Abstract: The subject technology generates a dataset based at least in part on a set of files. The subject technology generates, utilizing a machine learning model, a set of labels corresponding to the dataset. The subject technology filters the dataset using a set of conditions to generate at least a subset of the dataset. The subject technology generates a virtual object based at least in part on the subset of the dataset and the set of labels, where the virtual object corresponds to a selection of data from the dataset. The subject technology trains a second machine learning model using the virtual object and at least the subset of the dataset, where training the second machine learning model includes utilizing streaming file input/output (I/O), the streaming file I/O providing access to at least the subset of the dataset during training.