Marginal Alignment Does Not Guarantee Joint-Distribution Fidelity: An Official-Reference Audit of Nemotron-Personas-Korea with Cross-Locale Replication
A new research paper introduces the Independence-Assumption Footprint (IAF) to audit synthetic persona datasets. The IAF method compares synthetic joint distributions against official demographic references, revealing discrepancies even when marginal distributions align. When applied to NVIDIA's Nemotron-Personas-Korea dataset, the audit found significant mismatches in key joint distributions related to occupation, age, and gender representation, despite alignment with marginal demographics. The study also highlights that these diagnostic findings are locale-dependent and can be confounded by reference taxonomy cardinality. AI
IMPACT Highlights critical data integrity issues in synthetic persona datasets, impacting their reliability for downstream AI applications.