The Bias Pipeline Dataset: Connecting Data Production, Annotation Disagreement, and Model Fairness Across Modalities

Riabi et al. (2024) show how annotation variation and socio-demographic cues shape model predictions; Sekkat et al. (2024) demonstrate the value of controlled demographic labels and multivariate tests in speech systems. We propose a new “bias pipeline” dataset spanning text, audio, and image tasks in a shared domain (e.g., professional profile retrieval or content moderation), with three key innovations: (1) rich annotator metadata and repeated measures to model disagreement and its correlates; (2) explicit logs of curation decisions (filtering, balancing, prompt designs for LLM labeling), enabling causal attribution of bias; and (3) user- and subject-level demographics to evaluate multivariate fairness. We add “power annotations” (inspired by Barabas et al., 2020) to capture institutional role asymmetries, letting researchers test power-aware metrics from Idea 3. The benchmark includes protocols to compare technical debiasing (e.g., fairness-aware sampling/SMOTE-style balancing in tabular credit data: Le Duy Quang et al., 2025; fairness-aware imaging pipelines: Sufian et al., 2024) with socio-educational interventions (awareness raising: Jundan Wang, 2024) and governance patterns (Bahangulu & Owusu-Berko, 2025). This goes beyond existing datasets by linking the entire data-production chain to fairness outcomes, allowing researchers to answer questions like: Which stage contributes most to disparate impact? Do annotator-training programs outperform post-hoc debiasing? The impact is a shared resource that accelerates evidence-based standards for dataset curation, documentation, and auditability across sectors.

References:

1. Algorithmic bias, data ethics, and governance: Ensuring fairness, transparency and compliance in AI-powered business analytics applications. Julien Kiesse Bahangulu, Louis Owusu-Berko (2025). World Journal of Advanced Research and Reviews.
2. Studying up: reorienting the study of algorithmic fairness around issues of power. Chelsea Barabas, Colin Doyle, JB Rubinovitz, Karthik Dinakar (2020). FAT*.
3. An Empirical research of the Synergistic Reduction of Educational AI Algorithmic Bias by Technological Measures and Educational Awareness Raising. Jundan Wang (2024). ICAIE.
4. Beyond Dataset Creation: Critical View of Annotation Variation and Bias Probing of a Dataset for Online Radical Content Detection. Arij Riabi, Virginie Mouilleron, Menel Mahamdi, Wissam Antoun, Djamé Seddah (2024). International Conference on Computational Linguistics.
5. Sonos Voice Control Bias Assessment Dataset: A Methodology for Demographic Bias Assessment in Voice Assistants. Chloe Sekkat, Fanny Leroy, Salima Mdhaffar, Blake Perry Smith, Yannick Estève, Joseph Dureau, A. Coucke (2024). International Conference on Language Resources and Evaluation.
6. Mitigating Algorithmic Bias in Credit Scoring: A CNN-SMOTE Framework. Le Duy Quang, Nguyen Quang Dat, Ngo Dai Phong, Doan Tien Ban (2025). Asian journal of mathematics and computer research.
7. Mitigating Algorithmic Bias in AI-Driven Cardiovascular Imaging for Fairer Diagnostics. Md. Abu Sufian, Lujain Alsadder, Wahiba Hamzi, Sadia Zaman, A. S. M. S. Sagar, Boumediene Hamzi (2024). Diagnostics.