NVIDIA releases synthetic Japan finance dataset for AI development
NVIDIA and Elith have released 9 million synthetic Japanese personas for financial AI development, simulation and evaluation without real customer data.
NVIDIA and Elith have released Nemotron-Personas-Finance-Japan, an open synthetic dataset designed to help developers build and test financial AI without using real customer records. The announcement was published by NVIDIA on Hugging Face on 9 November 2026.1
What has been released
The dataset contains 9 million synthetic Japanese-language personas arranged across 1 million records, with nine personas per record. Each record contains 50 fields covering demographics, work, household circumstances, assets, personality traits and financial context. NVIDIA says the collection contains roughly 1.5 billion tokens, about 950,000 unique names and more than 1,600 occupation categories.1
The project extends Nemotron-Personas-Japan with information intended to model financial situations and behaviour. NVIDIA says the work combines Japan's official statistics with household-finance and financial-literacy data, using its NeMo SDG-PGMs framework and NeMo Data Designer. The company also says feedback from Japanese financial institutions and industry organisations helped shape the dataset.1
Why it matters
Financial AI developers often cannot use real customer data for experimentation because of privacy, compliance and security requirements. Synthetic data does not remove every modelling risk, but it can provide a safer test environment for conversational agents, customer simulation and model evaluation. The dataset is intended to represent varied financial circumstances rather than reproduce identifiable people, and NVIDIA says the personas are not modelled on real individuals.1
The release is available under the CC BY 4.0 licence for commercial and non-commercial use. It is also connected to NVIDIA's broader Nemotron-Personas collection and can be used with NeMo User Sim for simulations based on the distributions represented in the data.1
Our opinion
This is more useful than another generic synthetic-data announcement because it targets a specific gap: realistic financial context without exposing bank customers. The important caveat is that “synthetic” does not automatically mean unbiased or representative. Developers will still need to test how well the dataset reflects the populations and financial behaviours they care about, and to document where generated assumptions enter their models. As a freely licensed, million-record starting point, though, the release gives researchers and product teams something concrete to inspect rather than another promise of future tooling.