BCCard/pii-masking-openpii-finance

1. Overview

Korean·English financial-domain PII detection dataset for fine-tuning token-classification models, derived from ai4privacy/pii-masking-openpii-1.5m and re-labeled to an 18-class policy taxonomy, then augmented with synthetic finance/card/bank/insurance/security context. Every PII value is synthetic - either invalidated by construction or inherited from the synthetic ai4privacy corpus - so no real personal data is present (see Synthetic-Data Safety).

1.1. TL;DR

• 71,788 rows - train (57,494) / validation (14,294)
• 18 PII labels → 73 BIOES token classes (4N+1), tokenized with the o200k tokenizer (openai/privacy-filter base)
• Bilingual - en-share held to a 20-30% band (26.8% here); every label carries en coverage (US-format generators for en)
• 100% synthetic PII values - none drawn from real records (invalidated-by-construction for synthesized rows; inherited from the synthetic ai4privacy corpus otherwise); not pseudonymized personal data (see §2)
• Hard negatives built in - surface-similar non-PII decoys (label O) + label-confusion pairs (RRN↔FRN, CARD↔ACCOUNT)

1.2. Splits

split	rows	ko	en	en-share	role
train	57,494	42,072	15,422	26.8%	training input
validation	14,294	10,460	3,834	26.8%	HPO / checkpoint selection

Composition by source (train): openpii-1.5m-ko 28,572 (refined + perturbed) / domain 16,300 (synthesized) / openpii-1.5m-en 12,622 (replay). validation mirrors the same proportions (fixed seed = 42).

1.3. Schema

9 columns, all splits. Parallel-array form (NER-style HF Datasets convention) - 1 row is a self-contained example.

column	type	description
source_text	string	original (unmasked) text - the model input
masked_text	string	text with each PII span replaced by [LABEL_n] placeholders
privacy_mask	list<struct>	span annotations - each {label, start, end, value} over source_text
language	string	ko / en
split	string	train / validation (provenance; redundant with the parquet split)
uid	string	unique row id (encodes source + language + split + cluster)
source_dataset	string	openpii-1.5m-ko / openpii-1.5m-en / domain
tokens	list<string>	o200k subword tokens of source_text
token_classes	list<string>	per-token BIOES label (73 classes: B-/I-/E-/S- × 18 + O)

privacy_mask offsets index into source_text; token_classes is the token-aligned BIOES projection of those spans and is verified to round-trip against privacy_mask at build time.

1.4. Label Taxonomy (N=18)

Labels are reverse-derived from the masking policy so that model label = logging key = policy key (single source: label-taxonomy.yaml v1). data source marks where examples come from: ko/en (from openpii-1.5m) or domain (synthesized).

label	description	data source
PERSON	full name (surname + given, single span)	ko, en
RRN	resident registration number (Korea)	ko, domain
FRN	foreign registration number	domain
CARD_NUMBER	credit/debit card PAN	ko, en, domain
ACCOUNT_NUMBER	bank account number	domain
SECRET	auth secret (password / API key / token)	domain
USER_ID	online member ID	domain
EMAIL	email address	ko, en
PHONE	phone number (mobile / landline)	ko, en
PASSPORT	passport number	ko, en
DRIVER_LICENSE	driver's license number	ko, en
GENERIC_ID	generic identifier (no KO counterpart)	ko, en
ADDRESS	address (city / street / building, single span)	ko, en
ZIPCODE	postal code	ko, en
DATE	date / time	ko, en
CARD_EXPIRY	card expiry date	domain
CVC	card verification code	domain
IPIN	I-PIN number	domain

1.5. Quick Start

from datasets import load_dataset

ds = load_dataset("BCCard/pii-masking-openpii-finance")
train, val = ds["train"], ds["validation"]

row = train[0]
row["source_text"]      # model input
row["token_classes"]    # BIOES targets, aligned to row["tokens"]
row["privacy_mask"]     # [{"label": "USER_ID", "start": 30, "end": 38, "value": "..."}, ...]

Train a token classifier on (tokens, token_classes); report span-level precision / recall / F1 on validation.

1.6. Methodology

Built by the data_preprocessing module of the MoAI Privacy Filter pipeline. Four stages (playbooks/data_preprocessing/data-preprocessing-strategy.md §3.4.2):

ai4privacy/pii-masking-openpii-1.5m (ko + en)
   ↓ (1) refine     re-label to the 18-class taxonomy; sanitize values (RRN/card invalidation)
   ↓ (2) replay     carry the en subset (all-label coverage) into the 20-30% en band
   ↓ (3) synthesize domain templates × value generators (finance/card/bank/insurance/security)
   │                 · method A: mention-level perturbation of refined ko rows
   │                 · method B: static template × value-generator synthesis (no LLM at runtime)
   ↓ (4) mix        concat + shuffle(seed=42) → BIOES derivation → integrity + RRN audit
This dataset (train / validation)

• Value generators synthesize PII values that pass shape/format checks but are provably invalid or unassigned (see §2.2).
• Templates are statically authored (no LLM at generation time); {LABEL} draws a fresh value per occurrence, {LABEL_n} reuses one value (ai4privacy numbering convention).
• Value pool is md5-partitioned per split (train / validation) with empty pairwise intersection, enforced by a disjointness assertion.

2. Synthetic-Data Safety

All PII values are synthetic - none are drawn from any real individual's records. Safety rests on two provenance regimes, and it is worth stating precisely which applies where:

regime	rows	how values are kept safe
Inherited (openpii-1.5m-ko / -en)	~72%	values are the upstream ai4privacy corpus (LLM / faker-generated, CC-BY-4.0), re-labeled. They carry real-looking domains and formats (e.g. gmail.com emails, real-format phone numbers) but do not derive from any real record - synthetic by inheritance. RRN/FRN and CARD_NUMBER are additionally invalidated (§2.2)
Synthesized (domain)	~28%	values our generators produce - invalid or unassigned by construction (§2.2): reserved domains, unassigned prefixes, failing checksums

Note: the openpii-1.5m-ko bucket mixes two paths sharing that source_dataset tag - refined rows keep inherited faker values, whereas perturbed rows (method A, §1.6) have their PII re-generated by our value generators, so they are invalid-by-construction like domain. source_dataset does not separate the two, so the genuinely real-looking (inherited-faker) share is smaller than the ~72% row suggests (all en replay + only the refined part of ko).

2.1. Why this is not personal data

Under Korea's Personal Information Protection Act (PIPA), a value is personal information only if it (1) matches a real individual's real value and (2) singles that individual out, alone or via easy linkage. The claim fails on two independent grounds:

• Not sourced from real records. Every value is synthetic - generated by our rule/faker pipeline, or inherited from the synthetic ai4privacy corpus. There is no original record behind any value, so this is not pseudonymized personal data; there is nothing to re-identify.
• No linkage even on a coincidental match. If a synthetic value (say a faker mailbox at gmail.com) happens to equal a real one, the rest of the row - name, address, phone - is independently synthetic, so the tuple points to no real person. Almost every row also contains at least one structurally impossible value (checksum-failing RRN/card, unassigned-prefix phone, reserved-domain email) that marks the whole row as synthetic, collapsing any linkage attempt. This is the same basis on which ai4privacy itself is publicly distributed.

By label, identifiability breaks down further: names / addresses / dates / zipcodes do not single out a person alone (namesakes; a place is not a person); account-number / user-id / generic-id / secret / I-PIN / CVC / card-expiry are random values whose owners sit in non-public registries, so a value alone looks nobody up. The residual - a synthetic value coincidentally equal to a real one, with no way to link it - is the accepted, industry-standard risk of any synthetic-PII corpus, identical to the risk the upstream public dataset already carries.

2.2. Per-label invalidation

Two tiers. The first is applied in the refine stage to all sources (inherited and synthesized); the second is a construction property of synthesized (domain) rows only - for inherited rows these labels carry ai4privacy faker values and rely on the §2.1 identifiability argument.

Invalidated across all sources (public checksums - a mathematical guarantee):

label	mechanism
RRN, FRN	old-scheme unification + checksum digit +5 (deliberately fails the check digit; holds regardless of birth band)
CARD_NUMBER	Luhn check digit +5, for values ≥13 digits (real-PAN length)

Structural, synthesized (domain) rows only (inherited rows use faker values here):

label	mechanism
EMAIL	RFC 2606 reserved domains (example.com/net/org)
PHONE	unassigned prefix (leading 0/1 for ko; 555-01XX for en)
DRIVER_LICENSE	unassigned region code
PASSPORT	unissued prefix (avoidance - may weaken over time)

RRN is unified to the pre-2020.10 scheme so invalidity holds regardless of birth band (only the post-2020.10 number-change program leaves a documented residual). CVC is never generated standalone - always alongside an (invalid) PAN. The Luhn invalidation targets values ≥13 digits; a handful of ≤12-digit inherited values may still pass Luhn, but 12 digits is below real card-network length, so they are not real PANs.

2.3. Hard negatives

Two designed confusion types raise the decision boundary:

• Type 1 - surface-similar non-PII (decoy, label O): tokens that look like an identifier but are not (e.g. an order amount, a product code), left unlabeled so the model learns not to over-mask.
• Type 2 - label-confusion pairs: shape-adjacent values across labels (RRN ↔ FRN, CARD_NUMBER ↔ ACCOUNT_NUMBER) placed in disambiguating context so the model learns the correct class rather than the shape.

3. Distribution Stats

3.1. Language × split

train + validation = 71,788 rows; en-share 26.8% (target band 20-30%).

language	train	validation	total
ko	42,072	10,460	52,532
en	15,422	3,834	19,256
total	57,494	14,294	71,788

3.2. Label distribution

306,842 annotations across the 18 labels (train + validation).

label	train	validation	total
PERSON	43,407	10,569	53,976
DATE	41,880	10,205	52,085
ADDRESS	33,513	8,404	41,917
EMAIL	22,061	5,440	27,501
GENERIC_ID	20,302	5,012	25,314
PHONE	17,691	4,270	21,961
CARD_NUMBER	14,154	3,469	17,623
ZIPCODE	12,574	3,191	15,765
DRIVER_LICENSE	9,264	2,229	11,493
PASSPORT	7,346	1,837	9,183
RRN	6,176	1,507	7,683
USER_ID	2,986	728	3,714
FRN	2,800	696	3,496
SECRET	2,607	648	3,255
ACCOUNT_NUMBER	2,606	655	3,261
CVC	2,600	646	3,246
CARD_EXPIRY	2,600	646	3,246
IPIN	1,700	423	2,123

4. Limitations

• Domain: Korean·English financial services (card / bank / insurance / security context) - narrow domain; generalization to other domains is not guaranteed
• Synthetic distribution shift: synthesized rows follow authored templates, so surface variety is bounded by the template set - real-world phrasing diversity exceeds it
• PASSPORT safety is avoidance-based: the unissued-prefix guarantee can weaken if issuing authorities later allocate those prefixes - unlike the checksum/Luhn labels it is not mathematically permanent (flagged honestly)
• RRN residual: pre-2020.10 unification makes checksum invalidity time-independent, but the post-2020.10 number-change program is a residual case not fully covered
• IPIN value spec pending: I-PIN values are generated at pilot volume only, pending a finalized value-format confirmation
• Inherited values are real-looking: ~72% of rows carry ai4privacy faker values at real domains/formats (e.g. gmail.com emails, real-format phones, a few ≤12-digit Luhn-passing card values). These are synthetic (not from real records) but are not structurally invalidated - their safety rests on the §2.1 identifiability argument, not §2.2 construction. Upstream annotation conventions and residual noise also propagate into this portion

5. Meta Info

5.1. Source & License

Derived from ai4privacy/pii-masking-openpii-1.5m (CC-BY-4.0). Released under CC-BY-4.0; attribution to ai4privacy is required. Tokenized with the o200k tokenizer of the openai/privacy-filter base model; taxonomy = label-taxonomy.yaml v1 (18 labels / 73 BIOES classes).

5.2. Citation

@dataset{bccard_moai_privacy_filter_finance,
  title  = {MoAI Privacy Filter Finance (PII Masking, ko/en)},
  author = {BC Card AX Team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BCCard/pii-masking-openpii-finance},
  note   = {Derived from ai4privacy/pii-masking-openpii-1.5m (CC-BY-4.0)}
}

5.3. See Also

• Base model (fine-tuning target): openai/privacy-filter - the model this dataset fine-tunes; source of the o200k tokenizer used for tokens / token_classes
• Upstream dataset: ai4privacy/pii-masking-openpii-1.5m - the source corpus this dataset is derived from