Que veut dire « libre » (ou « open source ») pour un grand modèle de langage?

FWIW here's a mapping between the FSF's software freedoms and the stuff you can do with an LLM. Personally I think they can be "libre" as well as open source with the right permissions. http://marble.onl/posts/considerations_for_copyrighting_AI.h...

Also, auditing any big dataset given verbatim is practically impossible for now. Instead of including the dataset verbatim, the model that purports to be practically-usefully open-source should contain a relatively small procedure for deriving the dataset from some reputable general-purpose dataset, small enough that the resulting dataset can practically be audited.