NVIDIA releases 9M-persona synthetic dataset for Japanese financial AI

caolonghua · x · 2026-11-10

NVIDIA and Elith have released Nemotron-Personas-Finance-Japan, a synthetic persona dataset for financial AI use cases in Japan.

• 1 million records with nine personas each (9 million total), 50 fields per record
• Fully synthetic: demographic and household finance data follows official statistics and public surveys, not real individuals
• Built with NeMo SDG-PGMs and NeMo Data Designer, using gemma-4-31B-it for Japanese text generation
• Suggested uses: seeding multi-turn dialogue datasets, building financial AI assistants, user simulation, and testing model performance across population groups
• Licensed under CC BY 4.0 for commercial and non-commercial use

Original post →