What means GPL for a model ?

#8
by pycloux - opened

Thanks very much for this very important contribution to the community.

I would like though to understand what means "GPL" for an AI model.
It is possible to use a GPL software product for a commercial organization provided that it does not resell a product derived from this software.
If an business would use a server running the snoozy model "as a service" (accessible through a set of custom APIs) as it is, would this be considered as an infringement of the GPL license ?
If not can we thus consider that Snoozy, under certain conditions, is available for commercial use.

Thanks very much for your answer.

This is an interesting question, and I'm curious what the model creators think, too. Depending on the GPL version, commercial use may not be a problem (The Linux kernel is, after all, GPL2 and is used in many commercial products). A lot also depends on what constitutes a "derivative work" under GPL. For example, using the model directly from the HuggingFace repo wouldn't create a derivative work. Tuning and creating a LoRA? Maybe it is derivative, or maybe not -- it's a separate file that's applied after loading the GPL'ed model. But the new model created via PEFT is arguably a derived work.

Whether creating a pipeline, e.g., LangChain, creates a derivative work, is questionable, as it's not a model based on the original that's been trained or fine-tuned through any back-and-forth between the model and new front end; it's simply a pipe between two or more distinct pieces that interact through a well-defined API. I don't think you can call a program that queries a GPL'ed database via SQL or some custom search language derivative.

Now the model card says this model is derived from Llama, which has a Creative Commons Non Commercial license, which means it can't be used for commercial purposes, GPL not withstanding. In fact, iI have no idea whether the origina Llama license and GPL are compatible or not. Perhaps the creators could comment on this as well?

As much as I would love to see a 13b model for commercial use, a derivative work of a Creative Commons NonCommercial (CC-NC) license cannot be distributed under the GNU General Public License (GPL).

When creating a derivative work, you generally need to comply with the licensing terms of the original work. If the original work is licensed under a CC-NC license, you must ensure that the derivative work is also non-commercial in nature. Distributing a derivative work under the GPL would conflict with the non-commercial restriction of the CC-NC license.

If you're looking for something with less licensing issues, the efforts of this project may be something to keep in mind:
https://github.com/togethercomputer/RedPajama-Data

Thanks @panomity and @patrickhwood , I guess agree with both of you.
I do think that the license of the present model is debatable (it is labelled as "non commercial" on the GPT4All web site by the way).
I also think that GPL is probably not a very good license for an AI model (because of the difficulty to define the concept of derivative work precisely), CC-BY-SA (or Apache) is less ambiguous in what it allows or not.

I will keep the thread open one week if someone has a different position about it.

Well, looking at the model card for Llama (which this model links to through the Llama github repo), which this is based on: https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md, the license is listed as "Non-commercial bespoke license", so any derivative work (like this one) should defer to it. For example, the decapoda-research/llama-13b-hf model lists its license as "Other".

Sign up or log in to comment