NVIDIA announced Nemotron PII Japan on November 9, 2026: 100,000 synthetic Japanese business documents annotated for personally identifiable information (PII) detection. The release gives researchers a common corpus for comparing document-redaction systems, alongside explicit evaluation limits.
The announcement says it was released that day; the dataset card lists November 10.
What Nemotron PII Japan contains
The dataset covers 43 document types across healthcare, finance, government, employment, education, legal work, telecommunications and IT, and e-commerce. NVIDIA lists 56 native PII labels and 810,703 annotated spans. The documents are distributed in one Parquet file, with original text, character-position annotations and a version containing inline labels.
This is a different resource from the Nemotron Personas Finance Japan release. NVIDIA and Elith describe that separate dataset as one million financial persona records. PII Japan instead supplies documents for evaluating detection and redaction.
How the synthetic documents were generated
NVIDIA says synthetic personal information and business values are established before prose generation. Placeholders are replaced afterward, and the text is checked for consistency. That connects document generation to the annotations used for evaluation.
The workflow uses NeMo Data Designer. Its public repository documents a framework combining samplers, model-generated content, dependencies between fields and output validators. For university teams studying synthetic data, these materials expose the generation approach alongside the resulting corpus.
What the evaluation means for research teams
NVIDIA reports that fine-tuning GLiNER improved exact character-span-and-label micro F1 from 13.41% to 73.80% on Sumi test and from 8.13% to 65.30% on Sumi OOD. These are company-reported results, scored across eight common categories at a fixed threshold of 0.5.
For researchers and research offices assessing the release, three distinctions matter to computational reproducibility:
- Experimental split: the release has no fixed train/test split. NVIDIA used 95,000 training and 5,000 internal validation documents, using internal validation for checkpoint selection because the documents were highly similar.
- Annotation coverage: NVIDIA warns that labels may omit relevant information, so a correct detection can be scored as a false positive.
- Negative controls: every document has PII annotations. NVIDIA says separate PII-free documents are needed to evaluate false positives on documents without PII.
The dataset card identifies research, training, regression testing and red teaming as intended uses. It explicitly says synthetic-data performance does not guarantee privacy protection or complete detection on real data, and deployment requires additional validation.
Access and reuse
NVIDIA distributes the dataset under CC BY 4.0. The Creative Commons license summary permits sharing and adaptation, including commercial use, subject to attribution, a license link and an indication of changes.
Sources
- NVIDIA, Nemotron PII Japan announcement, November 9, 2026. Checked 2026-11-15.
- NVIDIA, Nemotron-PII-Japan dataset card. Checked 2026-11-15.
- NVIDIA and Elith, Nemotron-Personas-Finance-Japan announcement. Checked 2026-11-15.
- NVIDIA NeMo, Data Designer repository. Checked 2026-11-15.
- Creative Commons, Attribution 4.0 International. Checked 2026-11-15.
