not a lawyer, but i dont think it is enough to change the license, as its still derived from the LLaMa weights and so you'd still have to follow the rules.

>Meta grants you a non-exclusive, worldwide, non-transferable, non-sublicensable, revocable, royalty free and limited license under Meta’s copyright interests to reproduce, distribute, and create derivative works of the Software solely for your non-commercial research purposes. The foregoing license is personal to you, and you may not assign or sublicense this License or any other rights or obligations under this License without Meta’s prior written consent; any such assignment or sublicense will be void and will automatically and immediately terminate this License.

https://huggingface.co/decapoda-research/llama-7b-hf/blob/main/LICENSE

This is code for running the LLaMa model, sort of like llama.cpp.

It's a reimplementation of Facebook's original GPL-licensed open source client under a more permissive Apache license. The GPL requires all your other code to also be GPL, so you can't use it in closed-source projects.

This doesn't affect the license for the model weights, which you will still have to download from somewhere else.

Changing the video player you're using to watch a movie doesn't make the movie any less copyrighted; the same kind of mechanics would apply here.

Honestly at this point I am not sure weights can be copyrighted: they have no human "author". It is a total gray zone. Tribunals will rule knia few years that the habits taken now are the jurisprudence

I agree with this as well. Just because Meta isn't enforcing their license, this does not mean the license has gone away. At some point in the future, Meta could enforce it.

The weights are a part of the model. They will be under the same umbrella. You cannot use the weights with the model. It is created internally even if you cannot see it

If something lets you replicate someone's IP with a few trivial steps, and you use it to get financial gains, you will be in trouble.

I'm leaning in this direction myself.

IANAL, but I think about the Feist Publications ruling which dealt with raw listings of facts (white pages names, addresses, phone numbers organized in the most functional format - alphabetic.) SCOTUS ruled that the raw data was not copyrightable even though it took a lot of effort to collect and compile it. It seems to be that the raw data here are the weights which would make them not copyrightable. The structural design of the model might be, and more than likely the compiled model with weights would be copyrightable.

I suspect this will work its way through the courts just in time to be rendered moot.

How would either an architecture or a model be copyrightable?

Architectures are algorithms. If they aren't patentable and are in addition to that patented, they have no protection.

Model weights are a result of a mechanical procedure that fits a model to data, minimising some kind of error. That is not a work of human authorship.

Things that could be copyrightable are an article describing a model architecture, or a specific software implementation of a model.

As an argument why model weights are unlikely to be copyrightable consider the following parallel: we know that model output, for example, a story generated by ChatGTP based on a prompt is certainly not copyrightable, since it's not a work of human authorship, but then, how is the model? We can view the selection of training examples as something similar to a prompt and the training process as similar to the inference. I think giving copyright protection to model weights might be reasonable though, but I think it's unlikely that they have copyright protection.