README.md

metadata

language:
  - en
license: cc-by-nc-4.0
size_categories:
  - 100K<n<1M
task_categories:
  - conversational
  - text-generation
dataset_info:
  - config_name: default
    features:
      - name: username
        dtype: string
      - name: char_name
        dtype: string
      - name: bio
        dtype: string
      - name: context
        list:
          - name: text
            dtype: string
          - name: username
            dtype: string
          - name: char_name
            dtype: string
      - name: reply
        dtype: string
      - name: has_nameless
        dtype: bool
      - name: char_confidence
        dtype: float64
    splits:
      - name: train
        num_bytes: 1921588254
        num_examples: 140469
    download_size: 764073630
    dataset_size: 1921588254
  - config_name: grammar_filtered
    features:
      - name: username
        dtype: string
      - name: char_name
        dtype: string
      - name: bio
        dtype: string
      - name: context
        list:
          - name: char_name
            dtype: string
          - name: text
            dtype: string
          - name: username
            dtype: string
      - name: reply
        dtype: string
      - name: char_confidence
        dtype: float64
    splits:
      - name: train
        num_bytes: 371438765
        num_examples: 27053
    download_size: 166606326
    dataset_size: 371438765
  - config_name: high_confidence
    features:
      - name: username
        dtype: string
      - name: char_name
        dtype: string
      - name: bio
        dtype: string
      - name: context
        list:
          - name: text
            dtype: string
          - name: username
            dtype: string
          - name: char_name
            dtype: string
      - name: reply
        dtype: string
      - name: has_nameless
        dtype: bool
      - name: char_confidence
        dtype: float64
    splits:
      - name: train
        num_bytes: 949419370.7676569
        num_examples: 69403
    download_size: 386317057
    dataset_size: 949419370.7676569
  - config_name: pruned
    features:
      - name: username
        dtype: string
      - name: char_name
        dtype: string
      - name: bio
        dtype: string
      - name: context
        list:
          - name: text
            dtype: string
          - name: username
            dtype: string
          - name: char_name
            dtype: string
      - name: reply
        dtype: string
      - name: has_nameless
        dtype: bool
      - name: char_confidence
        dtype: float64
    splits:
      - name: train
        num_bytes: 782484734.2032762
        num_examples: 57200
    download_size: 326987882
    dataset_size: 782484734.2032762
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
  - config_name: grammar_filtered
    data_files:
      - split: train
        path: grammar_filtered/train-*
  - config_name: high_confidence
    data_files:
      - split: train
        path: high_confidence/train-*
  - config_name: pruned
    data_files:
      - split: train
        path: pruned/train-*
tags:
  - roleplay
  - not-for-all-audiences

Data scraped from roleplayerguild and parsed into prompts with a conversation history and associated character bio. Thanks to an anonymous internet stranger for the original scrape.

As usernames can be associated with multiple character biographies, assignment of characters is a little fuzzy. The char_confidence feature reflects how likely this assignment is to be correct. Not all posts in the conversation history necessarily have an associated character name. The column has_nameless reflects this.

Each row should fit into 4096 Llama tokens, depending on your prompt format - there's built in slack of 128 tokens + 8 per message.

There are a few configurations available. I highly recommend not using the default configuration as it contains a lot of questionable quality data. The options, in order of increasing usefulness:

• default - ocean of garbage with some gems
• high_confidence - only entries with no nameless posts that are highly likely to be assigned a correct char_name/bio
• pruned - Further filtered from high_confidence to remove common types of junk replies
• grammar_filtered - run through a grammar checker to remove rows with too many mistakes

The grammar_filtered configuration is almost certainly what you want to be using. (Unless you want to do your own processing and filtering.)