The output of a compiler is directly based on what you put in, the source code. That makes it a derivative work of the copyrightable source code, and thus copyright to the copyright holder of the source code, not the person who runs the compiler.
One might argue that model weights are derivative of the training material and copyright held be the copyright holder of the training material. The counter argument would be that the weights are significantly transformational.
Of course, model weights almost certainly are not copyrightable so the license isn't enforceable anyway, at least in the US.
The EU and the UK are a different matter since they have sui generis database rights which seemingly allows individuals to own /dev/random.