eaddario

/

Watt-Tool-8B-GGUF

like 0

I suppose it all depends on what you mean by "free" (beer vs speech). LLMs, and this also applies to other ML/AI algorithms BTW, tend to fall into one of three categories: really open source, open weights, and closed. On the other axis, you'll find the restrictions that the authors impose when releasing their IP freely (apologies for the childish pun 😂). Some license their research with almost no strings attached (e.g. MIT, Apache, etc.), others want to ensure access will continue on their terms (*GPL-like) while others impose restrictions based on some sort of tiering (e.g. ok if less than XXX milllion users, or fine to use but can't train/distill other LLMs, etc).

I'm sure there are many others but the closest to free-as-in-speech I can think of at the moment is the work that the Allen Institute for AI is doing with OLMo (https://allenai.org/open-models). Their models are really open source (training data, hyperparameters, mathematical optimizations, the whole nine yards...), but my interest is to find ways to reduce the size of LLMs in general, rather than focusing on a particular type. Having said that, I may try to shrink one of their models next.

On your last point, and you deserve a 🌟 for attention btw, I did pick up on the discrepancy between the foundational model (Llama) and what the Watt AI folk used when releasing their version. I don't know why they decided to release under Apache, but since I'm not qualified in IP Law, I opted use the same license as the parent model.

When we speak of free software, that is and never was about the beer, but freedom. Freedom to do what you wish with software. When we talk about "Open Source" we talk about Free Software. It is not related to "Open Weights", as contemplating definitions is wrong.

Back to point, you can't be changing licenses. I wish it would be Apache 2.0, but it is not. It has original license. It is illegal, and users who download it may think it is Apache 2.0 but you in first place didn't have permission for the re-licensing, and if I am wrong, then you could maybe show the permission? I don't know.

Even if you used wrongly re-licensed model, your awareness that it was wrongly re-licensed makes you "perpetrator". Why play with fire? It can backfire.

AllenAI before few months has shown to be one of best pushers of truly free as in free software LLMs, but then adopted Llama, and is now just deceptive organization. Truly free LLM, free as in freedom, are produced by IBM (Granite), Microsoft (Phi and others), Qwen, Deepseek, Alibaba, while some have restrictive licenses, those organizations have free models.

Always build on free models.

Re-licensing because upstream illegally re-licensed makes you liable in front of the law. We abide by good rules and respect to others, not so?

Re-licensing software without permission from the original license holders is generally illegal. This is because the original license dictates how the software can be used, modified, and distributed. If a project re-licenses software under a different license (e.g., Apache 2.0) without permission, it can lead to legal issues. Users who download the software under the assumption of a different license may be misled.

I think you don't want those conditions.