how do opt-outs work
There's supposedly an "opt-out pipeline", but I don't see any such thing anywhere, just people manually making pull requests and then squashing that in
How does this work?
The opt-outs are applied in batches on a roughly weekly basis, the dataset is filtered out offline and then reuploaded. The opt-out requests receive a reply after that is done. The current state is reflected in https://huggingface.co/spaces/HuggingFaceCode/in-the-stack
Squash commits release the references to old .parquet files to allow the server's garbage collector to do its job (usually 24-48h until the CDN cache is clear).
a roughly weekly basis
So it is manual, if it's a rough timeframe? Does that mean if the maintainers call in sick for a while the "opt-outs" will just pile up?
the dataset is filtered out offline
Yes, but how? I doubt it's an entirely manual process, given the several thousand requests that were handled at once, and the several hundred since then. How are the github issues turned into actual removals? What scripts are used for it?
So it is manual, if it's a rough timeframe? Does that mean if the maintainers call in sick for a while the "opt-outs" will just pile up?
For a bit of context: there are at most 2 active researchers working on The Stack, so this absolutely takes some time. As far as I know, it is the only code dataset that offers an opt-out mechanism among thousands of other open code datasets, exactly because it is non-trivial to maintain, given the many edge cases of opting out.
We wanted to offer a solution for developers to make sure that their repositories at least don't end up in the datasets built based on The Stack, and are actively working on it.
How are the github issues turned into actual removals?
We auto-add the requests to the queue if the author matches the requested accounts/repos, or is an owner of the significant portion of the requested org. The rest require some manual work on our end to verify.
Thank you for your feedback!
because it is non-trivial to maintain, given the many edge cases of opting out
Have you considered that there would be much less effort to opting out if you filtered out all the repositories without a valid licence :^)
thousands of other open code datasets
In the link you shared, I found mostly research done using specialized datasets that are not republished.
The two datasets I found are "Public Git Archive", which seems dead in the water and also doesn't have anyone processing removals, and their own tool to view the dataset doesn't even work anymore, and "GH Archive", whose stated primary goal is archival and that presents itself as a research tool.
(Even if other datasets illegitimately copying ARR source code exist, this does not excuse making one yourself)
The Stack v3, conversely, apparently has a grand total of 2 people maintaining the entire dataset right now, states plainly that it is "built to pre-train code LLMs with full-repository context", which is decidedly not archival or research, and contains a not-countable (because the weird Data Studio tab keeps breaking on me) amount of repositories explicitly labelled as not having a license, meaning they are All Rights Reserved, minus the exceptions granted by GitHub's platform terms.
GitHub additionally states that just because code is public, that does not mean other users gain any rights, other than to view and fork it.
I cannot find any source stating that huggingface -or the people maintaining this repository- are partners/affiliates of GitHub, which would give them permission to freely use public data for the sake of creating datasets for the purpose of training LLMs, under D.4 of GitHub's terms of service
@Toby222 if I understand you correctly, you do not have a problem with providing your data to GitHub and its Affiliates (explicitly: Microsoft, implicitly: OpenAI) to build closed-source paid AI models without an option to opt-out (besides the the Inputs and Outputs of their coding assistants), but have a problem with letting non-profit and research labs to benefit from having access to this data and publish free open models.
Let’s leave interpretations of the law to the lawyers and keep the conversation productive. Happy to answer any questions regarding your own data in the dataset if you see any after the last round of opt-outs.
@Toby222 if I understand you correctly, you do not have a problem with [...]
Yes I do, I just agreed to it for the benefits of using the platform, so I've resigned myself to it.
[...] have a problem with letting non-profit and research labs [...]
No, I have a problem with Hugging Face Inc., a for-profit company, redistributing code they have no rights to, which I, and the majority of people represented in the dataset, did not consent to.
Let’s leave interpretations of the law to the lawyers and keep the conversation productive
Ok, without legalese then, you took thousands of repositories without permission and can't even acknowledge why people would be annoyed at that.
your own data in the dataset if you see any after the last round of opt-outs
Hell if I know how to parse a multi-terabyte dataset. I tried using the duckdb thing in the description to scan for commits authored by me and ran into an API rate limit. :^)
Just made an access token to see if that manages to not get rate limited before the query finishes
After a solid 40 minutes and a bit of doing nothing but download at a fixed rate of 100Mibps, using up just about all available ram, and using most of a CPU core the entire time, I have now ran out of time to keep this thing running. If there's a way to search through the dataset in a way that doesn't take an access token and an hour on a dedicated machine, or several terabytes of disk space, so I can confirm that none of what I made is in this repo, I'd very much like to know about it
How am I to interpret a heart emoji reaction
