CoolFace
Datasetpublic

akaruineko/fantastic-offensive

fantastic-offensive dataset Binary offensive-language classification dataset combining several public sources and augmented with an obfuscation engine (leet-speak, separators, censoring, repeated chars, case shuffle, unicode homoglyphs, fullwidth) so classifiers learn to detect censored / mutated curse words. Schema column type meaning text string input sentence label int8 1 = offensive, 0 = clean source string originating dataset origin_label… See the full description on the dataset page: https://huggingface.co/datasets/akaruineko/fantastic-offensive.

sourceHugging Facecc0-1.0updated 1mo agoView on Hugging Face
0likes91downloads
Dataset Card

fantastic-offensive dataset

Binary offensive-language classification dataset combining several public sources and augmented with an obfuscation engine (leet-speak, separators, censoring, repeated chars, case shuffle, unicode homoglyphs, fullwidth) so classifiers learn to detect censored / mutated curse words.

Schema

columntypemeaning
textstringinput sentence
labelint81 = offensive, 0 = clean
sourcestringoriginating dataset
origin_labelstringlabel as given by the source
split_originstringoriginal train/test split of the source
mutatedint81 = obfuscation-engine variant
variantint8variant index within the source row
clusterint64near-duplicate cluster id (split safely)

Splits

  • —train: 2,456,268 rows
  • —validation: 20,000 rows
  • —test: 30,001 rows
  • —test_obfuscated: 14,062 rows

Sources

  • —Jigsaw Toxic Comment (tcapelle/jigsaw-toxic-comment-classification-challenge)
  • —Davidson et al. hate/offensive (contemmcm/hate-speech-and-offensive-language)
  • —ToxiGen (toxigen/toxigen-data, human toxicity scores)
  • —HateXplain (Hate-speech-CNERG/hatexplain)
  • —Wikipedia neutral filler (wikimedia/wikipedia 20231101.en)
  • —4chan casual conversation (fuzzy-g/4chan_pol_whole_ds)

Design notes

  • —Both classes receive orthographic mutations so obfuscation itself is not a cue for offense (prevents false positives on innocent leet like ex4mp1e).
  • —Split is performed at the near-duplicate cluster level so a mutated variant never leaks across train/validation/test.
  • —test_obfuscated holds mutated rows of the test set for robustness evaluation.

Disclaimer

This dataset contains raw offensive language. It is intended for research and moderation-model training only.