PLURAL: A Global Dataset for Value Alignment

PLURAL is a value-focused preference dataset derived from the Integrated Values Survey (IVS), a nationally representative survey spanning 92 countries. Its two-stage generation pipeline converts survey responses into realistic synthetic preference triplets while attempting to preserve normative signals. The initial release contains approximately 500,000 triplets covering 20 countries. The authors report that training with PLURAL reduces mean absolute error against target-country value profiles by up to 27.7% relative to strong baselines. A blind human evaluation with 176 evaluators in India, Brazil, and Japan also rated PLURAL-aligned responses as more representative of national values.