Currently, we unvendor tests via the build/unvendor-tests flag. It could be useful to additionally have a more generic unvendoring mechanism to unvendor other sub-components. For instance, some packages have included datasets (sometimes rather large) that it could be useful to unvendor: geopandas 1MB, scikit-learn 0.5MB, scikit-image: 7.5MB

Maybe it could be a flag of the form,

build:
    unvendor:
         - package-name: <regrex or glob>
         - package-name2: <regex or glob 2>

Also,

• ideally, the list of unvendored packages for a given package should be easily discoverable in the repodata.json (including the unvendored tests)
• we could add another MetaPathFinder similar to UnvendoredStdlibFinder which for a given ImporError, looks at the path, checks it's in a known top_level fields, and if it raises a meaningful error about how to install the missing component with micropip. This would improve a lot the usability of unvendored packages.
• The previous step would require that the top_level entries are correctly auto-determined from the unvendored packages (which is likely not the case currently). Related Parse top-level import name from wheel file #3006