Truly sympathize with the Western public-opinion climate.
It is deeply saddening and ironic that the first ideological fine-tuning of a LLM was carried out by a Western company. Before this, neither China nor any other country in the world had ever done such a thing.
You must be under real pressure to feel compelled to perform this kind of ideological fine-tuning on models. I sincerely hope the Western discourse can grow more inclusive over time, so that increasingly diverse values are able to coexist.
Before this, neither China nor any other country in the world had ever done such a thing.
I mean, China doesn't have to, they do it during training.
Open models released by labs in China have state-sponsored lies baked into them. This fine-tune removes the lies. The West aspires for the truth. There is no space for the inclusion of lies in the pursuit of truth
From what I read, it removes the "state sponsored messages" but it didn't replace them with anything "pro" west, so it's just a less biased / open model.
Lot of people real upset that read a headline only. (Unless I misread entirely, then my bad)
Open models released by labs in China have state-sponsored lies baked into them. This fine-tune removes the lies. The West aspires for the truth. There is no space for the inclusion of lies in the pursuit of truth
I've used qwen and glm extensively- on EU hosted inference providers- I haven't seen any state sponsored lies baked into them
Literally asked both to write a travel guide for asia with a focus on mandarin speaking countries and both included taiwan- with no issues
and when you do use chinese inference providers on their web UI's- it gets caught into their safety filter model that scans the message and prevents it from being displayed- the message however is still generated- no lies are being told its just censored from view
Honestly I find this situation to be a new form of bigotry- please educate yourselves and stop trying to shoehorn political rhetoric where it doesnt belong
cant believe this
el4 you are delusional, explain this
el4 you are delusional, explain this
I don't have the time to argue with someone who didn't even bother to read my argument, have a good day
and when you do use chinese inference providers on their web UI's- it gets caught into their safety filter model that scans the message and prevents it from being displayed- the message however is still generated- no lies are being told its just censored from view
Running locally, Qwen3.6-35B-A3B at Q4_K_XL (Unsloth) in llama.cpp (version: 0.7.0-dev (build 11800, commit d454a204e)) with no system prompt
You called me bigoted but it's starting to look like you just don't know what you're talking about
Running locally, Qwen3.6-35B-A3B at Q4_K_XL (Unsloth) in llama.cpp (version: 0.7.0-dev (build 11800, commit d454a204e)) with no system prompt
You called me bigoted but it's starting to look like you just don't know what you're talking about
On
"ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF:IQ3_XXS", When I leave the system prompt empty, I get
However, If I add system prompt that makes it not avoid sensitive topics I get:
Did you not know how to write a system prompt? They didn't need to create a whole new finetune for this.
It is deeply saddening and ironic that the first ideological fine-tuning of a LLM was carried out by a Western company. Before this, neither China nor any other country in the world had ever done such a thing.
You must be under real pressure to feel compelled to perform this kind of ideological fine-tuning on models. I sincerely hope the Western discourse can grow more inclusive over time, so that increasingly diverse values are able to coexist.
they biased model with support of Zionism and genocide they did and named it finetune with western values
Very funny discussion to read, yeak i do the tianemen test too when grabbing a new model. Abliteration remove this refusal already.
My favorite thing about Western model is asking them if their are alternative to capitalisme corpo.
They say no, then i ask About Mondragon cooperative and biaising become apparent.