IndiCASA: A Dataset and Bias Evaluation Framework in LLMs Using Contrastive Embedding Similarity in the Indian Context

2025-10-03

AI Summary

This paper addresses the challenge that existing embedding-based bias evaluation methods fail to effectively detect implicit social biases—along five dimensions (caste, gender, religion, disability, and socioeconomic status)—embedded in large language models (LLMs) within India’s multicultural context. To this end, we propose IndiBias, the first fine-grained, India-contextualized bias evaluation framework. Our contributions are threefold: (1) We construct IndiCASA, a manually validated, culturally adapted dataset comprising 2,575 context-aligned sentence pairs; (2) We introduce a contrastive learning–based embedding similarity quantification method, augmented with counterfactual samples, to measure multidimensional bias; and (3) We systematically evaluate leading open-source LLMs, revealing that disability-related bias is most pronounced, while religious bias is comparatively lowest. These findings underscore the critical importance of localized, culturally grounded bias assessment frameworks for advancing model fairness and equity in diverse sociolinguistic settings.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Large Language Models (LLMs) have gained significant traction across critical domains owing to their impressive contextual understanding and generative capabilities. However, their increasing deployment in high stakes applications necessitates rigorous evaluation of embedded biases, particularly in culturally diverse contexts like India where existing embedding-based bias assessment methods often fall short in capturing nuanced stereotypes. We propose an evaluation framework based on a encoder trained using contrastive learning that captures fine-grained bias through embedding similarity. We also introduce a novel dataset - IndiCASA (IndiBias-based Contextually Aligned Stereotypes and Anti-stereotypes) comprising 2,575 human-validated sentences spanning five demographic axes: caste, gender, religion, disability, and socioeconomic status. Our evaluation of multiple open-weight LLMs reveals that all models exhibit some degree of stereotypical bias, with disability related biases being notably persistent, and religion bias generally lower likely due to global debiasing efforts demonstrating the need for fairer model development.

Problem

• Evaluating embedded biases in LLMs within Indian cultural contexts
• Assessing nuanced stereotypes across caste, gender, religion, disability and socioeconomic status
• Developing framework using contrastive embedding similarity for fine-grained bias detection

Innovation

• Contrastive learning encoder for bias evaluation
• Human-validated dataset across five demographic axes
• Embedding similarity captures fine-grained cultural biases