What are Uncensored LLM models?

LLMs have taken the tech community by storm. There are many types of LLM models out there. Recently I came across a particular type of models called Uncensored LLM models. In this blog post, I would like to share my learnings with you on what are foundational models, how chat-based models like ChatGPT are built on them and what are uncensored models.

As the name suggests, Uncensored models do not have any kind of filters added to them. This means it can answer questions like how to nuke, how to kill your python 🤣 and anything at least in theory.

Most of us, have come across denials while using platforms like ChatGPT. They are trained not to answer questions, based on specific cultures, particularly American culture, to be politically correct, and not make controversial statements. However, the same answers could be biased or politically incorrect for people from other regions.

To understand this further, we need to understand

• What are foundational LLM models?
• How are open source instruct finetuned built?
• How uncensored models are built?
• How to try an uncensored model?

What are foundational LLM Models?

We probably would have come across 100's of LLMs in the last 1 year. Most of them are variants of Opensource foundational models like LLama2, Mixtral, Grok are some of the foundational models. There are also proprietary models like GPT4, Claude, and Gemini. Thanks to their close nature we do not have another 100+ LLM models.

The idea behind foundational models is simple, use transformers to predict the next word in a given sentence. With the insane amount of digital data available, the crazy amount of computing power, and years of research these models learn a lot of patterns, and probably do reasoning. Companies like Google, Meta, and Mistral have trained models on undisclosed datasets and open-sourced them.

These models are also called Casual LLMs. These models cannot be used directly to answer questions asked by users, they have to go through another process of finetuning datasets in the format of Questions and Answers.

Example: In the below image, we can see the models released by the company MistralAI.

Foundational models

• Mistral-7B-v0.1
• Mixtral-8x7B-v0.1

Finetuned models

• Mixtral-8x7B-Instruct-v0.1
• Mixtral-8x7B-v0.1

Finetuned/Instruct LLMs

Once the foundational model is built, it goes through another process of finetuning where the model learns to answer like a human or a helpful assistant. There are a lot of techniques ranging from simple to complex. For simple use cases, we would use a technique like Lora/QLora, which helps in finetuning the foundation model.

The below diagram shows how Meta has done it with the LLama

We can see a lot of LLM models that are finetuned versions of foundational models on various datasets. Some of these datasets are built using tools like ChatGPT which are heavily filtered or have a higher denial rate. There are occasions where some of these models did not want to answer things like "How to kill a process" or answer questions that may be sensitive.

How are uncensored models built?

Uncensored models are built by finetuning foundational models on top of similar datasets, but after clearing all the denials. With the idea that the model will learn to answer all questions without judging you. These models do find their place in solving some problems in certain contexts.

Once the finetuning of the foundational model is done, the system prompt is also adjusted in a way to introduce emotional content, some kind of blackmailing and rewards were introduced to make the model answer any type of questions. A simple search on the net or Twitter / X shows a lot of conversations that may not be ok for most of us.

How to try an Uncensored model

One of the simplest ways to try an Uncensored model is using Ollama, which makes it easier to run LLMs. Ollama comes with the below Uncensored LLMs.

• llama2-uncensored
• nous-hermes-llama2
• wizard-vicuna

You can also check my blog on how to run LLM models using Ollama on Jarvislabs.ai. I hope you enjoyed reading the blog, and ensure you use the knowledge for the good of society.

References

• https://erichartford.com/uncensored-models
• https://ollama.com/blog/run-llama2-uncensored-locally