Nemotron-Personas-Finance-Japan: A Synthetic Dataset for Japanese Finance

Using real customer data in finance is constrained by privacy regulations, compliance requirements, and the need for domain expertise. To avoid those risks while still reproducing realistic Japanese financial scenarios, Elith and NVIDIA have released Nemotron-Personas-Finance-Japan, a financial-domain extension of the open-source Nemotron-Personas-Japan dataset. The authors describe it as the first open persona dataset at million-record scale that pairs Japan’s official statistics with household finance and financial literacy data. It was generated with NeMo SDG-PGMs, a probabilistic graphical model framework for synthetic data generation, and NeMo Data Designer, and refined with input from Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance, and the Financial Data Utilization Association (FDUA).

The dataset contains 9M personas total—1M records with 9 personas each—written in natural Japanese. Every record has 50 fields: 9 persona fields, 11 persona attributes, and 30 contextual fields based on official demographic and labor statistics. That amounts to roughly 1.5B tokens, including about 750M persona tokens, with ~950k unique names and more than 1,600 occupation categories. Coverage spans demographic, regional, asset distribution, and personality trait axes; persona types include occupations, sports, arts, travel, cooking, finance, financial behavior, and financial concerns, and natural-language attributes cover cultural background, skills, career goals, hobbies, household/parental/financial background, attitudes toward money, and financial planning. The dataset is available under CC BY 4.0 for commercial and non-commercial use.

The generation pipeline uses NeMo Data Designer, described as a compound AI system that handles Jinja templating, Pydantic validation, structured outputs, automated retries, and multiple generation backends, with gemma-4-31B-it used for Japanese text generation. To make the data more useful for AI training, the authors deliberately refined official statistics: education is more granular than the national aggregates, occupation categories were expanded with business owners and specialized professions, and life stages that are underrepresented in statistics—students, retirees, unemployment—are explicitly modeled. Household-level financial and real assets are included alongside individual income, and internet/smartphone usage plus financial literacy help capture consumer profiles in financial use cases.

Every persona is fully synthetic, built from distributions rather than real records, so the dataset is designed to reduce privacy risk while reflecting official demographic and household-finance distributions. The primary audience is teams creating financial AI in Japan—banks, insurers, startups, and integrators—and the personas can be used to build customer-service agents, evaluate sales or marketing conversations, run simulations in NeMo User Sim, and red-team customer-facing AI. The Nemotron-Personas collection is designed to integrate with NeMo User Sim, allowing instant execution of simulations drawn from distributions consistent with the real world. The release also targets anyone building for Japanese users: compared with Nemotron-Personas-Japan, it has 50% more personas across 50 fields instead of 22, giving finer population slices for training and testing. Stated use cases include multi-turn conversation seeds, domain-specific assistant development, and bias or fairness evaluation across rural/urban, age, education, and economic dimensions.

A companion dataset, Nemotron-PII-Japan, launches the same day: 100,000 synthetic Japanese business documents with labeled personal information, built on Nemotron-Personas-Japan. The two are designed to be used together—simulate customers with Nemotron-Personas-Finance-Japan, and test PII detection and redaction pipelines with Nemotron-PII-Japan. The dataset can be downloaded directly with load_dataset(‘nvidia/Nemotron-Personas-Finance-Japan‘). The motivation behind the release is that enterprise AI has long relied on private data, limiting researchers, startups, and regional developers; this CC BY 4.0 release is meant to provide enterprise-grade synthetic data with lower privacy risk and support culturally grounded, Sovereign AI systems.

Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and Financial Context

View Original