When we're dealing with source code, the cost of getting from source -> binary is minimal. The entire Linux kernel builds in two hours on one modest machine. Since it's cheap to compile and the source code is itself legible, the source code is the preferred form for making modifications.
This doesn't work when we try to apply the same reasoning to `training data -> weights`. "Compilation" in this world costs hundreds of millions of dollars per compilation run. Cost of "compilation" alone means that the preferred form for making modifications can't possibly be the training data, even for the company that built the thing in the first place. As for the data itself, it's a far cry from source code—we're talking tens of terrabytes of data at a minimum, which is likewise infeasible to work with on a regular basis. The weights must be the preferred form for making modifications for simple logistics reasons.
Importantly, the weights are the preferred form for modifications even for the companies that built them.
I think a far more reasonable analogy, to the extent that any are reasonable, is that the training data is all the stuff that the developers of the FOSS software ever learned, and the thousands of computer-hours spent on training are the thousands of man-hours spent coding. The entire point of FOSS is for a few experts to do all that work once and then we all can share and modify the output of those years of work and millions of dollars invested as we see fit, without having to waste all that time and money doing it over again.
We don't expect the authors of the Linux kernel to document their every waking thought so we could recreate the process they used to produce the kernel code... we just thank them for the kernel code and contribute to it as best we can.