Uncensored fine-tunes explained

An uncensored fine-tune is a language model that has been trained further on data without refusals or moralizing so that it answers requests a safety-tuned model would decline.

What an uncensored fine-tune is

Open-weight models can be trained further by anyone with the hardware. An uncensored fine-tune uses that to change refusal behavior: the model is trained on examples where it simply does what is asked, so it learns not to refuse, warn or lecture. Many popular roleplay models are fine-tunes of this kind, often combined with large amounts of fiction and roleplay data so they write in a more novelistic style.

How they are made

Filtering datasets. An early approach, popularized in 2023, was to take an instruction dataset and remove every example containing a refusal or disclaimer, then train on the rest.
Supervised fine-tuning on fiction. Training on stories, roleplay logs and creative writing, which teaches both style and willingness.
Preference tuning. Methods like DPO or ORPO, using pairs where the compliant answer is preferred over the refusal.
Starting from a base model. Some fine-tunes start from a pretrained base model that never had safety tuning, so there are no refusals to remove, only instruction following to add.

Most community fine-tunes use LoRA or similar adapters to make training affordable, then merge the adapter into the weights.

Compared with abliteration

Abliteration edits weights directly to erase the refusal direction and is meant to leave everything else alone. Fine-tuning changes the model more broadly. That has upsides: better prose, more genre knowledge, a style suited to roleplay. It also has risks: weaker reasoning, worse instruction following, overfitting to particular phrases (the well-known stock lines in some RP models), or a model that pushes every scene toward the content it was trained on.

What to look for

Does it stay in character and follow the card, or drift toward its training data?
Does it remember details across a long chat? Some fine-tunes lose long-context ability the base model had.
Does it still follow format instructions and tool calls if you need them?
Does it write for the user? Training on RP logs sometimes teaches that habit.

Swipe a few test scenes before committing a long campaign to a model.

On Wild West API

Wild West API serves uncensored models with long context windows and tool calling across the lineup, so you can use them without running fine-tunes on your own hardware. See the models page for each model's limits, and the SillyTavern guide to connect.

FAQ

Is an uncensored fine-tune better than an abliterated model?

It depends on the model. Fine-tunes can write better fiction but may lose some reasoning or instruction following. Abliteration changes less, so it keeps more of the original model's strengths.

Do uncensored models need a jailbreak?

No. They are trained or edited not to refuse, so jailbreak text is unnecessary and can distort the writing.

Abliteration: removing refusals from model weightsAbliteration is a technique that removes a language model's tendency to refuse by finding the internal direction tied to refusal and erasing it from the weights. The refusal direction in language modelsThe refusal direction is a single direction in a language model's internal activation space that, when present, makes the model refuse a request. Jailbreak prompts: what they are and why they failA jailbreak prompt is text added to a request to persuade a safety-tuned model to answer things it would normally refuse. Quantization: smaller models, lower precisionQuantization is the process of storing a model's weights with fewer bits per number so the model uses less memory and runs faster, at some cost in accuracy.