BiasNavi: LLM-Empowered Data Bias Management

Bias in datasets undermines the fairness, transparency, and reliability of AI systems, presenting critical challenges across applications. Existing tools for managing data bias often remain inaccessible to non-experts or struggle with the complexities of domain-specific datasets. In this work, we introduce BiasNavi, an large language models (LLM)-empowered toolkit for data bias management. BiasNavi features an autonomous agent that seamlessly integrates with modules for bias identification, measurement, surfacing, and adaptation, reasoning over data and interactions to adaptively guide users through the bias management pipeline. With intuitive and personalized interfaces, BiasNavi empowers users to customize their workflows and address data bias effectively. A case study with the COMPAS dataset demonstrates how BiasNavi, by leveraging the advanced reasoning capabilities of LLMs, democratizes responsible AI practices, making bias management both accessible and effective for users with varying levels of expertise. BiasNavi is available at: https://github.com/CIRES-Hub/BiasNavi.