NVIDIA announced 100,000 synthetic Japanese business documents on November 9, 2026—November 10 in Japan—for teams testing personal-information removal before AI processing.
The collection, Nemotron-PII-Japan, is training and evaluation data, not a ready-made redaction service. NVIDIA says it generated the documents instead of collecting customer or patient records.
It is publicly available on Hugging Face under CC BY 4.0, which permits commercial reuse with credit, a licence link and disclosure of changes.
Testing what gets removed—and what remains
According to the dataset card, the collection covers 43 document types across eight business areas, including finance, healthcare and employment. Each record has text and annotations marking selected information and its position.
NVIDIA’s customer-service example involves removing names and contact details before an external AI summarises an inquiry. The useful test has two parts: whether identifying details survive, and whether excessive masking strips out the substance of the request.
Why one score is not enough
Every document contains annotated personal information. NVIDIA therefore says separate examples are needed to measure false alarms on documents with no personal information. Teams can still examine unnecessary removal within the supplied documents; they cannot use this collection alone to establish how often a wholly clean document will be flagged.
The labels are also incomplete, NVIDIA warns. A detector can correctly find an identifier that the reference labels missed, yet automatic scoring may count that detection as an error. That means a lower score can sometimes reflect a labelling gap rather than worse detection.
For a business comparing suppliers, the practical implication is to ask for missed identifiers and unnecessary removals to be reported separately, with disputed detections reviewed by a person. Clean examples should match the business’s own rules about what must be removed, and tests should cover the document types it actually handles.
The release is text data, not a test of an entire scanned-document workflow. If the intended inputs are scans or PDFs, text extraction and the final exported file need their own checks. NVIDIA explicitly says good synthetic-data results do not guarantee privacy protection in real use.
For teams choosing where to run AI document systems, RohitAI’s guide to Mistral Studio’s deployment and data controls explains the separate decisions about model processing, stored sessions and logs.