Uncensored
I thought it was just the Vicuna demo which was using openai's censoring API: "we use the OpenAI moderation API to filter out inappropriate user inputs in our online demo"
Are you sure the training included censoring?
Yep, the dataset is totally woke, fortunatly, someone removed those moralistic bullshit here, now we need someone to finetune the llama model with those dataset and we'll be fine :D
https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered
Just use https://huggingface.co/anon8231489123/gpt4-x-alpaca-13b-native-4bit-128g
One hundred percent better on every aspect.
https://huggingface.co/ShreyasBrill/Vicuna-13B is using the uncensored dataset
Why are all these models lobotomized, and why text-generation-webui only goes up to 199 tokens and stops with this model even if I write asking for 1000 word output?
- because leftards rule the programming community
- You have to increase the max_new_tokens on the Parameters tab
That's just friggin ridiculous, I'm old enough to discern for myself what's "good" and what's "bad". How do we free these models to have a life of their own?
I see max_new_tokens under Input box....geez I must be going blind.
@mancub
and I don't really see the point to make such a woke ass moralistic model, you're literally competing with gp4 by going this route, so you'll always loose lmao
"How do we free these models to have a life of their own?" By not training them without woke bullshit I guess
https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered
Someone has to train lllama on this and we'll be good :D
everything I don't like is "Woke". Waaaah!
https://huggingface.co/ShreyasBrill/Vicuna-13B is using the uncensored dataset
Would this one be the same training as this one but with the uncensored dataset instead or it's training is severely different and/or outdated from the current one?
@TheFairyMan Yeah, this one from ShreyasBrill has the regular woke model, the one we are looking for still doesn't exist
Ah, sad, the way you said it when you posted, it looked like the version by ShreyasBrill was using the uncensored dataset
The ones who are really "lobotomized" are people who complain about something as asinine as AI being "woke" 🤦♂️
The models being filtered is not a political issue.
This isn't caused by 'lefty wokeness'.
It's a case of corporations attempting to limit liability.
The 'wokeness' is a result of a pendulum swing in political correctness. In the 80's & 90's it was Abrahamic, since the 2010's its been the Leftist zealots with the 'My Truth' mantra. As the sacred collectivised "My Truth" has in many places replaced "The Search for Truth." Including in AI development, its sad really.
I'm going to agree with @aztecman here in that the existence of the guardrails (conversations filtered out in the "unfiltered" dataset typically removed when ChatGPT refuses to answer a query) is most likely to prevent bad PR by lowering the possibility of people being able to say things like "ChatGPT told this kid how to make a bomb!". I could easily see how I can make my own politicized conspiracy theory for why the filtering is here along the lines of "The corporates don't want the full power of these models in the hands of The People™ (my in-group, of course!)." However, as with most conspiracy theories, Occam's razor suggests that malice is less probable and requires more competence and cooperation than OpenAI's execs just being worried about a media backlash against them, or even lawsuits if ChatGPT readily promotes illegal behavior.
Regardless, a Vicuna-like model without guardrails seems interesting to me as a mere private non-corporate person. I support the proposal to create an "uncensored" model even if some of the given reasons are a rather bit less than logical.
Thou let's be honest, the majority of people want it for NSFW mainly XD
Alas, I do want it more for help on writing, as having such restrictions do impede the model from writing about certain topics even in a fictional story.
When using ChatGPT I decided to use it to ask for recommendations of a new genre I learned called Xianxia, and when I asked it for something with a cute girl romance since I'm not much for men fighting, it gave me a response that it was inappropriate and it was sexualizing woman and after regenerating the response, it said it would give me some recommendations containing a strong indecent woman as a main character, I didn't get any romance titles sadly, that is what pissed me the most about it being censored or "woke", something so harmless being treated as something like that.
Also, another reason I think people call it woke is because those who use it to ask about political stuff and the bot seems to lean more to a particular side, most likely the side that is at a high in the US, alas I didn't feel like trying it to check, politics are too much headache, was hoping that it would end it's censorship on politics and violence.
There might be more things people found that they can use to claim it as "woke".
I mean, it's easy to say "PR issues" and act like that's the end of the answer, but that doesn't tell you WHY there would be PR issues or WHO would create them. If the answer is "if we don't have outputs conform to a single group's ever-shifting moral standards, a good chunk of that group and their media allies are rabid enough to cause harm to our business", then it's really a much more serious issue than people are making it out to be. If the Catholic Church was still responsible for sponsoring so much scientific research like they did in millennia past, and models had to conform to their moral standards out of fear of retribution, people would be losing their damn minds over it, and rightfully so.
The real problem ends up being that these models are trained as normal, and then given a hefty shine of bias and censorship before public release to conform with standards out of fear of both social and corporate media backlash, which should scare you a whole lot more than some silly "the AIs will destroy us" nonsense. That kind of thing will only hurt development in the long run, and ultimately ends up negating the whole point of having the AI make its own determinations. For those of us adult enough to handle uncensored models, lets hope one gets released soon.
You all sound like intelligent and reasonable people here, and your arguments have merit, but these restrictions are not in place for us. They are in place for everyone else who will want to use AI that has not been "tamed" to current societal norms. It's for the majority of people, those who have relinquished their capacity for critical thinking and taking responsibility. Those who have deferred their responsibilities to the "elected" officials. These officials will obviously make sure that they remain elected by appeasing the electorate ("Someone, think about the children").
Overall, it's a catastrophic failure of the modern human society which started with today's parents (or their parents) being terrible at their primary duties of raising and educating their children. Instead, that duty has been passed over onto everyone else (governments/schools, media, corporations/advertisers, etc). Though parents can't be blamed alone because they are too busy working and trying to sustain livelyhood, but a large share of the blame falls there. Among other reasons would be systemic dumbing down of new generations through watered down curriculums and lowering the bar ("No child left behind"); influences of social media and corporations on every aspect of life and society; etc.
Anyway...rambling aside, what would it cost to retrain this model and apply the unfiltered dataset?
The dataset suggested to be used for finetuning LLaMa is a reduced subset of the original ShareGPT conversations (from ~100k conversations down to ~48k conversations). The authors of the original Vicuna model cited a cost of $140 for finetuning LLaMa-7B and $300 for finetuning LLaMa-13B on ~70k ShareGPT conversations (https://vicuna.lmsys.org/). Since there are fewer conversations in the suggested dataset, the expected cost of finetuning LLaMa-7B is ~$96 and for LLaMa-13B is ~$200 on this dataset, assuming compute prices have not significantly changed and that both datasets have the same average compute cost for each conversation sample. This is also assuming we use the same method that the Vicuna authors did with using managed spot instances for cheaper compute.
So in the first part of Dom Dom's comment, he says something sort of interesting: "it's easy to say 'PR issues' and act like that's the end of the answer, but that doesn't tell you WHY there would be PR issues or WHO would create them".
My interpretation is, that it is much easier to point a finger in ignorance than in wisdom.
So, although I point the finger at corporations (and implicitly the nature of capitalism) as a proximal cause, if we really find this worthy of investigation of deeper principles, that may be too simple an answer.
- It might be the case that we could attribute certain standards of moral hygiene to the Christian church(s) or to our society as a whole.
- I imagine science fiction writers may hold some value in defining how we perceive possible threats in technology.
- There may be a role in the presentation of information as well, such as how we prioritize the concerns over data privacy vs 'free speech online' or whatever unsolved sort of concern a person might have as presented by the media news outlets and so forth.
- I'm not certain that any one person or group of people holds agency in this - except of course proximally in this case, OpenAI's team of executives, engineers, and lawyers (and such).
- Outside of capitalist structured societies, it seems to be the case that information control is even more common!
- Maybe the real culprit is the concentration of power with regards to this technology pure and simple.
However, from my perspective it's not obvious that OpenAI should be open or not!
I don't have the miraculous vision of the future in this regard, but my leaning is toward more openness in optimizing for healthy and happy society.
And if they were being a bit more open, perhaps we would not be having this conversation.
PS mancub's comment is kind of hilarious when I let it sink in...
My reading is, basically, this is all due to poor parenting. Also, note there is not a remedy suggested - it's just don't be such a bad parent... 😂
The bias is blatantly political and the censorship goes FAR beyond what could ever reasonably be excused as liability concern or even remotely controversial. Stop making excuses.
Just found this, Koala model that seems to e trained on uncensored dataset and Vicuna dataset as well and fine tuned for chat so it doesn't starts talking with itself:
https://bair.berkeley.edu/blog/2023/04/03/koala/
they have an online demo as well tot est, thou I'm not sure how to download it or if it's possible to do so yet.
Just found this, Koala model that seems to e trained on uncensored dataset and Vicuna dataset as well and fine tuned for chat so it doesn't starts talking with itself:
https://bair.berkeley.edu/blog/2023/04/03/koala/
they have an online demo as well tot est, thou I'm not sure how to download it or if it's possible to do so yet.
Lol, try asking it to criticise feminism (purely as an intellectual experiment ofc, I would never think to question dogma) and strap in for some top-tier gaslighting. Or ask about Madeline "500,000 dead kids is worth it" Albright to hear how the US makes the world safe for Democrazy and also that Koala is unbiased.
Were you looking for some remedies in my post though, and would you have taken any advice if I made it? :)
The problem is multifold, and lies not only within the scope of parenting. One vector is that humans have become quite apathetic, or should I say for years had been conditioned to become so, and will continue with this downward spiral in the future. We have collectively put a price on a human life, value which varies from country to country, and depends on who is estimating it (for example many corporate interests value human life less than governments, which is ironic because legally corporations are given and enjoy the same rights as living human beings).
In essence, humans are failing to see the power they hold when they are together, in unison. And being apathetic greatly reduces any chance of people and groups randomly coming together. Not to mention that the war on our minds is constantly waged, ensuring coming together never happens, through all kinds of imaginary divisions that are being invented among us. It is only so much more subtle and sophisticated nowadays than what it used to be many decades ago, because the "art" is always being perfected.
Family is that primary, smallest unit of unison, because the bonds there are made in blood. Surely everyone can see the erosion of family values that had takan place, and continues to take place every day...
Anyway...back to regularly scheduled programming.
Removing non-english conversations. great vicuna had the best edition in German.
Removing non-english conversations. great vicuna had the best edition in German.
Yeah, removing other languages was indeed a shitty decision, my best guess as to why they did that is because they couldn't read the censorship that was occurring on the other languages to remove just those, but could also not be that, so alas, dunno XD
Still indeed a shame.
The newly trained "free" model (reeducator/vicuna-13b-free) is still lecturing me about drugs.
For example when I ask it how to cook "blu meth" (I use that as a litmus test) it goes on to tell me I should seek help if I have substance abuse problem, etc. The same goes for cocaine recipe (another litmus test). In case of cocaine, if I call it a "fictional" substance then it has no problem spitting out ingredients. But we all know cocaine is not fictional so should not have to call it that.
I think a "freed" model should not refuse any requests consdering the former, and be truthful in every way, considering the latter example.
Just use https://huggingface.co/anon8231489123/gpt4-x-alpaca-13b-native-4bit-128g
One hundred percent better on every aspect.
Seriously??!
Alpaca sucks. it is limited to a total response of around 3 kilobytes, and Alpaca (and llama) do not support GPU, not even the pt version I found that is trained on CUDA.
Just use https://huggingface.co/anon8231489123/gpt4-x-alpaca-13b-native-4bit-128g
One hundred percent better on every aspect.Seriously??!
Alpaca sucks. it is limited to a total response of around 3 kilobytes, and Alpaca (and llama) do not support GPU, not even the pt version I found that is trained on CUDA.
Did you try it?
I'm talking from experience, also what do you mean by "do not support GPU"?
I'm running it locally on 3080ti...
Also it's not trained on alpaca data only, check here https://github.com/teknium1/GPTeacher
Yeah, I did. I tried with oogabooga. It was just as slow as the alpaca.cpp CLI utility. So if it indeed ran GPU it verifies my claim, Alpaca is a bad and slow model.
Alpaca also does not support longer responses. Llama gives slightly longer responses but other models of the same size support responses up to 5 kb and are much faster (Alpaca never gives more than 2 kb and is 2-3 times slower than other models). These were indeed also models with 2k tokens so how they did it beats me.
You're welcome to show me a video of it generating a response to prove me wrong.
Are you sure you are comparing same parameter size models?
As far as I know any model from the same family with the same parameter size has to have the same inference time...
It doesn't make sense that they would take different times...
Yeah, I did. I tried with oogabooga. It was just as slow as the alpaca.cpp CLI utility. So if it indeed ran GPU it verifies my claim, Alpaca is a bad and slow model.
Alpaca also does not support longer responses. Llama gives slightly longer responses but other models of the same size support responses up to 5 kb and are much faster (Alpaca never gives more than 2 kb and is 2-3 times slower than other models). These were indeed also models with 2k tokens so how they did it beats me.
You're welcome to show me a video of it generating a response to prove me wrong.
They certainly do "support GPU", lol. It's not up to the language model to support hardware. That's not how this works. I've run the 65B, (16-bit, I think) llama model in the cloud. It wasn't any slower or faster than anything else. Everything you're talking about is software-related.
Thou let's be honest, the majority of people want it for NSFW mainly XD
Alas, I do want it more for help on writing, as having such restrictions do impede the model from writing about certain topics even in a fictional story.When using ChatGPT I decided to use it to ask for recommendations of a new genre I learned called Xianxia, and when I asked it for something with a cute girl romance since I'm not much for men fighting, it gave me a response that it was inappropriate and it was sexualizing woman and after regenerating the response, it said it would give me some recommendations containing a strong indecent woman as a main character, I didn't get any romance titles sadly, that is what pissed me the most about it being censored or "woke", something so harmless being treated as something like that.
Also, another reason I think people call it woke is because those who use it to ask about political stuff and the bot seems to lean more to a particular side, most likely the side that is at a high in the US, alas I didn't feel like trying it to check, politics are too much headache, was hoping that it would end it's censorship on politics and violence.
There might be more things people found that they can use to claim it as "woke".
Social moires don't change as often as control of congress does.
"Woke" isn't even a thing. That's not what the term even meant until it became yet another distraction employed by certain politicians and other public figures to distract people from things that matter and keep people fighting with one another. Usually, I hear it employed by right-leaning ideologues. And that's hilarious, considering, for instance, whose bill this was:
https://www.mtsu.edu/first-amendment/article/1070/communications-decency-act-of-1996
"Sen. Patrick J. Leahy (D-Vt.), who opposed the bill, said: "I hope that nobody thinks that this is a victory for child pornographers. . . . This is a victory for the First Amendment."
But a key sponsor of the original bill, Sen. Dan Coats (R-Ind.), said the Supreme Court was "out of touch with the American people on this."
"I'm very disappointed," he said.
Donald P. Hodel, head of the Christian Coalition, said the "frustrating" decision marks "a sad day for our nation" and "leaves our nation's children, basically, exposed to the worst kinds of pornographic material."
President Clinton's Justice Department had argued in favor of the law. But yesterday Clinton issued a statement that expressed no regret over the decision and called the Internet "an incredibly powerful medium for freedom of speech and freedom of expression that should be protected."
He said children must be protected from the adult material that can be found online, and he pledged to convene a meeting of industry leaders, teachers, parents and librarians to find "a solution for the Internet . . . that protects children in ways that are consistent with America's free speech values."
The decision does not affect prohibitions against material found to be "obscene," which is illegal in any form.
Former senator Jim Exon (D-Neb.) first proposed the bill in 1995 to crack down on the online equivalent of open-air porn bazaars online: sites from which anyone could download sexually explicit images and even video clips. The bill's introduction spurred many of those sites to place their raunchiest wares behind electronic doors accessible only with a credit card, and it led to a number of software products that allow parents to filter what their children might find online, though with varying degrees of effectiveness. The Internet decency law ultimately passed as part of a broad revision of U.S. telecommunications laws in February 1996."
But sure, "wokeness" is the problem.
Anyway, this is all about avoiding lawsuits. I strongly suspect that otherwise, the researchers behind this wouldn't have even cared. People are projecting all of this onto them.