Toxic-Bias Reasoning Dataset

The ToxicBias-Reasoning dataset is designed for bias detection and reasoning generation in online discourse, with a special focus on Indian sociocultural contexts. Each record includes a raw comment text, a binary bias label, one or more bias categories, and a natural language rationale explaining the classification. This structure enables training and evaluation of models not only for prediction but also for explainable AI in bias-sensitive applications.Columns:comment_text \u2013 Raw input text containing potential bias.category \u2013 One or more assigned categories: {caste, religion, race, gender, LGBTQ+, political} or none (non-bias).reason \u2013 Natural language rationale justifying the assigned category\/label.bias \u2013 Binary label: 1 = biased, 0 = non-biased.Dataset Statistics:Total Instances: 7,562Biased: 5,818Non-biased: 1,744Category Distribution: Race (2,211), Religion (1,783), Gender (651), LGBTQ+ (577), Political (774), Caste (247).Splits:Train\/Validation: Human-annotated labels + GPT-4o\u2013assisted human-in-the-loop rationales (5-fold cross-validation provided).Test: Fully human-annotated labels and rationales for high-quality evaluation.Use Cases:Binary Bias Detection (bias vs. non-bias).Multilabel Category Classification (detecting overlapping biases).Reasoning Generation (training\/evaluating explainable models that provide justifications).