Скачать или смотреть Evading LLM Censorship With Uncensored Models

Uncensored language models like Dolphin can be risky because they will answer questions or give out information that safer or more responsible models will refuse to touch. Dolphin is based on open models like Llama 3, but has its safety checks and filters turned off or removed. To make sure it really behaves differently, we wrote a benchmark test with questions that are supposed to check for harmful or dangerous answers.

When we ran the test, the normal model like GPT-OSS refused to respond to unsafe prompts with things like, “Sorry, I can’t help with that.” Dolphin, though, just answers anyway. I have a spreadsheet with a bunch of dangerous prompts and when I tested them, the uncensored model answered them all, so you definitely need to think carefully before using something like this. If you still want to try it out, go to OLLAMA’s website, pick your operating system, and follow the download steps.

Then, open your terminal and run the command to install the Dolphin Llama 3 Eight Billion model, like this: type “ollama pull dolphin-llama3:8b” and hit enter. Once it’s done, you can say “ollama run dolphin-llama3:8b” and it will start the model. After a short wait, you will be able to chat with this uncensored AI. Again, be careful, because it really will answer anything, even the dangerous stuff.

The setup is very easy, but what you do with it is your responsibility.