Re: Packages that work with LLMs - Marin seems problematic

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 just looked into Marin, seems problematic too. They have this site:

https://marin.community/

I followed the "retrospective" link for their Marin 32B model:

https://marin.readthedocs.io/en/latest/reports/marin-32b-retro/

This links to a "Marin Data Browser", which I assume is the training 
data they used:

https://marin.community/data-browser/experiment/?path=gs%3A//marin-us-central2/experiments/exp600_tootsie-9126ea.json

This lists "starcoder" data. While it doesn't directly link where that 
is from, it seems to be this: 
https://huggingface.co/datasets/bigcode/starcoderdata

This is just an older version of "The Stack", apparently, and the 
website for it says this:

"Any use of all or part of the code gathered in The Stack must abide by 
the terms of the original licenses, including attribution clauses when 
relevant."

As a result, I would again expect a model trained on this not to fulfill 
attribution of the original licenses, since models don't seem to be able 
to reliably trace back to what they're even quoting from. So I don't 
understand how the output wouldn't again be ignoring FOSS licensing of 
the original training data.

But I could be wrong, feel free to look into it yourself. Naturally, 
this email again isn't legal advice since I'm not  a lawyer.

Regards,

Ellie