But the source to train your own LLM equivalent is also released though (minus the data). Hence why there are so many variants of LLaMa. You also can’t fine tune it without the original model structure. The weights give the community a starting point so they don’t need literally millions of dollar worth of compute power to get to the same step.

Would Mistral's engineers be satisfied with the release if they had to rebuild from scratch?

But they built a llama equivalent + some enhancements that gives better performance…I’m not sure if this would be possible at all without Meta releasing all the required code and paper for LLaMa to begin with.

Meta didn't release all of the required code to build LLaMa, just enough run inference with their weights.

> Would Mistral's engineers be satisfied with the release if they had to rebuild from scratch?

Yeah, probably. But depends on what you're asking. In the exact same method to get the exact same results down to epsilon error? (Again, ML models are not deterministic) Probably not. This honestly can even change with a different version of pytorch, but yes, knowing the HPs would help get closer.

But to train another 7B model of the __exact__ same architecture? Yeah, definitely they've provided all you need for that. You can take this model and train it from scratch on any data you want and train it in any way you want.

I didn't ask if they'd be able to make do. I asked if they'd be satisfied.

Also, wrt

> Again, ML models are not deterministic)

ML models are absolutely deterministic if you have the discipline to do so (which is necessary at higher scale ML work when hardware is stochastically flaky).

It’s not a build artifact. The algorithm is the open source code.

So the engineers that built this edited this release directly? Or did they write code that built this?

Every engineer uses tools to create their products. Those products that are open source do not include the tools used to create them (usually).

"Open source" normally contains what the engineers working on it edit to submit a job to build into the build artifacts. This does not contain that, but instead the result of the build process.

That’s because you’re failing to differentiate a build process meaning compiling your stuff to an end result, vs. fine tuning an algorithm to refine code.

Explicitly.

I regard the distinction as meaningless and haven't heard a good reason as why I should reconsider when the process behind creating the weights is so integral to the overall engineering process here.

Well fortunately that’s just your opinion and it is neither popular nor relevant.

To quote stallman

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

That’s what this is. This is not a work that produces weights. So the code related to making weights would not be the source code of this project. It would be the source code of the tools used to make this project.

That quote is specifically designed to back up my position, hence why I've quoted it else where in this thread.

This is evidenced by the fact that if Mistral's engineers had to make modifications to this model, they would use other code to do so. This model is not the "preferred form of the work for making modifications to it", but simply the best we have in a world that isn't used open sourcing these models.

> This is evidenced by the fact that if Mistral's engineers had to make modifications to this model, they would use other code to do so.

It’s a PyTorch model. There is no secret internal code required to interact with it. You can write whatever you want. Unless you’re implying that in order to call this model open source they need to get all of PyTorch on the same license. You can write whatever you want to edit the weights.

Open source does not require babysit walking people through the basic competency steps of understanding how the tools of this domain work.

This model is 100% the preferred form of work for making modifications to it as evidenced by the large community oriented around sharing and adapting models from exactly this format.

And no, I’m calling you for changing your argument again. You were saying dumb shit about how they needed to open source all the data needed to train this thing from scratch, which is another wildly different argument.