hate-speech-and-offensive-language

Why we included this project

This repository holds the labeled Twitter dataset behind the Davidson et al. 2017 ICWSM paper, which separated hate speech from merely offensive language, a distinction that still informs how moderation systems are built and evaluated. Researchers and ML engineers working on abusive-language detection get the data as CSV or pickle files, plus the lexicon and a classifier script that reproduces the paper's three-way classification. The bundled notebook shows how the labels were derived, which helps you understand the dataset's assumptions before training on it. Note that the code targets Python 2.7 and is no longer maintained, so treat this as a data and methodology reference rather than a drop-in production pipeline. The authors' follow-up paper on racial bias in this and similar datasets is also worth reading before you rely on the labels.