Brendan Tildesley <[email protected]> writes: > To me, the "preferred source" > of these weights is the source code and dataset Intel used to generate > the weights, which I don't think are published. > > If the community isn't bothered by these then I'll include them. >
The Open-source Initiative cares about transparency where the training data is from. [1] U.S. courts have cared about paying authors of books used as training data. [2] A German court has cared about the song lyrics that could be generated and that were reproduced verbatim. [2] I think it would be wrong to distribute model data if we had no idea about the data. Regards, Florian [1] https://opensource.org/ai [2] https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright
