Under 'hotel-final-slots.json, for each hotel-booking only dialog, we have dialogwhich is the processed merged dialog,steps-summarywhich is the summarized steps of the dialog by InstructGPT,final-statewhich is the final ground truth state dictionary (from the dataset),binary-dictwhich is the slots converted to binary stored in a dictionary, andbinary-array` which is the binary values in binary-dict listed in an array.
Under hotel-slot-changes.json, for each hotel-booking dialog, we record the ground truth slot information change every time it's service's turn to speak. They are derived from the ground truth dialog state annotations.
(1) Under summaries-hotel folder is over 2k of summaries of dialogs from MultiWOZ 2.4 that is hotel-related
(2) Under summaries-hotel-booked is the subset of (1) where the booking was completed during the dialog (about 1.3k dialogs)
(3) Under summaries-hotel-only is the subset of (2) where there is no discussion of other types of service during the conversation (so hotel booking only). (230 dialogs)
(4) Under summaries-hotel-only-all is all dialogs from MultiWOZ 2.4 that satisfies the same criterion as (3), i.e. hotel-only and completion of booking. (432 dialogs)
(5) Under summaries-hotel-only-all2 is same dialogs as (4) but summaries generated with a different prompt (that seems to make InstructGPT less likely to make up steps that dont exist in dialogs)
See slides https://docs.google.com/presentation/d/1fSVIjLmEJyzBSvE9dNkO4d9uR7AMEcVHv8QR_usl9Tw/edit?usp=sharing for examples of (3)