ToxiGen

Large machine generated dataset of implicit hate speech across 13 minority groups.

Made by MicrosoftDatasets and benchmarks

About ToxiGen

This repo contains the code for generating the ToxiGen dataset, published at ACL 2022. The data, methods and trained checkpoints are intended to be used for research purposes only.

The dataset contains implicitly toxic and benign sentences mentioning 13 minority groups. ToxiGen is available on HuggingFace. This repository was archived by the owner on Jul 15, 2026, and it is now read-only.

Drafted from github.com · not yet confirmed by Microsoft

How it works

ToxiGen is a dataset of about 274,000 statements generated by prompting a large model to produce subtle, implicit hate speech and matched benign statements about 13 minority groups, plus a human labeled subset and a RoBERTa classifier trained on it. Its value is implicit toxicity, the kind keyword and slur based systems miss, and the balanced benign statements that help reduce false positives on identity terms.