PIIBench is a unified benchmark dataset for PII detection across multiple domains.
OEach example contains:
tokens: list of tokenslabels: BIO labelssource: original data source of the sampletext: original/source text when available; deterministic token-joined text for token-only sources| Split | Records |
|---|---|
train.jsonl | 799,948 |
validation.jsonl | 99,990 |
test.jsonl | 100,002 |
Ten datasets are downloaded from Hugging Face and consolidated into a unified BIO-tagged format:
| Dataset | Rows | Domain |
|---|---|---|
| ai4privacy/pii-masking-400k | ~400k | General, 63 PII classes |
| ai4privacy/pii-masking-300k | ~300k | General + Finance (FinPII-80k) |
| gretelai/synthetic_pii_finance_multilingual | ~56k | Finance (100 doc types) |
| nvidia/Nemotron-PII | ~100k | General (50+ industries) |
| wikiann (en) | ~20k | Wikipedia, PER/ORG/LOC only |
| Babelscape/multinerd (en) | varies | Wikipedia + news, 15 types |
| DFKI-SLT/few-nerd | ~188k | Wikipedia, 66 fine-grained types |
| conll2003 | ~14k | News (Reuters), 4 types |
| nlpaueb/finer-139 | ~1.1M | Finance (SEC filings), 139 XBRL tags |
| Isotonic/pii-masking-200k | ~200k | General, 54 PII classes |
finer-139 is capped at 150k records during data preparation. Entity types with fewer than 500 B-mentions globally are collapsed to O.
The current release was regenerated with the PIIBench preparation pipeline's
--include-text option so each example includes a text field in addition to
the trainable token/label/source fields.
This dataset is derived from multiple sources. Users must comply with the original dataset licenses of the constituent datasets.
@article{jha2026piibench,
title={PIIBench: A Unified Multi-Source Benchmark Corpus for PII Detection},
author={Jha, Pritesh},
year={2026}
}
PIIBench is a unified benchmark dataset for PII detection across multiple domains.
OEach example contains:
tokens: list of tokenslabels: BIO labelssource: original data source of the sampletext: original/source text when available; deterministic token-joined text for token-only sources| Split | Records |
|---|---|
train.jsonl | 799,948 |
validation.jsonl | 99,990 |
test.jsonl | 100,002 |
Ten datasets are downloaded from Hugging Face and consolidated into a unified BIO-tagged format:
| Dataset | Rows | Domain |
|---|---|---|
| ai4privacy/pii-masking-400k | ~400k | General, 63 PII classes |
| ai4privacy/pii-masking-300k | ~300k | General + Finance (FinPII-80k) |
| gretelai/synthetic_pii_finance_multilingual | ~56k | Finance (100 doc types) |
| nvidia/Nemotron-PII | ~100k | General (50+ industries) |
| wikiann (en) | ~20k | Wikipedia, PER/ORG/LOC only |
| Babelscape/multinerd (en) | varies | Wikipedia + news, 15 types |
| DFKI-SLT/few-nerd | ~188k | Wikipedia, 66 fine-grained types |
| conll2003 | ~14k | News (Reuters), 4 types |
| nlpaueb/finer-139 | ~1.1M | Finance (SEC filings), 139 XBRL tags |
| Isotonic/pii-masking-200k | ~200k | General, 54 PII classes |
finer-139 is capped at 150k records during data preparation. Entity types with fewer than 500 B-mentions globally are collapsed to O.
The current release was regenerated with the PIIBench preparation pipeline's
--include-text option so each example includes a text field in addition to
the trainable token/label/source fields.
This dataset is derived from multiple sources. Users must comply with the original dataset licenses of the constituent datasets.
@article{jha2026piibench,
title={PIIBench: A Unified Multi-Source Benchmark Corpus for PII Detection},
author={Jha, Pritesh},
year={2026}
}