CRD3 is a dataset released by Microsoft Research in 2020 for dialogue modeling and abstractive summarization. It aims to provide linguistically unique unscripted role-playing game conversation data to explore challenges related to semantics, pragmatics, content selection, and rewriting within long-form narrative storytelling.

The dataset comprises transcribed dialogues from 159 episodes of Critical Role, totaling 398,682 conversational turns. Accompanying summaries were collected from the Fandom Wiki. Each summary ranges between 10K and 100K tokens, resulting in an approximate total dataset size of 318 MB.

Dataset Composition

The dataset consists of three main splits: train, validation, and test. The specific sample counts are as follows: the training set includes 38,969 samples, the validation set contains 6,327 samples, and the testing set holds 7,500 samples.

Each data sample contains the following fields:

• chunk: String type containing the dialogue snippet text.
• chunk_id: Integer type representing the fragment ID.
• turn_start: Integer type indicating the starting index of the conversational turn.
• turn_end: Integer type indicating the ending index of the conversational turn.
• alignment_score: Float type representing the alignment score.
• turns: Dictionary type containing detailed information about the conversation:
  • names: List of strings containing speaker names.
  • utterances: List of strings containing the corresponding dialogue content.
  • number: Integer type denoting the turn number.

Citation

@inproceedings{
title = {Storytelling with Dialogue: A Critical Role Dungeons and Dragons Dataset},
author = {Rameshkumar, Revanth  and Bailey, Peter},
year = {2020},
publisher = {Association for Computational Linguistics},
conference = {ACL}
}