As described in the paper, DECODE includes 6 groups of dialogues: Train, Dev, Test, Human-Bot, A2T, RCT.

The format of the file is JSONL. Each line in the file is one dialogue example saved in a JSON.
Primary fields that are required for the contradiction detection task:
- record_id: It is the unique ID for the example.
- turns: The field contains a list of turns that presents a conversation between two speaker.
- is_contradiction: The field indicates the label of the example. true means that the last turn (turns[-1]) contradicts some turns in the dialogue history. false means that the last turn is not a contradiction.
- aggregated_contradiction_indices: The field is a list of the indices that gives the location of the supporting evidence for the contradiction (w.r.t. to the turns list). Notices that the turn_id of the last turn (turns[-1]) will always be in this list.

Other fields that is related to the contradiction task:
- num_of_turns_by_writer: The field indicates how many turns the annotator created for providing a contradiction utterance.
- writer_contradiction_indices: The original contradiction indicates provided by the annotator.
- verifications: For each examples, we asked another three annotator for verification. This field gives the results of the verification.

Field not described are not required for the contradiction detection task.
