You can't possibly think those models are only trained on open source data?