Re: Packages that work with LLMs - DaVinci seems problematic too

On 8/14/26 7:32 AM, Jean Louis wrote:
> I will verify more of the LLMs which claim to be trully free, such as:
> 
> - Moxin github.com/moxin-org/Moxin-LLM -- but trained with the non-free 
> LLM?
> - OLMo (from Allen AI) allenai.org/olmo
> - Marin github.com/dlwh/marin
> - DaVinci github.com/GAIR-NLP/daVinci-LLM
> - SmolLM github.com/huggingface/smollm (good candidate)
> - OpenCoder opencoder-llm.github.io
> 
> ...
> 
> Jean Louis

I have now taken a look at DaVinci, it seems problematic too. I've 
started here:

https://github.com/GAIR-NLP/daVinci-LLM

That page lists a "Training Data" link which leads you here:

https://huggingface.co/datasets/SII-GAIR-NLP/davinci-llm-data

This lists the following entries for the training data:

"Public GitHub repositories"

There is no indication given that those were filtered for licensing, let 
alone licensing that has no attribution requirements and is compatible 
with relicensing to most choices of FOSS licensing. I assume that would 
be required however, to use the model output in a way that doesn't 
violate FOSS licenses of the training data projects, given models seem 
to often quote larger sections from the training data without being 
aware of what they quoted from.

The training data also lists "MegaMath" which links here: 
https://huggingface.co/datasets/IFM/MegaMath

This seems to contain math related documents from "Common Crawl". As far 
as I know, "Common Crawl" is basically just a dump of openly accessible 
web pages that didn't ban data crawling: https://commoncrawl.org/ I 
don't think any of those are necessarily licensed in any way that would 
allow quoting from them verbatim for any desired use. It seems like a 
lot of that data might even be proprietary in terms of licensing.

Therefore, I don't see how DaVinci's output is open either. But I could 
be wrong, check for yourself, and as usual this isn't legal advice and 
I'm not a lawyer.

Regards,

Ellie