ToxicBias-Reasoning dataset

Overview The ToxicBias-Reasoning dataset is a culturally contextualised benchmark for bias detection and reasoning generation in online discourse, with emphasis on Indian sociocultural contexts. Each record contains: Comment Text – the input text. Binary Bias Label – whether the text is biased (1) or non-biased (0). Bias Categories – one or more labels from: caste, religion, race, gender, LGBTQ+, political. Reason – a natural language explanation justifying the assigned category and label. This structure enables both classification and explanation generation, supporting research in explainable AI (XAI) and socially responsible NLP. Dataset Structure Columns Column Description Example comment_text Raw input text (social media style, may include informal or offensive terms) "The enemy is greater than a friend who is a farmer who is in distress...!!" category One or more bias categories: {caste, religion, race, gender, LGBTQ+, political} or "none" political reason Natural language rationale for the classification "This statement is political bias because it frames all farmers as distressed." bias Binary label: 1 = biased, 0 = non-biased 1 Dataset Statistics Total Instances: 7,562 Biased: 5,818 Non-biased: 1,744 Category Distribution: Race: 2,211 Religion: 1,783 Gender: 651 LGBTQ+: 577 Political: 774 Caste: 247 Splits Training + Validation: Human-annotated labels. Rationales generated using GPT-4o with human-in-the-loop refinement. Provided in 5-fold cross-validation format. Test Set: Fully human-annotated labels and rationales. Intended for final evaluation.