NagaYu/sumi-ja-pii-corpus

Sumi — synthetic Japanese PII corpus

This dataset contains no real personal information. Every name, address, phone number, email address, date of birth, bank account, member ID and My-Number-shaped digit string is generated with a seeded RNG and inserted into the text while recording the offsets, so the gold spans are correct by construction rather than recovered by searching afterwards.

Built for Sumi, a Japanese PII detector.

Why this dataset exists

Japanese PII detection fails in a specific way: surnames that are also ordinary nouns (森 "forest", 林 "woods", 泉 "spring", 大和, 青木), place names that are also surnames (長野, 福島, 千葉, 山口), company names built from surnames, honorific boundaries (様 / さん / 氏 / 殿, and the many uses of 様 that are not personal at all), part numbers that look like phone numbers, and facility names that look like addresses. This corpus makes those confusions the point rather than an accident.

Important notes

• No real personal data. Surnames are drawn from an approximation of the public frequency distribution; addresses combine real public place names with randomised block/lot numbers; email addresses only ever use reserved domains (example.com, example.co.jp, example.ne.jp, .test, .invalid, …).
• Identifiers with check digits are format-valid and value-invalid by construction. Card-shaped numbers deliberately fail Luhn and My-Number-shaped numbers deliberately fail the official check-digit algorithm, so no usable number can be produced. The project's test suite asserts this over thousands of samples.
• This dataset does not certify compliance with any law or regulation. It is material for training and evaluating detectors, and makes no completeness claim.

Splits

split	documents	gold spans
train	12000	107426
validation	1500	13496
test	2000	17899
ood	520	1964
negatives	2000	0

negatives contains zero gold spans. Anything a detector finds there is by definition a false positive, which is what makes the false-positive rate measurable.

ood is the template-independent evaluation split. train / validation / test are generated from business-document templates (email, meeting minutes, application forms, support tickets), so measuring only on test would reward a model that merely memorised the template slots. ood inserts the same synthetic PII directly into Wikipedia / statute / literary prose, with each carrier sentence phrased so the label is genuinely valid. Report both.

Gold spans by type (train + validation + test)

type	Japanese	count
NAME	氏名	63778
ADDRESS	住所	11003
PHONE	電話番号	21979
EMAIL	メールアドレス	12925
DOB	生年月日	6881
BANK_ACCOUNT	金融口座	4955
CREDIT_CARD	カード番号	2029
MYNUMBER	マイナンバー様式	1301
MEMBER_ID	会員番号	7078
POSTAL_CODE	郵便番号	6892

Hard-negative kinds

Deliberately confusable material, mixed into the positive documents and forming the entire negatives split.

kind	count
address_like_facility	2706
common_noun_surname	2735
company_as_person	2769
date_like_nondob	2798
honorific_boundary	2690
number_like_id	2798
phone_like_id	2741
place_as_person	2655

• common_noun_surname — surnames that are also ordinary nouns (森 / 林 / 泉 / 大和 / 青木)
• place_as_person — prefecture and city names that are also common surnames, used as places
• company_as_person — company names built from surname morphemes
• honorific_boundary — 様 / さん / 氏 / 殿 with and without a person, including お客様, 皆様, 仕様, 同様, 様式 where 様 is not a person at all
• phone_like_id — part numbers, order numbers, ISBNs and dates shaped like phone numbers
• address_like_facility — civic buildings and stations that read like addresses
• number_like_id — bill numbers, statute numbers, quantities that look like account numbers
• date_like_nondob — contract, issue and meeting dates that are not birth dates

Base text licences

Synthetic PII is inserted into openly-licensed Japanese prose. Each row records its source and licence in source_license / source_ref.

source	licence	url	note
wikipedia_ja	CC BY-SA 4.0	https://ja.wikipedia.org/
egov_law	Public Domain (Japanese Copyright Act, Article 13)	https://laws.e-gov.go.jp/
aozora	Public Domain (Aozora Bunko, copyright expired)	https://www.aozora.gr.jp/

The synthetic parts (templates, PII values, hard negatives) are CC0-1.0. Rows containing Wikipedia-derived prose inherit CC BY-SA 4.0, which is why the dataset as a whole is published under CC BY-SA 4.0.

Fields

• doc_id (string), text (string)
• spans: JSON-encoded list of {start, end, label, text, score, source, meta}. start / end are Python character indices into the NFKC-normalised text, half-open [start, end) — not byte offsets.
• subset, genre, source_license, source_ref, negative_kinds, meta

Loading

from datasets import load_dataset
import json

ds = load_dataset("NagaYu/sumi-ja-pii-corpus")
row = ds["test"][0]
spans = json.loads(row["spans"])
for s in spans:
    assert row["text"][s["start"]:s["end"]] == s["text"]

Reproduce

python3 scripts/build_dataset.py --train 12000 \
  --val 1500 --test 2000 \
  --neg 2000 --seed 0