I don't think it's necessarily true that model size would need to be larger than dataset size. It's theoretically possible that the model encoding achieves sign

I don't think it's necessarily true that model size would need to be larger than dataset size. It's theoretically possible that the model encoding achieves significantly better compression than DEFLATE or GZIP or whatever compression algorithm is used to store the dataset itself.