README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: description
      dtype: string
    - name: group
      dtype: int64
    - name: domain
      dtype: string
  splits:
    - name: train
      num_bytes: 5586867
      num_examples: 1229
  download_size: 3369452
  dataset_size: 5586867
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

cols:

• text: str, the text that is being described in
• description: str, generated w/ a 5-shot prompt w/ hand-written descriptions, by deepseek v3.1 base
• group: int, unused in this dataset, always 0
• domain: str, the domain of the text, can be general, textfile, code, ao3, misc, metalure

domains*:

• general (400 samples) taken from crumb/askmistral-pile-2-15
• textfile (400 samples) taken from midwestern-simulation/textfiles (textfiles.com)
• code (200 samples) taken from bigcode/the-stack-smol
• ao3 (200 samples) taken from midwestern-simulation-active/ao3_random_subset
• misc (100 samples) taken from midwestern-simulation/random_texts
• metalure (66 samples) taken from https://metalure.neocities.org at date 9/25/2025

*we then filtered all blank generations out, and filtered all that were too long (improper prompt adherance, reiterating document for example)