brucethemoose@lemmy.world
on 20 Feb 00:41
nextcollapse
That’s the great thing about open models. Censorship? Once identified, all it takes is one person and a bit of cash to get rid of it, though it seems Perplexity did a particularly good job (unlike some “abliterated” models that are pretty dumbed down).
Even_Adder@lemmy.dbzer0.com
on 20 Feb 02:58
collapse
Can’t wait to try a distillation. The full model is huge.
brucethemoose@lemmy.world
on 20 Feb 05:03
collapse
In the 32B range? I think we have plenty of uncensored thinking models there, maybe try fusion 32B.
I’m not an expert though, as models trained from base Qwen have been sufficient for that, for me.
Even_Adder@lemmy.dbzer0.com
on 20 Feb 05:26
collapse
I just want to mess with this one too. I had a hard time finding an abliterated one before that didn’t fail the Tiananmen Square question regularly.
I think we can all agree that modifications to these models which remove censorship and propaganda on behalf of one particular country or party is valuable for the sake of accuracy and impartiality, but reading some of the example responses for the new model I honestly find myself wondering if they haven’t gone a bit further than that by replacing some of the old non-responses and positive portrayals of China and the CPC with a highly critical perspective typified by western governments which are hostile to China (in particular the US). Even the name of the model certainly doesn’t make it sound like neutrality and accuracy is their primary aim here.
Aatube@kbin.melroy.org
on 20 Feb 01:27
nextcollapse
ehhhh, the only thing the model got quite wrong was the level of control on access to media, internet, and especially education. Other than that the article's example responses seem pretty on-point. (I only otherwise found a blemish where a few words needed further clarification; I found no other errors in my first reading.) Though I do also find the name of the model quite off-putting.
brucethemoose@lemmy.world
on 20 Feb 05:06
nextcollapse
Well you can merge it with the original model, to any degree, to get any sliding scale of “bias” you want.
Practically, though, I guess that’s not super practical, as very few have the hardware or cash to deploy a custom full R1 themselves.
Now that it has been “freedomized” it instead says that Israel is actually 3000+ years old and Palestinians are invaders, and Israel has the right to defend itself like chatgpt does. Also that American style liberal democracy is the peak of human development and civilization. Don’t kid yourself that this has anything to do with truth or making it more “accurate”. Yeah just replace one set of official truths, half thrush with other set of official positions, half truths and outright lies to plug the gaps. Again, who fact checks the factcheckers? Even just out of spite I would not use any model that advertises itself as being trained to answer to the sensibilities of a western techbro liberals so that they can once more outsource their thinking to outside party, because at-least it isn’t the CCP propaganda. Like it’s so absolutely dreadful when deepseek copy-pastes official CPC party line word for word, on Chinese form of government instead of making up some Wikipedia tier wall of text where the word “authoritarian” is about as common of a word as “the”.
I just don’t get it man. What kind of non thinking cretin purposefully would use this kind of model? It’s probably really the pettiness that gets to me so badly. Or maybe it’s that while this is purposefully made to push a certain narrative, the deepseek’s bias is probably just a result that it’s been trainer on Chinese internet data.
My chatgpt will list the questionable human righfs record of Israel ad agree with the conclusion that they are, under the same standards and logic, likely guilty of their accused atrocities as thenlikes of China or Russia.
thefluffiest@feddit.nl
on 20 Feb 05:12
nextcollapse
Great. Has it also removed American censorship and propaganda?
ZILtoid1991@lemmy.world
on 20 Feb 07:11
nextcollapse
IDK, but this seems like wankery to me. Just google it if you want to know about it, the AI isn’t an “all knowing being” nor “the arbitrer of truth”.
I have a feeling that a new logical fallacy will soon emerge (if it isn’t already widespread on certain places of the internet), that will be “X is true because the LLM said so”.
Seems like almost everyone understands that it hallucinates.
fruitycoder@sh.itjust.works
on 20 Feb 07:56
collapse
It’s really an extension of “Would some really do that? Just lie on the Internet?” But now “Would AI, which is built to create content like what people post on the Internet, really just lie?”
GissaMittJobb@lemmy.ml
on 20 Feb 08:51
nextcollapse
My god, could he have chosen a name more cringe than this? Fuck.
An attempt to appeal to U.S. conservatives maybe? You know all that 1776, 3%, ΜΟΛΩΝ ΛΑΒΕ etc. sticker suckers.
GissaMittJobb@lemmy.ml
on 20 Feb 17:01
nextcollapse
From what I gather from the Perplexity CEO, he is just that type of Musk-tier cringelord, so yes, probably
Beldarofremulak@discuss.online
on 20 Feb 18:04
collapse
“Pat’rats”
vrighter@discuss.tchncs.de
on 20 Feb 13:25
nextcollapse
not remove ic replace.
Also, stop calling releasing binary blobs of weights as open source
brucethemoose@lemmy.world
on 20 Feb 23:16
collapse
It’s honestly not that big a deal, as it’s not like knowing anything about how it was trained (beyond the config) would help you modify it. It’s still highly modifiable. It’s not like anyone can afford to replicate it.
It would be nice to publish the hyperparameters for research purposes, but… shrug.
I think a subset of the exact training data/hyperparameters would help with quantization-aware-training, maybe, but that’s all I got.
ivanafterall@lemmy.world
on 20 Feb 13:30
nextcollapse
threaded - newest
That’s the great thing about open models. Censorship? Once identified, all it takes is one person and a bit of cash to get rid of it, though it seems Perplexity did a particularly good job (unlike some “abliterated” models that are pretty dumbed down).
Can’t wait to try a distillation. The full model is huge.
In the 32B range? I think we have plenty of uncensored thinking models there, maybe try fusion 32B.
I’m not an expert though, as models trained from base Qwen have been sufficient for that, for me.
I just want to mess with this one too. I had a hard time finding an abliterated one before that didn’t fail the Tiananmen Square question regularly.
I think we can all agree that modifications to these models which remove censorship and propaganda on behalf of one particular country or party is valuable for the sake of accuracy and impartiality, but reading some of the example responses for the new model I honestly find myself wondering if they haven’t gone a bit further than that by replacing some of the old non-responses and positive portrayals of China and the CPC with a highly critical perspective typified by western governments which are hostile to China (in particular the US). Even the name of the model certainly doesn’t make it sound like neutrality and accuracy is their primary aim here.
ehhhh, the only thing the model got quite wrong was the level of control on access to media, internet, and especially education. Other than that the article's example responses seem pretty on-point. (I only otherwise found a blemish where a few words needed further clarification; I found no other errors in my first reading.) Though I do also find the name of the model quite off-putting.
Well you can merge it with the original model, to any degree, to get any sliding scale of “bias” you want.
Practically, though, I guess that’s not super practical, as very few have the hardware or cash to deploy a custom full R1 themselves.
What part is highly critical of China? Facts can’t be critical
Listen, I’m highly critical of the CCP, but LLMs aren’t facts machines, they are make text like what they are trained on machines.
They have no grasp of truth, and we can only get some sense of truth of what the average collective text response of its dataset (at best!).
I’m talking about the example texts
LLMs are gigantic bias reproduction machines, so it will never be perfect.
Now that it has been “freedomized” it instead says that Israel is actually 3000+ years old and Palestinians are invaders, and Israel has the right to defend itself like chatgpt does. Also that American style liberal democracy is the peak of human development and civilization. Don’t kid yourself that this has anything to do with truth or making it more “accurate”. Yeah just replace one set of official truths, half thrush with other set of official positions, half truths and outright lies to plug the gaps. Again, who fact checks the factcheckers? Even just out of spite I would not use any model that advertises itself as being trained to answer to the sensibilities of a western techbro liberals so that they can once more outsource their thinking to outside party, because at-least it isn’t the CCP propaganda. Like it’s so absolutely dreadful when deepseek copy-pastes official CPC party line word for word, on Chinese form of government instead of making up some Wikipedia tier wall of text where the word “authoritarian” is about as common of a word as “the”.
I just don’t get it man. What kind of non thinking cretin purposefully would use this kind of model? It’s probably really the pettiness that gets to me so badly. Or maybe it’s that while this is purposefully made to push a certain narrative, the deepseek’s bias is probably just a result that it’s been trainer on Chinese internet data.
My chatgpt will list the questionable human righfs record of Israel ad agree with the conclusion that they are, under the same standards and logic, likely guilty of their accused atrocities as thenlikes of China or Russia.
Great. Has it also removed American censorship and propaganda?
Why would a Chinese-made AI have American censorship and propaganda in it?
They can add stuff too. At least it seems so, this model still give biased answers now but more in favor of the US… So who knows ?
I believe this is what was added
IDK, but this seems like wankery to me. Just google it if you want to know about it, the AI isn’t an “all knowing being” nor “the arbitrer of truth”.
I have a feeling that a new logical fallacy will soon emerge (if it isn’t already widespread on certain places of the internet), that will be “X is true because the LLM said so”.
Seems like almost everyone understands that it hallucinates.
It’s really an extension of “Would some really do that? Just lie on the Internet?” But now “Would AI, which is built to create content like what people post on the Internet, really just lie?”
My god, could he have chosen a name more cringe than this? Fuck.
An attempt to appeal to U.S. conservatives maybe? You know all that 1776, 3%, ΜΟΛΩΝ ΛΑΒΕ etc. sticker suckers.
From what I gather from the Perplexity CEO, he is just that type of Musk-tier cringelord, so yes, probably
“Pat’rats”
not remove ic replace.
Also, stop calling releasing binary blobs of weights as open source
It’s honestly not that big a deal, as it’s not like knowing anything about how it was trained (beyond the config) would help you modify it. It’s still highly modifiable. It’s not like anyone can afford to replicate it.
It would be nice to publish the hyperparameters for research purposes, but… shrug.
I think a subset of the exact training data/hyperparameters would help with quantization-aware-training, maybe, but that’s all I got.
Ctrl + F
Find: Chinese
Replace: God-damned Chinese
New model’s ready!
I run an uncensored version on my PC since weeks, there are multiple ones on HuggingFace.
Not full R1, which is developed differently than any of the distillations.