You mean like how the model itself is a derivative work of tons of copyrighted content? If the original model can sidestep the issue of being trained on copyrighted content, then it should be fair game to train a new model off of a copyrighted model.

Yes, the model is a derivative work of its training data, but the difference is that a model is transformative. Llama is not a replacement for reading some book. On the other hand a fine tuned model built upon llama is much less transformative.

That’s a legal unknown. And it’s also a technical unknown how you would even determine it was descended from the same model in a way that would hold up in court.

I would imagine that the weights of a finetuned model are highly correlated with the original weights. Having said that, simply permuting the neurons would make it way harder to match them up, I can't think of a straightforward way to reverse it.