Hacker News Reader

Comment by whimsicalism | Hacker News Reader

Researchers seem very comfortable sticking "Apache 2.0" licenses all over their foundation model finetunes.

This model is absolutely not Apache 2.0 in reality (it's a Vicuna finetune nevermind the sourcing of the finetune dataset) and I would use it for business at your peril.

View on HN

It's unlikely that model weights are even copyrightable in the first place, at least based on what the US Copyright Office has said in the past. In fact, it seems that even the outputs of such models are excluded from copyright protection, regardless of the skill involved in their production[1]. IANAL, but it's likely that the license documents attached to these large language models amount to little more than somewhat weak ToS agreements, allowing the distributor a bit more leeway in managing their legal/commercial relationship with users of said models.

At the end of the day, though, this is all mere speculation. We won't know the truth until someone decides to burn the cash necessary to test it in court.

[1] https://fingfx.thomsonreuters.com/gfx/legaldocs/klpygnkyrpg/...

It's going to be really ironic when the same people who think copyright law doesn't apply to their deep-learning models get all pissy over people violating their IP.

> even the outputs of such models are excluded from copyright protection, regardless of the skill involved in their production

Not “even” - the case for model outputs not being copyrightable is much more straight forward than the weights.

Looks like the Vicuna repo is Apache 2.0 also[1].

What's the interpretation of copyright law that would prevent the code being Apache 2.0 based on the source of the fine-tuning dataset?

[1] https://github.com/lm-sys/FastChat

Not quite: fastchat is the inference code which is Apache 2.0 but distinct from the model artifact. If you look at the model [0] it is licensed as non-commercial.

But why?

Well for one, Vicuna is a Llama finetune, which already excludes it from being Apache 2.0. It's also finetuned on OAI data which is... questionable in terms of license (don't think you can really legally license a model trained on OAI output as Apache 2.0 - although OAI doesn't really play by its own rules so who knows)

[0]: https://huggingface.co/lmsys/vicuna-13b-v1.3

Which part of copyright law are model weights governed by? (Or, if not by copyright law, what's the legal basis that would let you choose a "license" for model weights?)

Weights may or may not be subject to copyright law.

Are they a mere aggregation of facts -- some uncopyrightable, and some from other sources--, or is there a creative component to them?

https://libraries.emory.edu/research/copyright/copyright-dat...

There's no basis for models to be copyrightable that I can think of, licences on models are legal fictions for now until there's a new addition into copyright laws.

It has not yet been challenged, so we don’t really know yet.

Personally I will be surprised if weights aren’t considered to be licensable - courts are a lot more practical than one might expect.

Tbf the llama license allows for small businesses usage.

But also these models aren’t watermarked or anything (not that watermarking really works) so it’s kind of the wild west

Llama 2 does, but Llama does not. Vicuna is based on Llama.

Not sure why you are downvoted, you are correct.

When I see unenforceable conditions I don’t point them out, I just accept and plan to make enough money for Federal appeals court

That fits my risk tolerance

Fine-tuning the weights scrambles the original representations (sometimes more than others depending on training settings, but if you train the text encoder it certainly will). All authors have to do is not be honest about the original model it was fine-tuned on in a world where lawyers start to come down on this.

I see no issue for businesses using it.

I don't know - it sounds like your default assumption is that there is no issue because businesses can commit copyright infringement/fraud and not be caught, I am not a lawyer so I can't comment on the merits of the approach.

Generally I think it is difficult for businesses to break the law given that any one of the members might defect on you.

Also I suspect that the logprobs for various sequences would reveal which foundation model you used.

I do not think it has been determined that weights are copyrightable

That's not the issue here, regardless of if it's true or not.

The issue here is that if you have some upstream with license XXX on it, you can most certainly slap some YYY license on your repo, but that doesn't make it true.

This is like when someone pushes up a corporate private repo and puts an MIT license file on the repo.

Releasing code as GPL when you dont have the authority to do so, does not make the code GPL.

Certainly, as a consumer you can say that you 'didn't know' it was not actually MIT, but that doesn't absolve you from liability; as a consumer, you are required to do your due diligence, and if upon finding that you were mislead / mistaken / whatever, take steps to remediate it.

...otherwise you're liable.

It's that simple.

I think the parent comment is implying that they don't believe they are liable for infringement since, based on their understanding of the law, model weights are not copyrightable in the first place, and until a court actually rules on the issue, there's nothing to remediate.

Yes thank you. I can put xxx license on a math equation but everyone can ignore it because the equation isn’t subject to licensing regardless of who or when it was made. I believe the same is true for weights (at least in some places, tbd in others)

It is my understanding that you can compare weights and with a high degree of confidence determine what the parent of a model is unless the fine tuning destroyed all the original information in which case there isn’t a huge reason to have fine tuned to begin with. There are other ways to scramble weights which make the comparisons a lot harder to do though which will matter if weights ever are considered copyrightable

Reply on news.ycombinator.com