NVIDIA opens Japanese datasets for financial agents and privacy testing

NVIDIA has released synthetic data for two practical problems facing Japanese-language AI: simulating customers with different financial circumstances, and finding personal information before a system exposes it.

The announcements, dated 9 November, cover Nemotron-Personas-Finance-Japan and Nemotron-PII-Japan. Both are downloadable under Creative Commons Attribution 4.0, which permits commercial use with attribution. The first contains a million synthetic records with nine persona descriptions each. The second contains 100,000 synthetic business documents annotated for personal information.

For developers, the attraction is access to training and test material without putting actual customer files into the development process. The releases also show how much work remains between a plausible synthetic example and a trustworthy production system.

A more varied simulated customer

The financial dataset, co-developed with Japanese AI company Elith, combines demographic statistics with household-finance and financial-literacy distributions. Each record has 50 fields, covering matters such as employment, income, household assets and financial concerns. Its nine million persona descriptions represent nine views of each of a million records, rather than nine million distinct customers.

NVIDIA says the data can support conversational-agent training, customer simulation and evaluation through its NeMo User Sim framework. A developer could use different profiles to test whether an assistant explains the same product appropriately to people with different levels of financial literacy. That is a proposed application, not evidence that an agent trained on the data gives suitable financial advice.

The construction matters. Statistical distributions constrain the generated profiles, but the dataset still makes simplifying assumptions about relationships between attributes. Matching population statistics does not establish that every generated life story is realistic or that every group is represented adequately.

Testing what a redaction system misses

The separate PII dataset covers eight domains and 43 document types, with 56 labels for information such as names, addresses and account identifiers. Synthetic values are inserted into generated document text, producing labelled examples without collecting real customer records. This release uses the broader Nemotron-Personas-Japan dataset, rather than the new financial extension.

NVIDIA reports fine-tuning the multilingual GLiNER entity-detection model on 95,000 documents, with 5,000 reserved for internal validation. On separate Sumi evaluations, the fine-tuned model scored 73.80% on the business-document test and 65.30% on the out-of-distribution test, compared with 13.41% and 8.13% before fine-tuning.

Those are exact-span-and-label micro-F1 scores across eight shared categories, a measure combining precision and recall. They are vendor-reported results, not independent certification. The external evaluation sets were excluded from training, checkpoint selection and threshold selection, according to NVIDIA.

The much higher 98.08% internal validation result served a different purpose: selecting the checkpoint. It should not be read as a 98% success rate on real customer documents. Even after fine-tuning, reported recall on the external business-document test was 65.10%, leaving substantial room for missed information under that test’s scoring rules.

The datasets give teams inspectable, reusable material for building their own tests. They do not certify privacy, regulatory compliance or the quality of financial recommendations. The PII dataset’s documentation expressly warns that synthetic-data performance does not guarantee complete detection in deployment.

Source note: Both announcement pages display 9 November 2026. The financial dataset card lists its creation/version date as “11/10/2026”; that date discrepancy remains unresolved.

Featured image: NVIDIA’s Santa Clara headquarters sign, photographed by Will Buckner on 31 October 2020. CC BY 2.0. Cropped for display.