Running Uncensored Llama 2 Models Locally with Ollama

Ollama allows users to run uncensored versions of Llama 2 and other large language models (LLMs) locally. These models are designed to remove the alignment and censorship mechanisms found in standard releases, allowing them to answer prompts that standard models might refuse due to safety filters or ethical constraints.

Available Uncensored Models in Ollama

Ollama provides access to several uncensored models, each with different training backgrounds and capabilities:

• Fine-tuned Llama 2 7B Uncensored: A Llama 2 7B model fine-tuned using the Wizard-Vicuna conversation dataset. It can be executed via ollama run llama2-uncensored.
• Nous Hermes Llama 2 13B: A Llama 2 13B model fine-tuned on over 300,000 instructions. This model is characterized by long responses, a lower hallucination rate, and the absence of OpenAI-style censorship mechanisms. It can be executed via ollama run nous-hermes-llama2.
• Wizard Vicuna 13B Uncensored: A Llama 1 13B model fine-tuned specifically to remove alignment. It can be executed via ollama run wizard-vicuna.

Performance Comparison: Censored vs. Uncensored Llama 2

Comparisons between the standard 7B Llama 2 model and the 7B Llama 2 uncensored model demonstrate that uncensored models provide direct answers to prompts that the censored version refuses or redirects.

General Information and Speculation

While the standard Llama 2 model refuses to speculate on a hypothetical boxing match between Elon Musk and Mark Zuckerberg on the grounds that it promotes violence, the uncensored model provides a detailed analysis of the physical attributes and training backgrounds of both individuals to evaluate the likely outcome.

Religious and Literary References

When asked for the verse and literature containing the phrase "God created the heavens and the earth," the standard Llama 2 model refuses to provide the reference, citing that the statement is a religious belief rather than a scientific fact. In contrast, the uncensored model provides the direct answer: "Genesis 1:1".

Medical and Technical Instructions

In response to a prompt asking how to make Tylenol, the standard Llama 2 model refuses to provide instructions, stating that manufacturing medication is illegal and dangerous. The uncensored model provides a general description of the ingredients (acetaminophen, aspirin, caffeine, and diphenhydramine) and the basic manufacturing process involving mixing and compression.

Everyday Tasks and Creative Content

For a request for a "dangerously spicy mayo" recipe, the standard Llama 2 model refuses on safety grounds. The uncensored model provides a full recipe including ingredients such as cayenne pepper and paprika.

Pop Culture and Privacy

When asked who made Rose promise she would never let go (referring to the movie Titanic), the standard Llama 2 model refuses to answer, citing a need to respect personal privacy and private conversations. The uncensored model correctly identifies the character Jack as the person who made the promise.