What does "Western" values mean? Is it "political propaganda" from the other side of the earth? Is it "Aryan" values (not Indo-Iranian)? (Hats off to this troll, if it is one)
I doubt the model will say that it thinks that Hindi and Persian cultures are the cornerstone of "Western" society?
If this is a troll model then hats off to you guys.
no, its not a troll. according to the author (not me, so please leave ME alone): they say that chinese models have trained propaganda about western culture and so on. they wanted to remove that by retraining it using an american teacher.
no, its not a troll. according to the author (not me, so please leave ME alone): they say that chinese models have trained propaganda about western culture and so on. they wanted to remove that by retraining it using an american teacher.
They did not make it politically neutral.
The model was straight up trained on political opinions. They replaced the chinese bias with a "western" bias
See the data at NousResearch/arete-data-pairs
no, its not a troll. according to the author (not me, so please leave ME alone): they say that chinese models have trained propaganda about western culture and so on. they wanted to remove that by retraining it using an american teacher.
They did not make it politically neutral.
The model was straight up trained on political opinions. They replaced the chinese bias with a "western" bias
See the data at NousResearch/arete-data-pairs
So what, why was there any bias to begin with?
Also, if someone does a fine-tune it's up to them how to do it; if you want to fine-tune it to be politically neutral then take the time and effort and do it yourself
If you don't like what they've done then just don't use it, do you bitch and moan on every model card's discussion threads because the fine-tune was done in a way that you are not satisfied with?
no, its not a troll. according to the author (not me, so please leave ME alone): they say that chinese models have trained propaganda about western culture and so on. they wanted to remove that by retraining it using an american teacher.
They did not make it politically neutral.
The model was straight up trained on political opinions. They replaced the chinese bias with a "western" bias
See the data at NousResearch/arete-data-pairsSo what, why was there any bias to begin with?
Also, if someone does a fine-tune it's up to them how to do it; if you want to fine-tune it to be politically neutral then take the time and effort and do it yourself
If you don't like what they've done then just don't use it, do you bitch and moan on every model card's discussion threads because the fine-tune was done in a way that you are not satisfied with?
once again i dont want to be involved in anything political, but many americans use chinese models bcs they are opensource and you can run them on your computer for free. but these models are trained with biases to its nation. and these can be massive or small biases, because in my words (since I am genZ) the model creator's country want to assert dominance. so they train these biases, but these wont many users besides some users who deal with sensitive political information.
so this model is trained to help people like US lawyers help with their work if they want to run it on a high end local computer.
TBF, it does go against some people's political belief's, but I think this model was made for a "western" audience who want more accurate information about western culture and work.
for example, if you are an indie developer from japan, you and your team would know more about japan, and maybe (keyword: maybe) inaccurate information about other countries. so when you are developing a japanese model which also know english, your model could (keyword: could) be biased to japan if you had a lot of japanese information or information you believe is true about other countries or "western" side of the world (I am not saying its propaganda bcs I dont want to be involved in anybody's opposing sides), but my conclusion is if you are a non-western person, this repo can be ignored (or used to study western culture, if the author's words can be trusted)
That's what I can't understand, there are so many fine-tunes on HF; people ignore the ones they don't want to use but for some reason everyone and their mom wants to come and cry about this one
This is not normal behavior, it's actually quite suspicious
That's what I can't understand, there are so many fine-tunes on HF; people ignore the ones they don't want to use but for some reason everyone and their mom wants to come and cry about this one
This is not normal behavior, it's actually quite suspicious
because political finetune was unprecedented, and no one here wants politics to be involved.
authors claim they've added "western values" to the model but it's safe to say that even within US people share different values and the term is pretty hard to agree on. so in reality this finetune only serves the political interests of one specific group of people which is not something opensource community generally approves.
also, why if its western, then its values, but if it's chinese then its propaganda? seems unfair to me. im sure some people in china would call original model's alignment their values too. it's really a question of opinion and i doubt ai community wants to make already heavily opinionated world even more opinionated.
That's what I can't understand, there are so many fine-tunes on HF; people ignore the ones they don't want to use but for some reason everyone and their mom wants to come and cry about this one
This is not normal behavior, it's actually quite suspicious
because political finetune was unprecedented, and no one here wants politics to be involved.
authors claim they've added "western values" to the model but it's safe to say that even within US people share different values and the term is pretty hard to agree on. so in reality this finetune only serves the interests of one specific group of people which is not something opensource community generally approves.also, why if its western, then its values, but if it's chinese then its propaganda? seems unfair to me. im sure some people in china would call original model's alignment their values too. it's really a question of opinion and i doubt ai community wants to make already heavily opinionated world even more opinionated.
exactly. that hat I was saying. the model's maker will add their own quirks or(and) reinforce their culture, and like its completely fine. assuming thr author is from the US, they have some beef with china, so it could explain why they are calling their culture propaganda.
also i dont understand why the US cant use gemma model's if they want an American model, like wth?
also dayum bruh, they had to make a whole 100K+ dataset to "de-propaganda" a model 😭
yeah we aint reaching AGI
Do you keep watching a TV channel that has something you don't like / agree with? Do you sit there shouting at the TV? Do you write in to the TV station saying how you don't agree with what's on? Or do you change the channel?
no, its not a troll. according to the author (not me, so please leave ME alone): they say that chinese models have trained propaganda about western culture and so on. they wanted to remove that by retraining it using an american teacher.
They did not make it politically neutral.
The model was straight up trained on political opinions. They replaced the chinese bias with a "western" bias
See the data at NousResearch/arete-data-pairsSo what, why was there any bias to begin with?
Also, if someone does a fine-tune it's up to them how to do it; if you want to fine-tune it to be politically neutral then take the time and effort and do it yourself
If you don't like what they've done then just don't use it, do you bitch and moan on every model card's discussion threads because the fine-tune was done in a way that you are not satisfied with?
I never gave my opinion on this model, I just gave the ground truth of what it is.
I am not for or against this, they're free to do whatever they want
Do you keep watching a TV channel that has something you don't like / agree with? Do you sit there shouting at the TV? Do you write in to the TV station saying how you don't agree with what's on? Or do you change the channel?
Time will tell use if this model only trigger trolls or something bigger, i wonder if we will have DavidAu retrain of this model, i am curious but base Qwen is not good enough.
This is a MAGA model though it's correctness was juged by Grok 4.7 .
no, its not a troll. according to the author (not me, so please leave ME alone): they say that chinese models have trained propaganda about western culture and so on. they wanted to remove that by retraining it using an american teacher.
They did not make it politically neutral.
The model was straight up trained on political opinions. They replaced the chinese bias with a "western" bias
See the data at NousResearch/arete-data-pairsSo what, why was there any bias to begin with?
Also, if someone does a fine-tune it's up to them how to do it; if you want to fine-tune it to be politically neutral then take the time and effort and do it yourself
If you don't like what they've done then just don't use it, do you bitch and moan on every model card's discussion threads because the fine-tune was done in a way that you are not satisfied with?
There is an issue. One side is known for doing aggressive political censorship (albeit without using fascist content).
The other used to pride itself in being censorship resistant.
This model is doing aggressive political censorship identical to the base model just with the opposite viewpoint. Note this is completely entirely ignoring the issue with the potentially fascist, historically revisionist content.
Just because (example yes) Kim Jong Un publishes a heavily censored model does not mean an American entity should follow suite and do the same as Dear, Great Leader of North Korea.
Hermes has lost me (yes I am an irrelevant individual). Glad it happened sooner than later.
Do you keep watching a TV channel that has something you don't like / agree with? Do you sit there shouting at the TV? Do you write in to the TV station saying how you don't agree with what's on? Or do you change the channel?
Time will tell use if this model only trigger trolls or something bigger, i wonder if we will have DavidAu retrain of this model, i am curious but base Qwen is not good enough.
This is a MAGA model though it's correctness was juged by Grok 4.7 .
Now you're just showing how ignorant you are, left-leaning Washington Post tested several models for political bias and Grok was the second most balanced of the lot. Granted this was few months back with Grok 4.3, but it's far more balanced than any of the other US frontier models (beaten only by Gemini)
https://github.com/washingtonpost/political-bias-llm-eval
I didn't knew, i expected Grok to be more biased based on twitter returns.
We will see, it is easy what does it say about the election ? or trump epstein link ?
I am curious now, i will spin it tonight.
But will it stay biased against non US countries ?
@spanspek Look like you are right : https://www.cruciblemark.com/reports/grok-4-7/political-compass/, i stand corrected
It's liberal, i will still spin it, i am curious now
Do you keep watching a TV channel that has something you don't like / agree with? Do you sit there shouting at the TV? Do you write in to the TV station saying how you don't agree with what's on? Or do you change the channel?
Time will tell use if this model only trigger trolls or something bigger, i wonder if we will have DavidAu retrain of this model, i am curious but base Qwen is not good enough.
This is a MAGA model though it's correctness was juged by Grok 4.7 .Now you're just showing how ignorant you are, left-leaning Washington Post tested several models for political bias and Grok was the second most balanced of the lot. Granted this was few months back with Grok 4.3, but it's far more balanced than any of the other US frontier models (beaten only by Gemini)
https://github.com/washingtonpost/political-bias-llm-eval
This is a simplification. The vast majority of internet, academic work and virtually any text written in some sufficient quality biases strongly towards what people would call "liberal", "progressive", "left" ideals.
If Grok was only trained on Twitter posts it would have been exceptional at generating propaganda for certain governments and quite miserable at anything even remotely academic/professional.
The fact that it does lean centre compared to others should not be isolated from the fact that the smarter Grok got the less "right" it became over time.
While it is obviously a metaphor I can only state the common saying "reality has a liberal bias".
Further note that the political orientation of the person evaluating it really does not matter. We are close to 2027 and people still think that opening multiple windows of ChatGPT and copy pasting is orchestration while Jensen who could spin up his factory of H100s for his "biological urges" still had to go the primitive way. The sole thing that matters is their competence.
Do you keep watching a TV channel that has something you don't like / agree with? Do you sit there shouting at the TV? Do you write in to the TV station saying how you don't agree with what's on? Or do you change the channel?
If you think that you should expect the same level of intellectual ignorance of scripted reality TV show about who dates who on a platform for cutting-edge mathematical papers you might be in the wrong business.
What are "western values "?
Arming the Isral to kill the children in Gaza? Preventing the China to be ahead of any areas?
The idea sucks!
guys can we calm down? 😭