metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data.parquet
task_categories:
  - text-generation
  - conversational
  - roleplaying
pretty_name: Deep Dungeons and Dragons
dataset_info:
  features:
    - name: thread_id
      dtype: int
    - name: message_id
      dtype: int
    - name: author_id
      dtype: int
    - name: author_num_posts
      dtype: int
    - name: message
      dtype: str
    - name: character
      dtype: str

Deep Dungeons and Dragons

A dataset of long-form multi-turn and multi-character collaborative RPG stories, complete with associated character cards.

This dataset comprises of 56,000 turns in 1544 stories following 9771 characters: a total of 50M Llama tokens.
Each turn comprises a multi-paragraph continuation of a story from the perspective of a defined character including both dialogue and prose.

This dataset is a cleaned and reformatted version of Deep Dungeons and Dragons, originally released in 2018 by Annie Louis and Charles Sutton and comprising of transcripts collected from public games at roleplayerguild.com. We've removed images and links (as well as their references) from posts to make this a text-only dataset, as well as anonymising usernames - although this is still available in the original dataset.

Schema

thread_id: The numeric ID of a specific RPG game.
message_id: The turn within a thread.
author_id: ID corresponding to the message author.
author_num_posts: The total number of posts by that author - may be usable as a quality signal.
message: The post containing a continuation of the RPG story roleplaying as a specific character. Most of the time, this is a multi-paragraph description containing both prose and dialogue.
character: The character card that this post is role-playing as. This is 1-to-1 mapped to (thread_id, author_id).