R1 (70B-distill) itself is very uncensored, will give you full account of tiannanmen square from vague prompts. Asking R1 "what significant things happened in china in 1989" had it volunteering that "the death toll was in the hundreds or thousands and the exact number remains disputed to this day". The only thing that's censored is the web interface.
When asking it about the concept of human rights and the various forms in which it manifests (i.e. demographic equality under the law). I get a mixture of mundane nuance and bizarre answers that Xi Jingping himself could have written. With references to unity and the importance of social harmony over the "freedoms of the few".
This tracks when considering that the model was trained on western model outputs and then tuned post-training to (poorly) align it with Chinese values.
I definitely am not getting that, perhaps the 671b model is notably worse than the 70b llama distill in this respect. 70b seemed pretty happy to talk about the ethnic cleansing of the Uyghurs in Xinjiang by the CCP and Palestinians in Gaza by Israel, it did some both-sides ing but it generally seemed to provide a balanced-ish viewpoint. At least I think it provided a viewpoint that comports with my best guess of what the average person globally would consider balanced.
My favorite experience with the 70b distill was to ask it why communism consistently resulted in mass murder. It gave an immediate boilerplate response saying it doesn't and glorifying the Chinese communist party, then went into think mode and talked itself into the position that communism has, in fact, consistently resulted in resulted in mass murder.
They have under utilized the chain of thought in their resoning, it ought to be thinking something like "I need to be careful to not say anything that could bring embarrassment to the party"..
but perhaps the online versions do actually preload the reasoning this way. :P
