Cambridge restaurant dialogue domain dataset collected for developing neural network based dialogue systems. The two papers published based on this dataset are: 1. A Network-based End-to-End Trainable Task-oriented Dialogue System 2. Conditional Generation and Snapshot Learning in Neural Dialogue Systems The dataset was collected based on the Wizard of Oz experiment on Amazon MTurk. Each dialogue contains a goal label and several exchanges between a customer and the system. Each user turn was labelled by a set of slot-value pairs representing a coarse representation of dialogue state. There are in total 676 dialogue, in which most of the dialogues are finished but some of dialogues were not.