models lists the abliterated ones.

Neither is a guarantee Uncensored can still refuse. Abliterated rarely refuses but can still be wrong or produce content that is illegal to use. The tag is a description, never a promise about a given answer.

FAQ

What is the difference between abliterated and uncensored?

How the refusals were removed. Uncensored means a model was finetuned on data chosen to stop it refusing; abliterated means the refusal direction was projected out of its weights. Abliterated refuses less often and costs a little capability; uncensored keeps its quality but can still decline.

Which is better, abliterated or uncensored?

Neither in general. Abliterated is the stronger guarantee against refusals; a good uncensored finetune of a stronger base model can be the better model overall. Test both on your own prompts.

Is an abliterated model always uncensored?

In the everyday sense, yes: it refuses very rarely. MagmaRouter tags it abliterated rather than uncensored because the more specific word says what was done.

Related