end-to-end dialog systems, in which all components are learnt simultaneously, have recently obtained encouraging successes. However these were mostly on conversations related to chit-chat with no clear objective and for which evaluation is difficult. This paper proposes a set of tasks