BCCard/privacy-filter-openpii-masking

1. Overview

privacy-filter-openpii-masking is a Korean and English entity-detection dataset for fine-tuning token-classification models. It is derived from ai4privacy/pii-masking-openpii-1.5m, relabeled to a 29-label taxonomy, and supplemented with statically authored or contextualized financial, customer-service/VOC, security, identity, and infrastructure scenarios.

The dataset provides entity annotations rather than application-specific redaction output. masked_text replaces every annotated span with a placeholder, including the non-PII disambiguation labels PORT and ORGANIZATION. A consuming system must decide how each detected label is handled after inference.

All rows are designed as synthetic training records. Values are either inherited from the synthetic ai4privacy corpus or generated locally under the controls described in Section 3. Public organization aliases are used only for the non-PII ORGANIZATION label. No operational sentence, identifier value, or complete customer record is copied into the artifacts.

1.1. TL;DR

• 72,375 rows - 57,851 train and 14,524 validation rows.
• 333,741 annotated spans across train and validation.
• 29 model labels - 117 BIOES token classes under 4N+1.
• Korean and English - English represents 25.74% of train and 26.03% of validation.
• 1024-token limit - the observed maximum is 758 tokens in train and 801 in validation, with no truncation drops.
• All labels in both languages and splits - every one of the 29 labels has ko and en support in train and validation.

1.2. Quick Start

from datasets import load_dataset

dataset = load_dataset("BCCard/privacy-filter-openpii-masking")
train = dataset["train"]
validation = dataset["validation"]

row = train[0]
row["source_text"]      # Unmasked model input.
row["tokens"]           # o200k subword token strings.
row["token_classes"]    # BIOES targets aligned with tokens.
row["privacy_mask"]     # Character-offset span annotations.

Train a token classifier using tokens and token_classes. The dataset stores token strings, not model input IDs. Convert them with the tokenizer described in Section 5.1. The published token sequence was produced with add_special_tokens=False; the 1024-token limit applies before any special tokens are added. If a model requires special tokens, add them in the training collator and assign an ignored label such as -100 to those positions without changing the published token-to-label alignment. Use validation metrics for checkpoint selection and experiment comparison, and assess final generalization on an independently held-out evaluation set.

To reproduce the published token strings, load openai/privacy-filter with a fast tokenizer and call it with add_special_tokens=False and truncation=False.

2. Data Preparation

2.1. Pipeline

The v1 release combines relabeled ai4privacy rows with locally generated perturbation and synthesis rows.

ai4privacy ko raw
  -> refine: relabel, merge spans, normalize Korean text, and apply safety controls
  -> perturbation: replace selected spans with split-specific generated values

ai4privacy en raw
  -> refine
  -> replay: retain rows for labels with at most 30 spans, then fill the quota with seed-42 sampling

static ko/en templates
  -> synthesis: materialize one row per authored template with local value generators

refine + perturbation + replay + synthesis
  -> mix: concatenate, shuffle with seed 42, tokenize, derive BIOES classes, and audit
  -> train / validation

The main rules for relabeling the upstream ai4privacy source labels - 37 labels, the union of 36 Korean and 34 English source labels - into the v1 taxonomy are:

• GIVENNAME and SURNAME are merged into a full PERSON span.
• CITY, STREET, and BUILDINGNUM are merged into ADDRESS; ZIPCODE remains separate.
• Korean TAXNUM values matching the resident-number form map to RRN; other values fall back to GENERIC_ID.
• SOCIALNUM maps to SSN in both languages.
• ORGANISATION, BANKNAME, and HOSPITALNAME map to ORGANIZATION.
• TIME maps to DATE, while TIMEZONE remains O.
• Source labels outside the label taxonomy in Section 2.3 are removed from privacy_mask and learned as O.

2.2. Splits

split	rows	ko	en	spans	maximum tokens
train	57,851	42,963 (74.26%)	14,888 (25.74%)	266,846	758
validation	14,524	10,743 (73.97%)	3,781 (26.03%)	66,895	801
total	72,375	53,706 (74.21%)	18,669 (25.79%)	333,741	801

The final mix consists of the following pipeline components.

component	train	validation	augmentation_type	source_dataset
ko refine	21,038	5,230	none	openpii-1.5m-ko
ko perturbation	13,038	3,230	perturbation	openpii-1.5m-ko
ko synthesis	8,887	2,283	synthesis	domain
en replay	12,623	3,138	none	openpii-1.5m-en
en synthesis	2,265	643	synthesis	domain
total	57,851	14,524	-	-

source_dataset records the source bucket, while augmentation_type records the transformation path. For example, a perturbed Korean row remains in openpii-1.5m-ko because its sentence originated upstream, even though its selected values were generated locally.

Across both splits, the source buckets contain 42,536 openpii-1.5m-ko rows, 15,761 openpii-1.5m-en rows, and 14,078 domain rows - locally authored synthesis rows covering financial, customer-service/VOC, security, identity, and infrastructure scenarios.

2.3. Label Taxonomy

The 29 labels produce 117 BIOES classes: B-, I-, E-, and S- for every label, plus O.

label	definition
PERSON	Full personal name as one span.
RRN	Korean resident registration number.
FRN	Korean foreign resident registration number.
SSN	Social-security-number family inherited from ai4privacy SOCIALNUM; not limited to the US 9-digit form.
GENERIC_ID	Identity-card or tax identifier that cannot be assigned to a more specific country-level label.
CARD_NUMBER	Credit or debit card PAN.
ACCOUNT_NUMBER	Bank account number.
SECRET	Password, API key, access token, or similar authentication secret.
USER_ID	Online account or member identifier.
EMAIL	Email address, including intentionally obfuscated forms in targeted synthesis rows.
PHONE	Mobile or landline telephone number, including intentionally verbalized forms in targeted synthesis rows.
PASSPORT	Passport number.
DRIVER_LICENSE	Driver's license number.
ADDRESS	City, street, and building components merged into one address span.
ZIPCODE	Postal code kept separate from ADDRESS.
DATE	Date or time; timezone-only strings remain O.
CARD_EXPIRY	Payment-card expiration date.
CVC	Card verification code, generated only with a card context.
IPIN	Korean I-PIN identifier.
TRANSACTION_APPROVAL_ID	8-digit payment authorization or transaction approval identifier.
BUSINESS_ID	10-digit business registration number or 9-digit merchant identifier.
VIRTUAL_CARD_NUMBER	11-digit alternate or virtual card number.
CI	Korean identity-linkage information value.
IPADDRESS	IPv4 network address.
MACADDRESS	48-bit MAC address.
IMEI	15-digit mobile-equipment identifier.
PORT	Network service port in the range 0 to 65535; a non-PII disambiguation label.
ORGANIZATION	Company, bank, hospital, or other organization name; a non-PII disambiguation label.
URL	Full web URL, including path, query, and fragment when present.

The taxonomy defines annotation labels. O means that no entity span is annotated at that position; it does not by itself mean that the text is non-sensitive or safe to retain. Applications can handle detected labels according to their own requirements.

2.3.1. Label Distribution

label	train	validation	total
PERSON	45,718	11,417	57,135
DATE	44,319	10,922	55,241
ADDRESS	38,152	9,556	47,708
EMAIL	25,374	6,372	31,746
PHONE	20,625	5,122	25,747
CARD_NUMBER	14,839	3,672	18,511
ZIPCODE	14,475	3,676	18,151
GENERIC_ID	14,335	3,512	17,847
DRIVER_LICENSE	9,914	2,405	12,319
SSN	9,371	2,412	11,783
PASSPORT	8,019	2,015	10,034
RRN	7,168	1,797	8,965
ORGANIZATION	2,113	588	2,701
IPADDRESS	1,103	305	1,408
TRANSACTION_APPROVAL_ID	1,066	308	1,374
USER_ID	1,060	309	1,369
BUSINESS_ID	992	268	1,260
ACCOUNT_NUMBER	966	277	1,243
PORT	958	217	1,175
URL	805	215	1,020
MACADDRESS	783	203	986
CI	761	225	986
SECRET	655	214	869
IPIN	645	201	846
FRN	600	158	758
IMEI	544	135	679
CARD_EXPIRY	522	138	660
VIRTUAL_CARD_NUMBER	508	136	644
CVC	456	120	576
total	266,846	66,895	333,741

The distribution is intentionally not uniform. Common free-form labels require more examples, while format-constrained labels use smaller targeted sets. Per-label strict Precision, Recall, and F1 must still be measured after training, particularly for the least frequent labels.

2.4. Synthesis and Hard Examples

Synthesis uses statically authored templates and local value generators. Each static template contributes one synthesis row to one split, and templates are not reused for value-only variants.

The initial 12,500 authored train and validation templates used the following design targets.

axis	target distribution
purpose	70% positive, 20% confusion, 10% pure hard negative
cue level	20% explicit, 35% contextual local, 30% contextual distant, 15% adversarial
unique labels per row	0 to 7 labels, with 2-label and 3-label rows forming the largest groups
repeated labels	15% of labeled rows repeat at least one label
long documents	20% of synthesis rows, with 3 to 6 unique labels

The published synthesis subset contains 14,078 rows: 12,500 initial authored rows, 1,528 contextualized legacy-template rows, and 50 independently authored email-disclosure rows. Its cue-level distribution is shown below.

final cue level	rows	share
explicit	2,500	17.76%
contextual_local	5,903	41.93%
contextual_distant	3,778	26.84%
adversarial	1,897	13.47%
total	14,078	100.00%

Each cue level describes what evidence identifies a span as an entity.

• explicit - a direct field-name cue accompanies the value, as in application forms, tables, and logs. This is the only group where such direct hints are allowed.
• contextual_local - the action, role, or relationship in the same clause identifies the value, without a direct field-name hint.
• contextual_distant - later sentences or the document-wide event flow identify the value. Long synthesis documents are placed in this group.
• adversarial - confusable values, O decoys, and other labels appear in the same context, and direct hints for the target label are forbidden. Pure hard-negative rows belong to this group.

Hard examples include:

• DATE versus CARD_EXPIRY.
• CARD_NUMBER or PHONE versus VIRTUAL_CARD_NUMBER.
• DRIVER_LICENSE versus TRANSACTION_APPROVAL_ID.
• GENERIC_ID versus BUSINESS_ID.
• SECRET versus CI.
• IPADDRESS versus URL.
• MACADDRESS versus IMEI.
• PORT versus ZIPCODE or an unlabeled operational number.
• PERSON or ADDRESS versus ORGANIZATION.
• Order numbers, product codes, industry codes, amounts, status phrases, and application settings as O decoys.

2.5. Long-Context Coverage

The financial, customer-service/VOC, and CI targeted subset intentionally places target spans at the front, middle, and end of long documents.

split	supplement rows	long rows	long share	front	middle	end	last-96-token rows
train	241	154	63.90%	37.35%	16.27%	46.39%	58
validation	60	39	65.00%	35.00%	12.50%	52.50%	17

These figures describe the targeted subset, not the entire dataset. They establish data coverage only; they do not establish long-context model accuracy.

2.6. Schema

All rows have ten columns. privacy_mask uses half-open character offsets over source_text: source_text[start:end] == value.

column	type	description
source_text	string	Unmasked model input.
masked_text	string	Annotation view with every labeled span replaced by a [LABEL_n] placeholder. Numbering follows left-to-right span order and restarts at 1 for each label in each row.
privacy_mask	list<struct>	Span records containing label, start, end, and value.
language	string	ko or en.
split	string	train or validation.
uid	string	Unique row identifier.
source_dataset	string	openpii-1.5m-ko, openpii-1.5m-en, or domain.
augmentation_type	string	none, perturbation, or synthesis.
tokens	list<string>	o200k subword token strings for source_text.
token_classes	list<string>	BIOES label aligned one to one with tokens.

masked_text is a rendering of the span annotations, not application output. Consumers can retain PORT or ORGANIZATION, partially mask selected labels, or apply channel-specific handling after inference.

3. Synthetic-Data Safety

The dataset is designed as synthetic training data, not as pseudonymized operational data. Its safety argument relies on provenance, independent generation, lack of linkage to a real source record, and additional structural controls where public validation rules exist.

3.1. Row Transformation Paths and Value Provenance

The following counts describe row transformation paths. They do not claim that every value in a row has the same provenance.

row path group	rows	share	interpretation
Rows retaining an upstream sentence structure without perturbation - ko refine and en replay	42,029	58.07%	Values primarily originate in ai4privacy, although refine locally transforms qualifying RRN and card values.
Rows with local replacement or full synthesis - ko perturbation and ko/en synthesis	30,346	41.93%	Synthesis is fully local. Perturbation replaces selected target spans but can retain other upstream values.
total	72,375	100.00%	-

Value provenance must therefore be interpreted per span rather than by assigning one provenance to the whole row. The row path groups are also not identical to source_dataset. Perturbed Korean rows retain source_dataset=openpii-1.5m-ko because their sentence scaffold originated upstream, even though selected values were generated locally.

Public organization aliases and public geographic vocabulary can appear as non-personal reference material. They are not linked to an actual individual's record.

3.2. Structural Controls and Audits

The following controls are applied where a public algorithm or a reserved range makes a meaningful check possible.

value family	control
RRN and locally generated FRN	Generate a valid structural body, then alter the check digit so the public checksum fails.
CARD_NUMBER	Alter the Luhn check digit for plausible PAN lengths covered by the final audit.
EMAIL	Use reserved example.com, example.net, and example.org domains on locally generated paths.
North American PHONE	Use the reserved 555-01XX range on locally generated paths.
IPADDRESS	Use documentation-only IPv4 networks on locally generated paths.
MACADDRESS	Generate locally administered unicast addresses.
IMEI	Alter the final digit so the generated value fails Luhn validation.
DRIVER_LICENSE and PASSPORT	Use avoidance rules and unassigned formats where defined; these are not permanent mathematical guarantees.

Current mix audits found:

• 8,965 RRN spans and zero checksum-valid values.
• Zero RRN values in the excluded post-2020 birth-band construction.
• 18,511 CARD_NUMBER spans and zero Luhn-valid values at lengths of 13 digits or more.
• 17 inherited card-number values shorter than 13 digits that pass Luhn - 14 in train and 3 in validation. These are a known limitation and rely on synthetic provenance.

For identifiers backed by private registries, such as account, telephone, passport, user, and social-security identifiers, actual issuance cannot be exhaustively checked. Their safety basis is synthetic provenance and the absence of linkage to a real person, not a claim that every generated surface form is impossible in the real world.

3.3. Split Independence and Integrity

The published artifacts passed the following structural checks.

• No duplicate UID or exact source_text across train and validation.
• No exact completed-value overlap between synthesis train and synthesis validation for any of the 29 labels. This compares literal span-value strings within the same label without case, punctuation, or format normalization.
• No span offset or value mismatch across 333,741 annotations.
• No taxonomy violation, split-field mismatch, or token/BIOES length mismatch.
• No row-level mismatch between the JSONL and Parquet build artifacts.
• No row dropped for exceeding 1024 tokens.

4. Limitations

• No model-performance claim. This card describes data construction and integrity. Precision, Recall, F1, and boundary accuracy remain to be measured after training.
• Synthetic distribution shift. Authored text is cleaner and more explanatory than many production VOC messages, logs, OCR fragments, and email chains. Some long rows contain phrases that explicitly explain why a value should or should not be treated as an identifier.
• Rare-label imbalance. CVC, VIRTUAL_CARD_NUMBER, CARD_EXPIRY, IMEI, and FRN have far fewer spans than PERSON, DATE, and ADDRESS. Format regularity can reduce the required volume, but adequate generalization is not established by count alone.
• Organization boundary ambiguity. The v1 taxonomy labels companies, banks, hospitals, and other legal organizations as ORGANIZATION, while brands, products, services, and personal trade names can remain O. Real text does not always make that distinction explicit.
• Inherited-value realism. Upstream ai4privacy values can resemble real identifiers. Except for audited checksum-controlled families, safety rests on synthetic provenance and absence of source linkage.
• Korean translationese. The refined Korean source retains sentence structures inherited from English-oriented upstream templates. Perturbation preserves those scaffolds, and the resulting effect on natural Korean generalization has not been measured.
• Short card-number residual. The 17 inherited values described in the structural audit section are not covered by the current 13-or-more-digit Luhn invalidation guarantee.
• I-PIN format maturity. IPIN is a pilot-format subset and may not cover all production variants.
• Domain coverage. Financial services, customer service/VOC, identity, security, and infrastructure contexts are emphasized. Performance in unrelated domains is not guaranteed.

5. Reproducibility and Metadata

5.1. Release Metadata

This card describes dataset version v1. The published token strings were produced with the fast tokenizer associated with the openai/privacy-filter model, using add_special_tokens=False and truncation=False. The sequence-length limit during dataset construction was 1024 tokens before special tokens.

Padding and special-token insertion are not included in the published alignment. Configure them in the training collator and preserve one label for each published token.

5.2. Source and License

This dataset is derived from ai4privacy/pii-masking-openpii-1.5m, distributed under CC BY 4.0. This derivative dataset is also released under CC BY 4.0, and attribution to ai4privacy is required.

Token strings and BIOES targets use the o200k tokenizer associated with the openai/privacy-filter base model.

5.3. Citation

@dataset{bccard_privacy_filter_openpii_masking_v1,
  title  = {Privacy Filter OpenPII Masking: Korean and English Entity Detection Dataset},
  author = {BC Card},
  year   = {2026},
  url    = {https://huggingface.co/datasets/BCCard/privacy-filter-openpii-masking},
  note   = {Version 1; derived from ai4privacy/pii-masking-openpii-1.5m under CC BY 4.0}
}

5.4. Related Artifacts

• Base model: openai/privacy-filter.
• Upstream dataset: ai4privacy/pii-masking-openpii-1.5m.

6. Legal Disclaimer

No Warranty and Use at Your Own Risk: This dataset is provided "as is" without warranties of any kind, express or implied. BC Card makes no representations or warranties regarding accuracy, completeness, or suitability for any purpose. Use is at your own risk.

No Liability: BC Card and its affiliates shall not be liable for damages arising from the use of, or inability to use, this dataset.

Compliance and Responsibility: Users are responsible for ensuring that their use complies with applicable laws, regulations, contractual obligations, and ethical requirements, including relevant data-protection and AI regulations.

Synthetic Data Statement: This dataset is designed as synthetic training data and is structurally audited where public checks exist. Values are either generated locally under the documented controls or inherited from the synthetic ai4privacy corpus. Actual issuance cannot be exhaustively checked for identifiers backed by private registries. Any coincidental match with real-world information is unintended and has no source linkage to an operational record. These statements describe design intent and verification results and do not constitute a legal warranty.