It is a fair point though - there's no utility in training an openly available model from scratch. Finetuning is far more practical.