Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

01 / AI · Claim vs. Check · Analysis

Elith and NVIDIA release 9M synthetic Japan finance personas

With 50 fields per record, the dataset targets financial AI development without real customer records, but the companies did not publish how closely it matches official statistics.

Empty customer consultation counters with vacant chairs line a Japanese bank branch after closing, as a staff member walks away out of focus in the background. AI-generated image

Elith and NVIDIA released 9M synthetic Japanese financial personas, or fictional person profiles, under a CC BY 4.0 licence on 9 Nov 2026 for teams building financial AI without real customer records. The dataset is called Nemotron-Personas-Finance-Japan. The Hugging Face announcement states that commercial use is permitted.

The two companies said privacy regulation and compliance requirements make real data hard to use in the financial sector. In their account, developing financial agents requires synthetic data that carries demographics, household structures and asset status.

The announcement leads with the rise to 50 fields per record and the exclusion of real customer records. It contains no figure measuring how closely the data matches official statistics.

Japanese population data expands to 50 fields

The announcement's contents list gives 1M records with nine personas each, in company-published figures. Each record has 50 fields, and the contextual fields are based on official demographic and labour statistics. Tokens total about 1.5B, and occupation categories number more than 1,600.

The two companies said they added household-level financial and real assets to individual income and spending propensity. Internet and smartphone use and financial literacy are also included. Education bands that official statistics group broadly were divided more finely, and students, retirees and periods of unemployment were modelled.

The dataset has 50% more personas than its predecessor, Nemotron-Personas-Japan, and fields rose from 22 to 50, according to figures the two companies published. The companies wrote that it is, to their knowledge, the first open persona dataset at this scale to pair Japan's official statistics with household finance.

A generation process that excludes real customer records

The two companies said they used NeMo SDG-PGMs, a probabilistic graphical model framework, for statistics-based generation. The model gemma-4-31B-it handled Japanese text generation. The full pipeline was built with NeMo Data Designer.

The announcement states that the personas were built from statistical distributions, "not from records about real people". It describes the dataset as designed to reduce privacy risk, framing that reduction as an aim. It reports no measurement of re-identification risk, the chance that a synthetic record can be linked to a real person.

The checks that remain after financial firms' input

The announcement states that feedback from financial firms including Sompo Japan and Mizuho Financial Group informed the design. It does not say that these institutions use the dataset in live services.

The announcement lists customer-service agents, evaluation of sales conversations and vulnerability checks on customer-facing AI as uses. Nemotron-PII-Japan, released the same day, labels personal information in 100,000 synthetic Japanese business documents. The companies' guidance is to use the persona dataset for customer simulation and Nemotron-PII-Japan for checking privacy protection.

The two companies rest their case on alignment with official statistics. The announcement does not say how the distributions for the finer education bands were set, and it carries no performance figures for financial AI built on the dataset. What it supplies is material for developing and evaluating such systems.