• Description:

This version of the Wikipedia Toxicity Subtypes dataset provides access to the primary toxicity label, as well the five toxicity subtype labels annotated by crowd workers. The toxicity and toxicity subtype labels are binary values (0 or

1) indicating whether the majority of annotators assigned that attributes to the comment text.

The comments in this dataset come from an archive of Wikipedia talk pages comments. These have been annotated by Jigsaw for toxicity, as well as a variety of toxicity subtypes, including severe toxicity, obscenity, threatening language, insulting language, and identity attacks. This dataset is a replica of the data released for the Jigsaw Toxic Comment Classification Challenge on Kaggle, with the training set unchanged, and the test dataset merged with the test_labels released after the end of the competition. Test data not used for scoring has been dropped. This dataset is released under CC0, as is the underlying comment text.

See the Kaggle documentation or for more details.

Split Examples
'test' 63,978
'train' 159,571
  • Features:
    'identity_attack': tf.float32,
    'insult': tf.float32,
    'obscene': tf.float32,
    'severe_toxicity': tf.float32,
    'text': Text(shape=(), dtype=tf.string),
    'threat': tf.float32,
    'toxicity': tf.float32,
  • Citation:
  author = {Wulczyn, Ellery and Thain, Nithum and Dixon, Lucas},
  title = {Ex Machina: Personal Attacks Seen at Scale},
  year = {2017},
  isbn = {9781450349130},
  publisher = {International World Wide Web Conferences Steering Committee},
  address = {Republic and Canton of Geneva, CHE},
  url = {},
  doi = {10.1145/3038912.3052591},
  booktitle = {Proceedings of the 26th International Conference on World Wide Web},
  pages = {1391-1399},
  numpages = {9},
  keywords = {online discussions, wikipedia, online harassment},
  location = {Perth, Australia},
  series = {WWW '17}