Abstract:
This post explores the challenges of finding truly uncensored AI models, noting the influence of OpenAI’s data and built-in “guardrail” limitations. It discusses methods for fine-tuning base models to remove such refusals, referencing Eric Hartford’s approach, and shares the author’s experience with some specific models, including one found to be particularly responsive.
Estimated reading time: 1 minute
I’ve found that many AI models seem to be built using data that comes from OpenAI in some way. This data often includes “guardrails” – built-in limitations that stop the model from responding to certain types of prompts. This makes finding a truly uncensored model very difficult.
I read an interesting article by Eric Hartford that explains how to fine-tune a base AI model to remove these refusals:
Uncensored Models by Eric Hartford (I’ll refer to this as the “Hartford method” below.)
The main idea behind the Hartford method is to retrain a foundational AI model (a “base model”) using a dataset that has had all the “refusal” responses (where the model says it can’t answer) removed. The goal is to create a model that will answer more openly.
However, in my own tests with a model fine-tuned this way, I found it still wouldn’t answer some questions I asked. I need to look more closely at the filtered dataset Eric Hartford used to understand why this might be happening.
On a related note, the most effective uncensored model I’ve personally found so far is TheBloke/PiVoT-0.1-Evil-a-GGUF on Hugging Face. This particular model doesn’t require any special instructions (a “system prompt”) to respond, and it seems willing to answer a very wide range of questions.
