> 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.
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.