To get it to run on different platforms and gain new features like super-resolution and so on, yes. Rom files are the preferred form for modifying old games. No one bothers digging up old source code and assets to reconstruct things when they can use an emulator to essentially spit out a derivative binary with new capability. See every re-release of a 16 bit era or earlier game.

Now that I've beat my head against this issue for a while, I think it's best summed up as: weights are a binary artifact, not source of any kind.

If what you say is true—if the ROM is the preferred form for making modifications (even for the original company that produced it) and the ROM is released under a FOSS license—then sure, I have no problem calling it open source.

Preferred by who? It sounds like these people have a strong say in what constitutes open source.

Preferred by anyone who's actually using and modifying the work.

No one trains an existing model from scratch, even those who have access to all of the data to do so. There's just no compelling reason to retrain a model to make a change when you have the weights already—fine tuning is preferred by everyone.

The only people I've seen who've asserted otherwise are random commenters on the internet who don't really understand the tech.

> Preferred by anyone who's actually using and modifying the work.

> ...fine tuning is preferred by everyone

How do you know this? Did you take a survey? When? What if preferences change or there is no consensus?

> The only people I've seen who've asserted otherwise are random commenters on the internet who don't really understand the tech.

There are lots of things that can be done with the training set that don't involve retraining the entire model from scratch. As a random example, I could perform a statistical analysis over a portion of the training set and find a series of vectors in token-space that could be used to steer the model. Something like this can be done without access to the training data, but does it work better? We don't know because it hasn't been tried yet.

But none of that really matters, because what we're discussing is the philosophy of open source. I think it's a really bad take to say that something is open source because it's in a "preferred" format.

> I think it's a really bad take to say that something is open source because it's in a "preferred" format.

Preferred form and under a free license. Llama isn't open source, but that's because the license has restrictions.

As for if it's a bad take that the preferred form matters—take it up with the GPL, I'm just using their definition:

> The “source code” for a work means the preferred form of the work for making modifications to it.

Today, the weights may be the preferable format indeed, due to the cost. Are you going to change the definition tomorrow, when the cost drops?

Sure, why not?

A good definition should not depend on transient state of affairs.