CRD3 is a dataset for dialogue modeling and summarization released by Microsoft Research in 2020, designed to provide linguistically unique non-scripted role-playing game dialogue data for exploring challenges such as semantics, pragmatics, content selection, and rewriting in long-form story narration.

The dataset contains transcribed dialogue from 159 episodes of Critical Role, totaling 398,682 dialogue turns, along with corresponding summaries collected from the Fandom wiki. The data scale ranges from 10 K to 100 K, with a total data size of approximately 318 MB.

Dataset Composition

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

Each data sample includes the following fields:

• chunk: string type, containing the dialogue fragment text.
• chunk_id: integer type, fragment ID.
• turn_start: integer type, starting index of the dialogue turn.
• turn_end: integer type, ending index of the dialogue turn.
• alignment_score: float type, alignment score.
• turns: dictionary type, containing detailed dialogue information:
• names: list of strings, speaker names.
• utterances: list of strings, corresponding dialogue content.
• number: integer type, turn number.

Citation```bibtex

@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} }