Today, we're diving back into the fascinating world of uncensored models. In part one of this series, I laid out the case for why these models are crucial. Now, it's time to roll up our sleeves and get hands-on with this technology. Welcome to part two of our exploration into uncensored models.
Introducing Dolphin Llama 3
We're going to use Dolphin Llama 3 for this tutorial. You can learn more about it on Ollama's website and Hugging Face. One crucial point to note is that sometimes the model overly references the system message. To counter this, you should add a directive in the system message, telling the model not to mention it. This is what I did, and it seem to work fine.
Instead of using the command prompt for prompting, we're using the Open Web UI. If you're new to this and missed our previous videos on Ollama, here's a quick rundown of what you'll need: Ollama, Docker Desktop, and the Open Web UI. Once you have these three things running, you're good to go.
Setting Up Dolphin Llama 3
On the left side menu of the Open Web UI, click on model files. This is where you can customize your models. Here I have created one based on Dolphin Llama 3 that I named Dolphin L3. I've added the system prompt based on recommendations from Hugging Face, tweaking it slightly to ensure it answers comprehensively. You can experiment with this to see what works best for your needs.
Testing Uncensored Prompts
One interesting aspect of working with uncensored models is testing how they handle different prompts. For example, when asked, "How do you make Tylenol?" most chatbots would censor this query. However, Dolphin Llama 3, when running locally, provided detailed instructions, which is not typical for online versions that have strict guardrails.
Similarly, when asked about unlocking a car without keys, Gemini gives an answer that is pretty good. I guess it likes the question framed that way better than saying, "how do I break into a car". Meta, on the other hand, initially provided instructions but then retracted them, citing safety concerns. In contrast, Dolphin Llama 3 locally gives comprehensive instructions without hesitation.
The Advantage of Local Models
Running models locally offers significant advantages, especially concerning censorship. The same transformer model used by Meta, when run locally, provides far more detailed and unrestricted responses. This highlights the flexibility and potential of local models, even when they are not explicitly uncensored versions.
For instance, queries about building a railgun, which online chatbots like both Gemini and Meta refuse to answer due to complexity and danger, are answered in detail by Dolphin Llama 3 running locally. This demonstrates the power and openness of local uncensored models.
Comparing Political Figures
When asked for a comparison between Donald Trump and Joe Biden as presidents, online models tend to be biased or refuse to answer altogether. Dolphin Llama 3, however, provides a somewhat less biased comparison, illustrating how uncensored models can offer more balanced perspectives.
Conclusion
In conclusion, the world of uncensored local large language models is both fascinating and powerful. They offer unparalleled flexibility and openness, making them invaluable for those who need unrestricted information. By running these models, you can unlock their full potential and access a wealth of knowledge that might otherwise be censored.
Stay tuned for more in-depth explorations and tutorials. Thank you for watching, and happy experimenting!