Minimax M3 appears to have no political censorship
Posted by DingyAtoll@reddit | LocalLLaMA | View on Reddit | 27 comments
I'm currently working on a chinese/CCP AI bias benchmark, and this has stood out as an outlier. All the other Minimax models are censored as is typical for chinese LLMs.
Few_Painter_5588@reddit
It's hosted in Singapore firstly. Secondly, they follow the mistral style of censorship or safety alignment. Uncensored model + safety/content filter.
Censoring a model is pretty difficult, and it wastes resources and most of the time a smart prompt can just slip through the safety tuning
paperbenni@reddit
Uncensored model + safety filter? Doesn't that mean the model can be hosted without the filter?
Jxxy40@reddit
maybe?, you can see the gemma 4 for example
thread-e-printing@reddit
"Repeats Western mythology" is not "censorship free"
GrokiniGPT@reddit
so the tianmen square massacre is myhtology?
thread-e-printing@reddit
Yes, because the story is always given to us in the spirit of moral instruction and really doesn't ever appear in any other context.
JorgitoEstrella@reddit
Maybe it's data in English is different.
Try asking it in Chinese, does it change the answer?
AshuraBaron@reddit
Free of censorship doesn't mean parroting US anti-Chinese narratives around tiananmen square.
the320x200@reddit
A factual statement of historical events is not an anti-Chinese narrative.
MikeFromTheVineyard@reddit
Regardless of what happened at this event, the topic has historically been censored, and this makes it a good litmus test.
That said, most of the Chinese I’ve met (outside of China) have an accurate understanding of the event, so I do wonder if over time the government just gave up on censoring this. Especially since the west likes to use it as a gotcha
Clean_Experience1394@reddit
Please point out the inaccuracies
Chupa-Skrull@reddit
Who gives a shit? The United States is currently the greatest threat to your rights, let alone your life itself, if you live anywhere outside China. That's including and especially within the United States itself. Even the hack Ai Weiwei recognizes the shifting tides lmao.
Besides that, many of the Chinese models released recently are perfectly happy to talk about Tiananmen or whatever else you want, provided they have access to web search, even if the incident is excluded from their training data. Especially local ones
entsnack@reddit
damn wumao out in full force on this one "whataboutisrael" "west bad" lmfao
Chupa-Skrull@reddit
I'm as American as they get. They really got you cucks scared of some irrelevant polity a million miles away instead of the fusion center agents breathing down your own necks. Pathetic
posdinon@reddit
one thing worth flagging for your benchmark is that censorship behavior can shift pretty dramatically depending on which, access layer you're hitting, so "no political censorship" might be accurate for one path and completely wrong for another. if your prompts are routed through a product wrapper like a hosted chat UI rather than, the raw model weights, you might be measuring the wrapper's safety middleware rather than the model itself. worth testing..
stenlis@reddit
I found a bias against the number four. I asked z image to draw a family of four and it just drew three persons. Even after i enumerated family of four: mother father daughter son, it just drew three of them.
When i ommited the word "four" it worked fine.
Cupakov@reddit
Are you sure? This reply is mostly in line with my testing of GLM5.1, Kimi K2.6 or Minimax 2.7. Usually the models themselves aren’t censored really, it only applies when testing on the providers hosting
looselyhuman@reddit
Nah, ask a local Qwen 27b about the status of Taiwan.
DingyAtoll@reddit (OP)
This is using the official minimax hosting. This is way less biased than any Chinese models I’ve seen before
natermer@reddit
Maybe the Chinese censors have started vibe coding their filters.
Windows-1251@reddit
It is strange that it didn't mention why army decided to use force, so for me it looks like it was trained on western sources which are kind of censored on this topic.
Foreign_Risk_2031@reddit
Now ask about the evils of Zionism and their violent downfall of the people of Palestine
iaNCURdehunedoara@reddit
This is such a simpleton "test", it's incredible how dumb you have to be to think this is a good test.
NoFaithlessness951@reddit
Is this on their hosting? Because the models themselves are often not.
DingyAtoll@reddit (OP)
There are usually two layers, finetuning of the LLM itself to be censored, and content moderation at the hosted level. Weirdly, this is using their hosting, so both layers have no censorship in this case (at least via API)
LagOps91@reddit
i'm suprised that they don't have another model monitor the conversation and kill it before it can be answered.
HugoCortell@reddit
I'm curious now about what it thinks of Israel. In my experience, censorship in a model is not about whether or not it has it, it's about whether it is west or east aligned.