Re: Packages that work with LLMs

"Dr. Arne Babenhauserheide" <[email protected]> Fri, 14 Aug 2026 12:57:11 +0200
Newsgroups	gmane.emacs.devel
Message-ID	<[email protected]>
Jean Louis <[email protected]> writes:

> Hello Dr. Arne,
>
> On 2026-08-14 12:27, Dr. Arne Babenhauserheide wrote:
>> Jean Louis <[email protected]> writes:
>> 
>>> Their logic is delightfully simple: if a website doesn’t explicitly
>>> tell a bot to go away in its robots.txt, then the data is "openly
>>> available."
>> This complies with the EU copyright directive article 4, so it is
>> legally sound in the EU (but might not be elsewhere in the world):
>> https://en.wikipedia.org/wiki/Directive_on_Copyright_in_the_Digital_Single_Market#Article_3_and_4
>
> Thanks, I got you. Though that copyright exception is for purposes of
> training LLM (text and data mining). The LLM can be trained on it, no
> issues.
>
> But that LLM is trained on copyrighted datasets.

Yes, and in the EU that’s legal if there wasn’t an opt out. Though that
wasn’t the intention of the copyright directive.

> This creates issue for the user using the LLM to generate text. When
> text is generated it may potentially infringe on someone's copyrights,
> especially if text is given verbatim.

Yes, that can happen. 

And then the user has a legal problem.

> That means we have to find LLM which is trained on fully free and
> known list of datasets. Apertus isn't "aperto" enough.

My preference would be LLMs trained only on AGPLv3 or later or cc by-sa
compatible data that correctly attributes which projects contributed
materially to the output.

And marks them properly as AGPLv3 with creativecommons as proxy¹ whether
using new versions of AGPL is allowed.

That way most FLOSS source code and Wikipedia and Stackoverflow could be
used for training data. Also some websites like mine ☺

cc by-sa is the special case here, because cc by-sa is one-way
compatible to GPLv3 only², and for future versions of GPL Creative
Commons must agree that they apply (the GPLv3 allows for such proxies).

¹ https://wiki.creativecommons.org/wiki/ShareAlike_compatibility:_GPLv3#Option_to_comply_with_later_versions
² https://creativecommons.org/compatible-licenses/

Best wishes,
Arne
-- 
Unpolitisch sein
heißt politisch sein,
ohne es zu merken.
https://www.draketo.de
signature.asc (application/pgp-signature, 1.1 KB)