NoisyHate: Hate Speech Detection with Text Perturbations Analyzed
A NoisyHate (hate speech benchmark) dataset of 1,339 labeled sentences with text perturbations and toxicity scores, collected by NoisyHate Authors via ICWSM 2025. Each record contains sentence text, binary hate label, toxicity scores (toxicity, severe, obscene, identity_attack, insult, threat), identity group indicators, and engagement metrics. The dataset is notable for its benchmark dataset with human-written text perturbations and fine-grained toxicity metrics. Intended for research into hate speech detection, model robustness, adversarial testing, toxicity analysis, and identity-based hate speech research.