Just make it a LJM (Large JSON Model) capable of predicting the next JSON token from the previous JSON tokens and you would have massive savings in file storagre and network traffic from not having to store and transmit full JSON documents all in exchange for an “acceptable” error rate.