4 ms·

Thanks! I missed that. The attribution by training data source category (arxiv vs wikipedia vs nemotron etc.) is an interesting approach.