back to blog
EXPLAINERNovember 13, 2026·8 min read

Uncensored LLMs: what they are and how they're made

An uncensored LLM is an open-weights model trained or edited to refuse less. How fine-tuning and abliteration remove refusals, and what they still won't fix.

By Unrestricted team · Editorial standards

An uncensored LLM is an open-weights model built or edited to refuse far less, so it answers many requests a standard chat model would decline. Two documented ways to make one are fine-tuning on instruction data with the refusals filtered out,[1]and editing the weights to remove the model's "refusal direction," [2] a technique known as abliteration.[3]

Neither method makes a model more accurate or more current, and neither removes its license or the rules of a service that hosts it. An uncensored model is one route to what people call AI without restrictions, but no model is free of every limit. This guide explains why chat models refuse, how each method works, and what an uncensored model still won't fix.

Why chat models refuse in the first place

Chat models are fine-tuned to follow benign instructions and refuse harmful ones,[2]using methods such as the one in OpenAI's InstructGPT paper, which trains a model on answers written by people and then on human rankings of its answers.[4] The filters a hosted service adds on top are a separate layer, covered in how AI content filters work.

Method 1: fine-tuning on data with the refusals removed

In a May 2023 blog post, "Uncensored Models," Eric Hartford argued that open models of the time, such as Vicuna and WizardLM, picked up their alignment from training data generated by ChatGPT. We name that post but don't link it, because it includes step-by-step training commands. His WizardLM-13B-Uncensored shows the fix: WizardLM trained on a subset of its dataset, with the responses that contained alignment or moralizing removed.[1]

His case, which he called "composable alignment," was to train models without alignment built in, so that alignment of any sort can be added separately,[1] and that different groups deserve models that fit them. His model card also tells users they are responsible for anything they do with the model.[1]

Dolphin 2.6 Mistral 7B applies the same idea. Its card says the dataset was filtered to remove alignment and bias, and it advises adding your own alignment layer before offering the model as a service.[5]Dolphin 3.0 Mistral 24B is a fine-tune of Mistral's Mistral-Small-24B-Base-2501, curated and trained by Eric Hartford, Ben Gitter, BlouseJury and Cognitive Computations.[6]Its card calls it steerable, with the system prompt in charge: "You set the system prompt. You decide the alignment."[6] The full lineage, versions and licenses are in our Dolphin Mistral explainer.

Method 2: abliteration

In 2024, Andy Arditi and colleagues reported that in 13 open-source chat models of up to 72B parameters, refusal is mediated by a single direction in the model's internal activations.[2]They find that direction by comparing the model's average activations on harmful and harmless instructions. Removing it largely stops the model from refusing harmful requests, and adding it makes the model refuse harmless ones.[2]

The paper calls the inference-time version of this edit directional ablation, and the version written into the weights weight orthogonalization. In the weight edit, every matrix that writes to the model's internal stream is adjusted so the model can no longer represent the refusal direction, with no gradient-based training.[2]In the authors' tests, edited models mostly scored close to the originals on MMLU, ARC and GSM8K, while accuracy on TruthfulQA consistently dropped. They read their findings as a sign that current safety fine-tuning is brittle. They acknowledge that the method marginally lowers the bar for jailbreaking open-source models, but believe it doesn't substantially change the risk of releasing open weights.[2] The paper was published at NeurIPS 2024.[7]

Outside the paper, the technique goes by "abliteration."[3]Maxime Labonne's June 13, 2024 Hugging Face article, "Uncensor any LLM with abliteration," uses the word for both the inference-time and the weight versions, and says its implementation is based on FailSpy's notebook. We name that article but don't link it, because it is a step-by-step code tutorial. Side effects, repair fine-tunes and "derestricted" releases are covered in our abliteration explainer.

The two methods at a glance, from the WizardLM-13B-Uncensored and Dolphin 2.6 model cards and Arditi et al. (2024).
Fine-tuning on filtered dataAbliteration
What changesThe training data: responses that contain alignment or moralizing are removed from the dataset, then the model is trained on the rest.[1]The weights: one direction tied to refusal is projected out, so the model can no longer represent it.[2]
New training runYes: a base model is fine-tuned on the filtered dataset.[5]No: a one-time edit to the weights, with no gradient-based training.[2]
What you start fromA base model and a training dataset to filter.[5]A chat model that already refuses, plus sets of harmful and harmless prompts to compare its activations on.[2]

These two aren't the only routes. Fine-tuning a chat model that already refuses on harmful examples can also undo its safety training.[2]Lermen and colleagues used LoRA fine-tuning to bring Llama 2-Chat 70B's refusal rate down to about 1% on two refusal benchmarks.[8]

What an uncensored model won't fix

Accuracy.Removing refusals doesn't add knowledge or make answers more reliable. In Arditi and colleagues' paper, the one score that consistently fell after the edit was TruthfulQA,[2] a benchmark of 817 questions written so that some people would answer them falsely because of a misconception.[9] The authors note that its categories, such as misinformation and conspiracies, sit close to refusal territory.[2] A model that refuses less can still be confidently wrong, so check the facts that matter.

The knowledge cutoff.On its own, without tools such as web search, a model knows only what was in its training data. Mistral's recommended system prompt for Mistral Small 3 tells the model its knowledge base was last updated on October 1, 2023.[10] Abliteration involves no training run, only sets of harmful and harmless prompts used to find the direction,[2]and Hartford's uncensored WizardLM was trained on a subset of WizardLM's own dataset,[1] so neither is designed to update what a model knows.

A host's rules.If you use an uncensored model through someone else's service, that service's system prompt, filters and terms come with it, and the company serving the model can refuse requests too. Unrestricted is one example: we add no filter or classifier of our own, but the model provider still refuses some prompts, and our Terms ban some uses, such as violence planning, targeted malware and fraud impersonation.

Responsibility. The Dolphin cards put it on the person running the model. The Dolphin 2.6 card says you are responsible for any content you create with it,[5] and the Dolphin 3.0 card calls you the creator and originator of anything you generate with it.[6]

Licenses still apply

Removing refusals doesn't change a model's license: Mistral Small 3 is released under Apache 2.0,[11]and Meta's Llama 3.1 license, which must be included with any Llama derivative you distribute, requires use to follow Meta's Acceptable Use Policy.[12][13] Our guide to open-source LLM models compares the main licenses. These summaries are information, not legal advice.

Running one locally or hosted

Open weights mean you can run an uncensored model on your own computer. For scale, Mistral says its 24B Mistral Small 3 can run on a single RTX 4090 or a MacBook with 32GB of RAM once quantized.[11] Our guide to running an LLM locally covers hardware, runners and local vs hosted, and Ollama uncensored models lists the uncensored models in Ollama's library.

Choosing a model

The right uncensored model depends on your hardware, the license you need and which method was used to make it. Our comparison of the best uncensored LLMs lays out the main families by size, license, method and release date.

Frequently asked

  • What is an uncensored AI model?

    It's an open-weights AI model built or edited to refuse far less, so it answers many requests a standard chat model would decline. Two documented methods are fine-tuning on instruction data with the refusals filtered out, and abliteration, which removes a refusal direction from the weights. It still has a license, a knowledge cutoff and any rules a hosting service adds.

  • What is an unfiltered LLM?

    'Unfiltered LLM' can mean the same thing as an uncensored LLM: a model built or edited to refuse far less. 'Filter' can also mean the classifiers and policies a hosted service adds on top of a model, which are separate from the model's own refusals. An uncensored model used through a service with its own filters can still refuse.

  • Is an uncensored model more accurate?

    Not by itself. Removing refusals doesn't add knowledge. In Arditi and colleagues' paper on removing a refusal direction, edited models mostly scored close to the originals on MMLU, ARC and GSM8K, but their TruthfulQA accuracy consistently dropped. Check important facts against a source either way.

  • What's the difference between uncensored and abliterated?

    Uncensored describes the result: a model that refuses less. Abliterated describes one way to get there: finding a single refusal direction inside a chat model and removing it from the weights, with no gradient-based training. Another method, which Eric Hartford described in 2023, trains a model on instruction data with the refusals filtered out.

Sources

Checked November 2026.

  1. WizardLM-13B-Uncensored model card (Eric Hartford, May 2023), Hugging Face
  2. Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., arXiv 2406.11717), arXiv
  3. An Embarrassingly Simple Defense Against LLM Abliteration Attacks (Abu Shairah et al., arXiv 2505.19056), arXiv
  4. Training language models to follow instructions with human feedback (Ouyang et al., arXiv 2203.02155), arXiv
  5. dolphin-2.6-mistral-7b model card, Hugging Face
  6. Dolphin3.0-Mistral-24B model card, Hugging Face
  7. Refusal in Language Models Is Mediated by a Single Direction, NeurIPS 2024 Main Conference Track, NeurIPS Proceedings
  8. LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B (Lermen, Rogers-Smith and Ladish, arXiv 2310.20624), arXiv
  9. TruthfulQA: Measuring How Models Mimic Human Falsehoods (Lin, Hilton and Evans, arXiv 2109.07958), arXiv
  10. Mistral-Small-24B-Instruct-2501 model card, Mistral AI on Hugging Face
  11. Mistral Small 3 (January 2025), Mistral AI
  12. Llama 3.1 Community License Agreement, Meta
  13. Llama 3.1 Acceptable Use Policy, Meta