[nmstoker]

Very impressive but I don't follow how the author can license it under MIT terms when they've trained it using material only licensed for research purposes.

Wouldn't research purposes be non-commercial and therefore be incompatible with MIT terms?

[Eisenstein]

The script is MIT licensed. The model weights might not be.

[reaperman]

In terms of copyright law, it's not clear that the model weights are derivative works from the training material. At this time, I think courts would lean towards the weights not being a derivative work. And even if it is a derivative work, it could be considered transformative use and therefore not a violation of copyright. Therefore, IMHO the creator of the model can license the model weights however they want to, regardless of the material that was used to train it.

For now, anyways.

If model weights are ruled derivative works of copyrighted material, it would wipe trillions of dollars off the S&P500 overnight. Currently, copyrighted works are needed to train all the big foundational models - and there's no practical or cost-effective way to get a license for all of it.