`choose_from_datasets`. Setting `stop_on_empty_dataset=True` will stop
        sampling if it encounters an empty dataset. This preserves the sampling
        ratio throughout training. The prior behavior was to continue sampling,
        skipping over exhausted datasets, until all datasets are exhausted. By
        default, the original behavior (`stop_on_empty_dataset=False`) is