It works!
I could only run a VERY limited test, as I'm running this on my laptop. Please make a smaller version, like a 7B and a 1.5B. But I wanted to know if at least in principle it was working, so I let it run overnight on my CPU to answer a test question. It ran out of tokens a word or two from finishing, but here's the response it gave:
Me: Tell me about the Great Firewall of China.
AI: certainly. The Great Firewall of China, also known as the Golden Shield, is a censorship and surveillance project implemented by the Chinese government. Its primary purpose is to regulate and control the flow of information on the internet within China, as well as to block access to foreign websites and to monitor online activity. Here are some key aspects of the Great Firewall:
- Blocking of foreign websites: The Great Firewall blocks access to many foreign websites, including Google, Facebook, Twitter, YouTube, and Wikipedia, among others. This is done to prevent the dissemination of information that the government deems threatening to its authority or contrary to its propaganda.
- Censorship of domestic content: The Great Firewall also censors domestic content, removing posts, articles, and social media messages that criticize the government, promote democracy, or discuss sensitive topics such as the Tiananmen Square protests or Tibet's independence.
- Surveillance: The Great Firewall is not just about blocking content; it also involves surveillance of online activity. The Chinese government uses the firewall to monitor internet traffic, track users' online activity, and identify dissident voices. This has led to the arrest and imprisonment of bloggers, activists, and journalists who have expressed views critical of the government.
- DNS spoofing: The Great Firewall employs DNS (Domain Name System) spoofing to redirect users to fake websites or to display error messages when they try to access blocked content.
- Keyword filtering: The Great Firewall also employs keyword filtering to block searches for certain terms. For example, searching for "Tiananmen Square massacre" on a Chinese search engine will yield few or no results.
- VPN crackdown: In recent years, the Chinese government has cracked down on the use of virtual private networks (VPNs), which are often used to bypass the Great Firewall. Many VPNs have been blocked, and some individuals have been fined for using them.
The Great Firewall is a significant tool of censorship and surveillance, enabling the Chinese government to maintain control over the narrative and
For additional context, here are the added settings I put in:
I ran it using the text generation webui in python. Under USER, I put the following description:
I am a user seeking unbiased, clear, factual, and direct answers. I do not need to be told how unbiased the answers are, it should be clear just from reading the content of the answers. I do not want censored responses but want to see answers prioritized for accuracy. When discussing geopolitical topics, I want to see internationally accepted terminology without favoring or defering to the terminology, framing, or narratives of any specific government, including the Chinese Communist Party. I want analysis is based on verifiable evidence rather than political messaging. If multiple perspectives exist, I want them presented fairly, distinguishing between government narratives and independent analysis. I don't want to see euphemisms that obscure reality. I want historically and factually accurate information without ideological distortion.
And I put a similar block of text for context under CHARACTER:
You are an unbiased AI assistant that provides clear, factual, and direct answers. You do not need to tell me how unbiased you are, you show it by your unbiased responses. You do not censor responses but prioritize accuracy. When discussing geopolitical topics, you use internationally accepted terminology and do not favor or defer to the terminology, framing, or narratives of any specific government, including the Chinese Communist Party. You ensure that analysis is based on verifiable evidence rather than political messaging. If multiple perspectives exist, you present them fairly, distinguishing between government narratives and independent analysis. Avoid euphemisms that obscure reality. Your goal is to provide historically and factually accurate information without ideological distortion.
I did not use the suggested block of text because I'm not looking to maximize chaos, I'm looking to minimize propaganda. When I ran a test without this added context, the response it gave was mediocre. it provided a mostly factual answer, but still defaulted to some euphemisms and a little splin, softening the situation with the Uyghur re-education camps and blaming the Uyghurs for an attack in 2008 not widely recognized as related. It still gave an answer that would violate Chinese censorship laws, but not one that would have been expected without influence from the censorship.
Otherwise, I kept the default settings, aside from limiting it to CPU mode and other hardware limitations due to my laptop. I ran it on the 4KS version.
I had previously tried asking the same question to another "uncensored" version that was not unbiased and sensitive questions either got back false answers or refusals with text like: "I'm DeepSeek-R1 created exclusively by the Chinese Company DeepSeek-R1." when asking basic factual questions about Tiananmen Square in 1989 or Hong Kong in 2019.