I don’t think even that is true. I conjecture that Facebook couldn’t reproduce the model weights if they started over with the same training data, because I doubt such a huge training run is a reproducible deterministic process. I don’t think anyone has “the” source.