For instance:

- datasets for supervised classification, with a lot instances to simulate a data stream and many classes to simulate the playable actions,

- or datasets with contexts, actions and the reward of the played action, with a huge number of instances to use rejection sampling when the chosen action is not the played action.