PyTextAD benchmark datasets

Text anomaly detection datasets with a 0/1 anomaly label for every word, used by the PyTextAD library. A document is anomalous if any of its words is anomalous, so each dataset serves both token-level and document-level evaluation.

from pytextad.datasets import load_dataset     # pip install "pytextad>=0.4"
ds = load_dataset("restaurant_review")
ds.tokens, ds.token_labels, ds.labels

Dataset	Documents	Anomalous	Words	Anomalous words	Anomaly
sms_spam	4,518	393	82,197	424	injected gibberish
restaurant_review	1,100	50	36,222	859	negative sentiment
grammar_correction	300	30	2,746	47	grammatical errors
hate_speech	4,302	140	99,835	484	hateful or offensive words
olid	650	30	21,256	133	offensive words
restaurant_review2	520	25	33,764	208	negative sentiment

Each line of a file is one document: {"id", "text", "tokens", "labels"}, where labels has one 0/1 value per entry of tokens.

Word segmentation

Every document is split into words by the regular expression \w+|[^\w\s], and every word has its own label: an annotated span such as "not fresh" is two words, each labelled 1. The evaluation unit therefore does not depend on the labels.

Before PyTextAD 0.4.0 (commit 2af4a01 of this repository), spans in anomalous documents were stored as single items while normal documents were split into words. An item made of several words was then almost always anomalous, which inflates token-level scores that grow with the number of sub-words of an item. That version is kept in the history only for reference and should not be used for token-level evaluation. In the current version, every span is split with the same expression as the normal documents, each word keeping the span's label, and seven lone invisible characters (U+2066, U+200D) that produce no sub-word were removed from olid. Document counts and document labels are unchanged.

Sources

• sms_spam: SMS Spam Collection, taken from NLP-ADBench; meaningless character sequences were injected into some messages.
• restaurant_review: Google Maps reviews of a restaurant in the USA.
• grammar_correction: Kaggle grammar-correction.
• hate_speech: tweets from Davidson et al., Automated Hate Speech Detection and the Problem of Offensive Language, ICWSM 2017.
• olid: tweets from Zampieri et al., Predicting the Type and Target of Offensive Posts in Social Media (OLID), NAACL 2019.
• restaurant_review2: reviews of a second restaurant.

The word-level annotations are released under CC BY 4.0; the texts remain subject to the terms of their original sources.

Citation

If you use these datasets, please cite PyTextAD:

@software{cao2026pytextad,
  author = {Cao, Yang},
  title  = {{PyTextAD}: Text Anomaly Detection in Python},
  year   = {2026},
  url    = {https://github.com/charles-cao/pytextad}
}

sms_spam, restaurant_review and grammar_correction were introduced in Towards Token-Level Text Anomaly Detection (with spans stored as single items; see above); please also cite:

@inproceedings{cao2026tokenlevel,
  title     = {Towards Token-Level Text Anomaly Detection},
  author    = {Cao, Yang and Yu, Bicheng and Yang, Sikun and Liu, Ming and Yang, Yujiu},
  booktitle = {Proceedings of the ACM Web Conference 2026 (WWW '26)},
  year      = {2026},
  doi       = {10.1145/3774904.3792952}
}