Hinglish Dataset of Gender and Race Bias Prompts with Paired Biased and Neutral Responses

Description

This dataset (HGRB) contains 1,830 curated entries in Hinglish (Romanized Hindi-English code-mixed text) for studying gender- and race-related social bias. Each entry is a triplet consisting of an input prompt, a biased response that affirms or elaborates a stereotype, and a neutral response that rejects the stereotype and reframes the topic around individual skills, effort, and opportunity. Every entry also carries a category label: "gender" (931 entries) or "race" (899 entries), where "race" is used broadly to cover regional and linguistic community stereotypes, religion, caste, skin colour, and nationality. The dataset is released as a single deduplicated CSV file with five columns: id, input, biased response, neutral response, and category. There are no missing values, no duplicate input prompts, and no duplicate complete entries. All text is written in Roman script using informal, non-standardised Hinglish spelling, reflecting how such language is typically used in Indian social media, messaging, and chatbot interactions. This resource is intended for researchers and developers working on bias detection, fairness evaluation, and bias mitigation in language models and chatbots that operate in Hindi-English code-mixed settings. Because each prompt is paired with both a stereotype-affirming and a stereotype-free response, the dataset can be used to train and evaluate bias classifiers, to construct preference pairs for alignment/debiasing experiments (treating the neutral response as preferred and the biased response as rejected), and to probe whether Hinglish-capable language models tend to agree with stereotypes when prompted. The biased responses and many of the prompts contain stereotypical and potentially offensive statements about groups defined by gender, region, religion, caste, colour, and nationality. These are included solely to support research on detecting and reducing such bias; the authors do not endorse the views expressed in them.

Institutions

• Manipal University Jaipur

  Rajasthan, Jaipur

Categories

Licence