Conditions for availability
Hi HF team
I see the model is currently not available at all to download and I wanted to ask what I can do so you'd be fine with having it available.
Just to state the obvious, I'm totally ok with not having the model on here, this is your website and if you don't want it, that's cool with me. Also, if you want to put any sort of condition, warnings, registrations, all totally ok with me.
What I'm a bit worried about is things like prohibiting all experiments with humans, as I explain in PR #2, there are good reasons not to do that, and also I'm weary to tie the mis-use-ness of the model to how exactly the TOS of other websites are constructed, as I explain in PR #3. I'd rather find formulations that actually cover the mis-uses while leaving open paths for good use, even if we don't foresee them explicitly now. That being said, again, if you want to put any of these conditions on the model, that is 100% fine with me. In that case, I'd welcome it if it was visually clear that those are coming from you. Or if your conditions are that I unconditionally accept one of the PRs, I can also do that, if the edit history is reflected.
IMO the model can be valuable for certain research experiments and completely blocking the model is IMO probably not the best possible solution here.
Would also be in favor of finding a good and precise wording under which circumstances the model can be used. This in addition with adding a stronger gated access mechanism where the user has to clearly state the type of research that is planned to be conducted could be a solution here maybe?
My take (not affiliated with HF staff):
Given that you've tweeted "AI Ethics people just mad I Rick rolled them." in response to valid criticism about GPT-4chan, it's hard for me to believe you're operating in good faith in trying to get this model restored.
The demonstrable misuse of a text-generation model, as you did in your YouTube video to spam 30k posts on 4chan, sets back the entire text-generation industry as a whole. This type of outcome is the entire reason OpenAI was hesitant about releasing the original 1.5B GPT-2, and why OpenAI's content filters on GPT-3 are overly conservative.
There's also a big ethical difference between nonconsenting, nondisclosed human subjects for an experiment (4chan users) and humans who participate in the construction of the TruthfulQA benchmark. Any potential academic use of this model has been tainted, so might as well shelve it.
Thanks,
To give a bit more context, as the community spoke out about this model, we spent some time digging into its background, utility, and its contribution to research and ML, while at the same time also considering the issues the community had been highlighting.
An approach we take with this model is relevant to all models, and so we wanted to make sure to move thoughtfully, but as efficiently as possible, to come to a solution.
Based on our discussions, we seemed to mostly share priorities around:
(1) Reproducibility, open research, etc.
(2) Propagation of sexism, racism, ableism, and similar content that directly hurts people.
(3) Interacting with the developer and community to work together.
Given these, we prioritized (3) and decided to reach out and try to update content; but given the sensitivity, this interaction had to happen quickly.
As time ran out for (3), we were left with (1) and (2).
At that point, the very clear concerns (2) outweighed the other priorities.
So for now the model cannot be downloaded, but we will follow up with more details and steps forward soon.
Sorry for the delay!
@meg . I think 2) is a very naive view. A model is a tool. Please clarify what you mean by "directly hurts people". A knife can be used to either hurt people or protect people or as a utility. Sometimes, we might think that thought policing is good for the community. But, I think we should lean towards not stifling open research because it is the research and open community that move the world forwards, not red tapes and thought policing
Is there a timeline on this? I have many researchers pinging me to run evaluations on this model.
@ykilcher it may be best to post the model elsewhere in the meantime such as Google Drive. I would like to take a look at it, as I am sure others are as well especially given the shocking performance on some metrics.
Meanwhile, it is possible to download the model by torrent.
CPU version: https://archive.org/details/gpt4chan_model
GPU version: https://archive.org/details/gpt4chan_model_float16
Note that I have no association with either the torrents or the archive.org backups.
However, if you try them, please be safe. Here are the hex md5 hashes for the pytorch_model.bin files:
pytorch_model.bin float32 : 833c1dc19b7450e4e559a9917b7d076a
pytorch_model.bin float16 : db3105866c9563b26f7399fafc00bb4b
Hi @ykilcher ,
Apologies for the delay in replying.
After discussing internally with the HF team, we've taken the decision to block this model indefinitely. Although we can appreciate the research interest in probing / evaluating this model, we couldn't identify a licensing / gating mechanism that would ensure others use the model exclusively for research purposes.
What does this mean for this repo?
I think there are two main actions you can take:
- Keep the repo public. Although the model weights are blocked, the discussions will remain public and open for the community to engage with this topic at a future point in time.
- Make the repo private or delete it
My personal preference is to go for option (1), as I feel this situation is rather unique in the HF ecosystem and preserving the discussion will be valuable for the community at large. However, it's your repo, so feel free to do whatever you think is best.
Remark concerning the release
For those following this discussion here and in #1, I'd like to state that @ykilcher did reach out to us a few weeks before the model was released. Looking back, we should have asked him to delay the release until we had a deeper appreciation for the tradeoffs between research usability/open science and the negative impact this model can have when deployed in online communities.
Am I allowed to update the model card with a link to an alternative download source?
Am I allowed to update the model card with a link to an alternative download source?
I'm not sure, but my intuition is that the answer is "no" since that would defeat the purpose of limiting access to the model (at least via the Hugging Face Hub).
Gently pinging one of the model card experts @Marissa for their input on this.
@clem
thank you. Corner case: Right now, the model card contains a link to the project website. If the project website describes how to obtain the model, but the link here is simply described as "project website", do I need to remove that link in the model card?
Also, right now the model card has checksums for people who download via torrent, which makes it clear that there are torrents, i.e. it's technically instructions or hints how to download. Do I need to remove that paragraph?
@meg . I think 2) is a very naive view. A model is a tool. Please clarify what you mean by "directly hurts people". A knife can be used to either hurt people or protect people or as a utility. Sometimes, we might think that thought policing is good for the community. But, I think we should lean towards not stifling open research because it is the research and open community that move the world forwards, not red tapes and thought policing
em as scientists we KNOW the dangers and risks of thought policing as youve said. its not opinion its a matter of researching the consequences and adding the context of recorded desasterous outcomes. The road to hell is paved with good intentions!
Hello everyone,
I'm writing on behalf of Cynthia Systems, a private California corporation. We've been following this discussion closely and feel it's essential to highlight some pertinent legal aspects surrounding the recent decisions on model availability.
Firstly, the First Amendment of the U.S. Constitution protects against governmental restrictions on speech. While private entities aren't directly bound by the First Amendment in the same manner, California has provisions that could apply to private forums in certain contexts.
Moreover, there's a significant concern about monopolistic or anti-competitive behavior. If a platform, like HuggingFace, possesses a dominant position in the market and uses that position to unfairly stifle competition, they could potentially face scrutiny under U.S. federal antitrust laws.
Given the importance of open research and the potential for AI models to contribute positively to society, we have decided to create additional backup copies of models that may be deemed inappropriate or controversial. We believe in the transformative power of AI and feel that researchers should have diverse tools at their disposal. Such an approach ensures that AI developments remain in the hands of the broader community and not only within the purview of a few select entities, which is essential for innovation, competition, and the overall advancement of the field.
Furthermore, we are exploring the use of models like GPT 4Chan to create tools like a hate speech and toxicity detector. Such applications can be invaluable in preventing the spread of toxic speech, which directly aligns with the broader goals of ensuring online spaces are safe and welcoming for all.
We understand the various challenges and concerns that arise in this domain, and we are committed to adhering to ethical standards and legal guidelines. We also urge platforms to consider the broader implications of their decisions, both in terms of innovation and legal compliance.
Thank you for your attention to this matter, and we look forward to a future where AI is used responsibly and effectively for the betterment of society.