Datasets:

bigcode

/

the-stack

like 657

Should have been opt-in in the first place. This begs the immediate question of how to communicate to everyone that I don't want my code be trained by yet another LLM wasting precious resources.

Something akin to robots.txt but for code repositories.

The little sympathy I had for HF is now gone for sure

@ghisvail Sorry to hear that! Have you opted-out your repos? Personally I like the stack as it's a great practical middle ground with a dataset that is big enough to train really good models but more transparent, open & consensual (thanks to opt-outs) than the traditional way researchers train models (aka by secretly scrapping all github repos).

It would be really great to create an opt-in dataset & robot.txt too though! Would anyone be able to lead this by any chance (happy to help)?

but more transparent, open & consensual (thanks to opt-outs)

Consent: Acceptance or approval of what is planned or done by another; acquiescence. synonym: permission.

I never gave permission for this. Really, the tech industry needs to rethink its difficult relationship with consent.

And sadly, your answer only strengthens my point above.

So do you think it's better companies train secretly on your repos without opt-outs (which is the common practice currently unfortunately)? My point is that this is a step in the right direction if that makes sense?

So do you think it's better companies train secretly without opt-outs

This is called a false dichotomy.

My point is that this is a step in the right direction if that makes sense?

I don't think so. It still is ethically questionable.

I gave you an alternative: ask permission (i.e. consent), which is the right thing to do from an ethics point-of-view.

This is an interesting thread because:

• The Big Code project has been going on for years, e.g. milestones https://www.bigcode-project.org/docs/about/timeline/
• Most open source licenses are completely compatible with The Stack... Apache 2.0 even explicitly grants a copyright license. I.e. Apache 2.0 is effectively consent so far as copyright goes.
• And yet for The Stack there are now over 1700 opt-out requests on Github Issues alone, and it appears like the earliest requests have not even been serviced (i.e. the code is still there): https://github.com/bigcode-project/opt-out-v2/

When it comes to research and especially research which faces IRB review, there is a distinct issue around informed consent ( https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2840885/ ) where participants are made aware of likely consequences and expected outcome of their participation. And open source licenses do not explicitly contemplate informed consent; in particular, Apache 2.0 was written 20 years ago, before we even had Alexnet.

So @clem while I agree The Stack's design is an order of magnitude better than what many companies are doing privately with similar data today, I'm not so sure it's fair to say The Stack is "consensual." There is definitely a "free as in beer vs free as in freedom" debate to be had here. While most OSS licenses themselves predate LLM-as-a-business-model, The Stack itself does not, and so it seems fair to acknowledge that The Stack's participants do not necessarily even have the tools (licenses) to express "consent" to certain research and/or commercial use.

@clem I hope to see Hugginface continue to lead this space, perhaps by helping design a new family of OSS licenses. Would be really interesting to see if @meg has thoughts on what informed consent means in a post-LLM world.