Ludovic Courtès <[email protected]> writes: > 4. Packages in Guix will always be **built from source**, the only > exceptions being compilers or build systems for which a bootstrap > has yet to be found (a notable example is GHC). For software that > includes neural networks, we consider the Corresponding Source to > include all the training data; software for which training data is > unavailable, or for which re-computing weights from training data > is infeasible, *cannot be included in Guix*. > For example, LibreTranslate, which [downloads models for local > > use](https://lists.gnu.org/archive/html/help-guix/2026-04/msg00004.html), > may not be included in Guix. Conversely, GNU Backgammon, [which > can recompute its neural network > weights](https://mastodon.nz/@gtw/115851324313436125), is > acceptable for inclusion in Guix. Should applying this rule lead > to package removal, the [Deprecation > > Policy](https://guix.gnu.org/manual/devel/en/html_node/Deprecation-Policy.html) > must be followed.
I'd propose a minor change: I agree that we should not package models whose weights cannot be (realistically) recomputed, but I don't see as much of an issue in packaging software that is intended to be used with such models, or even software that facilitates downloading and using such models. For me, this is akin to packaging libre software intended to work with non-libre datasets, or to packaging libre interfaces to a non-free API. Heck, even wine is packaged in guix. Apart from this caveat, I wholeheartedly support the initiative. Kind regards, pinoaffe
