again, it's all about the data (and we need better open datasets to close the gap w the closed sota models)