Meddies/meddies-pii-v2

Meddies PII v2 Dataset

Character-level PII annotations for training and evaluating multilingual de-identification systems across 17 languages and nine entity families.

The dataset is public but access-gated. The training configs are machine-generated and the held-out configs must not enter training. Do not treat either as proof that a system is safe for clinical deployment.

If you want to use this dataset in commercial work, contact contact@meddies.ai.

Why this dataset

PII datasets often stop at names, addresses, and contact details. Clinical text also carries hospital identifiers, access-bearing portal links, credentials, local date formats, and language-specific boundary cases. Those gaps become brittle rules or silent misses in a de-identification pipeline.

Meddies PII v2 publishes exact character spans rather than tokenizer-specific tags. That keeps the annotations readable, reviewable, and reusable across tokenizers. BIOES labels are derived only when a model is trained or evaluated.

New in v2

V2 expands company_name examples to cover hospitals, clinics, insurers, pharmacies, legal suffixes, and organization names with punctuation. The held-out eval-challenge config also adds adversarial patterns such as character spacing, spelled-out separators, full-width punctuation, line breaks, voice and optical character recognition renderings, and structured payloads.

These synthetic annotations are shown exactly as stored in eval-challenge:

Challenge pattern	Annotated text	Label
Complex organization name	Phòng khám Đa khoa An Bình Từ Liêm	company_name
Character-spaced email	m i a . k e l l e r at m e d i f a x dot t e s t	email_address
Delimiter-heavy phone number	0 9 1 2 - 6 4 8 - 3 7 5	phone_number

What is in this dataset

The repository has 17 language-specific training configs and two held-out configs. Every config uses the physical split name train; eval and eval-challenge are config names, not training data.

Config group	Purpose	Rows
17 language configs	Model training and domain adaptation	317,846
eval	Pinned checkpoint-selection control set	1,700
eval-challenge	Harder held-out challenge set	3,400

Each row contains:

• text: the source string.
• label: a list of {category, start, end, text} character spans.
• info: generation, language, scenario, source, and validation metadata. Optional fields vary by config.

The label set is address, company_name, date, email_address, human_name, id_number, phone_number, private_url, and secret. A private_url is an access-bearing link, such as a signed patient-portal URL. A secret is a credential, token, password, or key.

Coverage and diagnostics

The release covers German, English, Spanish, French, Indonesian, Japanese, Korean, Lao, Malay, Burmese, Portuguese, Russian, Tamil, Thai, Filipino, Vietnamese, and Chinese.

Rows by language config

Language config	Rows
Vietnamese	38,963
English	27,757
Burmese	20,508
Chinese	19,154
Spanish	18,451
Russian	17,260
Lao	16,578
Filipino	16,515
Indonesian	16,512
German	16,490
Malay	16,303
Tamil	16,121
Thai	16,018
Korean	15,718
French	15,459
Japanese	15,124
Portuguese	14,915

The language configs intentionally target hard and underrepresented labels, including private_url and secret. Their label frequencies do not represent natural clinical prevalence. Do not use raw class counts from this dataset as an estimate of real-world risk.

Where it fits in the pipeline

The dataset is the character-span source of truth between data generation and model-specific tokenization. Text and span annotations pass schema, boundary, and label checks before publication. Training code then derives tokenizer-specific BIOES tags without changing the original annotations.

The public release contains Meddies-generated data. External public corpora used by the model are rebuilt from their original sources and are not rehosted here.

Access and quick start

The repository uses manual access approval. After accepting the access conditions and authenticating with Hugging Face, load one config explicitly:

from datasets import load_dataset

dataset = load_dataset(
    "Meddies/meddies-pii-v2",
    "vietnamese",
    split="train",
    token=True,
)

row = dataset[0]
print(row["text"])
print(row["label"])

Load a held-out config the same way:

control = load_dataset(
    "Meddies/meddies-pii-v2",
    "eval",
    split="train",
    token=True,
)

Despite the physical split name, never mix eval or eval-challenge into training.

Good fits

• Training span-based PII detectors for multilingual clinical and general text.
• Deriving tokenizer-specific BIOES labels while keeping character spans auditable.
• Testing annotation tools, constrained decoders, and redaction interfaces.
• Studying hard labels and boundary cases before collecting local evaluation data.

Limits

• The training data are machine-generated. Fluent text, varied scenarios, and validation checks do not make them equivalent to real clinical records.
• Language coverage is not the same as site coverage. Local names, identifiers, templates, and optical character recognition errors will differ.
• The label distribution is deliberately shaped for model learning and is not a prevalence estimate.
• The nine-label taxonomy is closed. Downstream teams must define how unsupported identifiers are handled.
• The dataset is not a substitute for a held-out, human-reviewed evaluation from the intended deployment environment.
• The CC BY-NC 4.0 license does not permit commercial use without separate permission.

Feedback

Send us annotation errors, invalid offsets, taxonomy ambiguities, duplicated patterns, language-specific boundary problems, or config-loading failures. Include the config name and row ID when possible, but never post real patient data or credentials.

Open a thread in this repository's Community tab or email contact@meddies.ai.

Collaboration and sponsorship

Public data should make privacy research easier to inspect, not just easier to train. We welcome hospitals, research groups, and privacy teams that can contribute safe failure patterns, independent audits, or language review.

Sponsorship supports more clinician review, better held-out evaluations, and safer coverage of underrepresented languages. Contact contact@meddies.ai to collaborate.

Citation

@misc{meddies-pii-v2-dataset,
  title={Meddies PII v2: Multilingual Character-Span Annotations for PII Detection},
  author={MeddiesAI},
  year={2026},
  url={https://huggingface.co/datasets/Meddies/meddies-pii-v2}
}