Dataset at a glance
Elith and NVIDIA announced Nemotron-Personas-Finance-Japan, a synthetic persona dataset intended for financial AI development in Japan. According to the release, the collection includes 9M personas (described as 1M records × 9 personas each), spans 50 fields per record, and totals roughly 1.5 billion tokens, including about 750 million persona tokens.
The dataset contains ~950,000 unique names and more than 1,600 occupation categories. Each record combines 9 persona fields, 11 persona attributes and 30 contextual fields derived from official demographic and labor statistics. Personas incorporate attributes such as education, household finances, digital and financial literacy, hobbies, and attitudes toward money.
How it was built
The team says the dataset was generated using NeMo Data Designer and NeMo SDG-PGMs, a probabilistic graphical model framework for statistically grounded synthetic data generation. The pipeline also used gemma-4-31B-it for Japanese text generation and supports complex templating, validation and multiple generation backends.
Elith and NVIDIA report that the design aligns persona distributions with Japan’s official statistics while adding finer granularity for education, expanded occupation categories, and life-stage scenarios such as retirement and unemployment. Names follow real Japanese naming patterns but, the publishers emphasize, all personas are synthetic and built from distributions rather than records about real people.
Use cases, partners and availability
The release notes collaboration and feedback from domestic financial institutions and associations, including Sompo Japan, Tokio Marine Holdings, Bank of Fukuoka, Mizuho Financial Group, Meiji Yasuda Life Insurance and the Financial Data Utilization Association (FDUA). The dataset is intended for tasks such as conversational agent training, customer simulation, model evaluation, bias testing and red-teaming customer-facing AI.
Nemotron-Personas-Finance-Japan is available under the CC BY 4.0 license and is cleared for commercial use. The publishers also launched Nemotron-PII-Japan the same day, a separate set of 100,000 synthetic Japanese business documents with labeled personal information for detection and redaction testing.
To download the persona dataset, the release provides this example:
from datasets import load_dataset
ds = load_dataset("nvidia/Nemotron-Personas-Finance-Japan")
Original source: Hugging Face Blog