README.md

---
pretty_name: MultiPII-X
language:
- sl
task_categories:
- token-classification
size_categories:
- 10K<n<100K
source_datasets:
- extended
license: 
- cc-by-4.0
tags:
- text
- pii
- privacy
- anonymization
- named-entity-recognition
- fine-grained-ner
- slovene
- multi-domain
- medical
- legal
- telecom
- news
- finance
- banking
- gliner
- gliner2
configs:
- config_name: canonical
  default: true
  data_files:
  - split: train
    path: canonical/slovene_pii_train.parquet
- config_name: gliner
  data_files:
  - split: train
    path: gliner/gliner_train.parquet
- config_name: gliner25
  data_files:
  - split: train
    path: gliner25/gliner25_train.parquet
---

# MultiPII-X

## Description

**MultiPII-X** is an extensible fine-grained dataset for personally identifiable information (PII), sensitive-entity recognition, and text anonymization.

Version **1.0** contains Slovene (`sl`) data collected and adapted from multiple source corpora and domains under one unified annotation scheme.

The release contains:

- **15,180** text records
- **89,812** entity annotations
- **56** fine-grained entity types
- **8** normalized source collections
- general, news, telecom, legal, medical, healthcare, finance, and banking domains

The dataset is intended for training and research on fine-grained PII detection, anonymization, and named entity recognition.

MultiPII-X does **not** prescribe an official train/validation/test partition. The complete release is exposed as one `train` split so users can construct splits appropriate to their experiments.

## Dataset representations

MultiPII-X provides three configurations.

### `canonical`

The canonical representation is the authoritative version of the dataset.

```python
from datasets import load_dataset

dataset = load_dataset(
    "Nan0NJ/MultiPII-X",
    "canonical",
    split="train",
)
```

Each record contains:

- `text` — original text
- `tagged` — text with annotated spans replaced by `[ENTITY_TYPE]`
- `source` — normalized provenance/source identifier
- `domain` — text domain
- `language` — ISO language code (`sl`)
- `entities` — character-level entity annotations
- `gliner_tokenized_text` — derived word tokens for classic GLiNER
- `gliner_entities` — derived token-level entities where exact representation is possible

Each canonical entity contains:

```json
{
  "start": 0,
  "end": 5,
  "text": "Janez",
  "entity_type": "NAME_GIVEN",
  "entity_group": "PERSON"
}
```

Character offsets are zero-based and use an exclusive end offset:

```python
text[start:end] == entity["text"]
```

The `entities` field is the source of truth for annotation boundaries.

### `gliner`

A convenience representation for classic [GLiNER](https://github.com/urchade/GLiNER).

```python
dataset = load_dataset(
    "Nan0NJ/MultiPII-X",
    "gliner",
    split="train",
)
```

The representation contains:

- `tokenized_text`
- `ner`

Entity spans use zero-based token indices with an inclusive end index.

The classic GLiNER export contains **14,631** complete examples. The remaining **549** canonical rows contain character-level entity boundaries that cannot be represented exactly by the classic word-token representation without modifying the annotation. They remain available in `canonical` and `gliner25`.

The native JSONL representation is also provided in:

```text
gliner/gliner_train.jsonl
```

### `gliner25`

A convenience representation for [GLiNER2 / GLiNER2.5](https://github.com/fastino-ai/GLiNER2).

```python
dataset = load_dataset(
    "Nan0NJ/MultiPII-X",
    "gliner25",
    split="train",
)
```

All **15,180** records are available in this representation.

The native JSONL representation is also provided in:

```text
gliner25/gliner25_train.jsonl
```

## Dataset sources

MultiPII-X is a derived and reannotated resource. The original source datasets remain separate works and should be cited according to their respective terms.

The normalized `source` field contains the following values:

- `tanaos` — **7,612** records
- `gigafida` — **2,125** records
- `eeisenzopf` — **1,494** records
- `kurkowski` — **1,026** records
- `COLESLAW` — **1,000** records
- `PoVeJMo-VeMo-Med` — **1,000** records
- `telekom_synthetic` — **500** records
- `e3-jsi` — **423** records

### Upstream resources

- [Tanaos Text Anonymizer Training Dataset](https://huggingface.co/datasets/tanaos/synthetic-text-anonymizer-dataset-v1)
- [Contextual Text Anonymizer Dataset](https://huggingface.co/datasets/kurkowski/synthetic-contextual-anonymizer-dataset)
- [Corpus of Written Standard Slovene Gigafida 2.0](https://www.clarin.si/repository/xmlui/handle/11356/1320)
- [Telecom Conversation Corpus](https://huggingface.co/datasets/eisenzopf/telecom-conversation-corpus)
- [E3-JSI Synthetic Multi-Domain PII NER Dataset](https://huggingface.co/datasets/E3-JSI/synthetic-multi-pii-ner-v1)
- [PoVeJMo-VeMo-Med 1.0](https://www.clarin.si/repository/xmlui/handle/11356/1983)
- [COLESLAW 1.0](https://www.clarin.si/repository/xmlui/handle/11356/2095)

`telekom_synthetic` is additional telecom-oriented synthetic material included in the MultiPII-X construction process rather than a separately published upstream dataset.

Full citation metadata for upstream resources is provided in `CITATION.cff`.

## Domains

The current Slovene release contains the following domains:

- `general` — 8,742 records
- `news` — 2,125 records
- `telecom` — 1,994 records
- `legal` — 1,065 records
- `medical` — 1,000 records
- `banking` — 94 records
- `finance` — 82 records
- `healthcare` — 78 records

The `e3-jsi` subset retains its original domain assignments.

## Entity types

MultiPII-X v1.0 defines **56 fine-grained entity types**.

The fine-grained `entity_type` is the model target. `entity_group` is additional metadata and should not replace the fine-grained label during training.

### PERSON

`NAME_GIVEN`, `NAME_FAMILY`, `NAME_ALIAS`

### TEMPORAL

`DATE_OF_BIRTH`, `DATE`, `AGE`, `PASSPORT_EXPIRATION`, `CREDIT_CARD_EXPIRATION`

### CONTACT

`EMAIL_ADDRESS`, `PHONE_NUMBER`

### LOCATION

`LOCATION_ADDRESS`, `LOCATION_CITY`, `LOCATION_ZIP`, `LOCATION_OTHER`, `LOCATION_COORDINATES`

### FINANCIAL

`BANK_ACCOUNT`, `CREDIT_CARD`

### IDENTIFIER

`SSN`, `PASSPORT_NUMBER`, `INTERNAL_ID`, `IMEI`, `IMSI`, `RIO`, `HEALTH_INSURANCE_ID`, `DRIVER_LICENSE_NUMBER`, `STUDENT_ID`, `INSURANCE_ID`, `REGISTRATION_NUMBER`, `VEHICLE_REGISTRATION_NUMBER`, `SERIAL_NUMBER`, `RESERVATION_NUMBER`, `TRANSACTION_ID`, `VISA_NUMBER`, `BIRTH_CERTIFICATE_NUMBER`, `TRAIN_TICKET_NUMBER`, `IDENTITY_DOCUMENT_NUMBER`

### DIGITAL

`USERNAME`, `URL`, `IP_OR_MAC`

### CONTEXTUAL

`ORGANIZATION`, `PRODUCT`

### SECRET

`TAX_ID`, `PIN`, `PUK`, `CVV`, `PASSWORD`, `SIM_CODE`, `ACTIVATION_CODE`, `VERIFICATION_CODE`, `AUTH_TOKEN`

### MISC

`CURRENCY_AMOUNT`, `MEASUREMENT`, `RATIO`

### HEALTH

`MEDICAL_CONDITION`, `MEDICATION`, `BLOOD_TYPE`

Detailed label definitions are available in `label_registry.json`.

## Annotation conventions

MultiPII-X uses exact span annotations under a unified fine-grained ontology.

Important conventions include:

- **Full entity spans.** The complete intended entity is annotated.
- **Slovene suffix absorption.** Inflectional endings attached to protected entities are included in the entity span.
- **Honorific exclusion.** Titles such as `gospod`, `gospa`, `dr.`, and `prof.` remain outside personal-name annotations.
- **No clock-time tagging.** Clock times are not annotated as `DATE`.
- **Fine-grained secret labels.** Tax IDs, PINs, PUKs, CVVs, passwords, SIM codes, activation codes, verification codes, and authentication tokens use dedicated labels.
- **Fine-grained numeric labels.** Monetary amounts, measurements, and ratios use `CURRENCY_AMOUNT`, `MEASUREMENT`, and `RATIO`.

Raw source text is preserved in `text`. Character-level entities are authoritative.

## Construction

The source material was translated, adapted, sampled, reannotated, or otherwise transformed as appropriate for each upstream resource and normalized into the common MultiPII-X ontology.

The final release contains all **15,180** prepared records and all **89,812** final annotations.

The construction pipeline validated:

- entity labels against the frozen ontology
- ordered and non-overlapping character spans
- exact `text[start:end]` recovery
- deterministic reconstruction of `tagged`
- source, domain, and language normalization
- derived classic GLiNER token spans
- GLiNER2/2.5 representation compatibility

Rare labels and zero-entity records are intentionally retained.

The release contains **634** zero-entity records.

## Intended use

MultiPII-X is intended for:

- fine-grained PII and sensitive-entity recognition
- text anonymization and pseudonymization research
- NER model training and fine-tuning
- cross-domain PII detection
- evaluation of domain adaptation and label generalization
- GLiNER and GLiNER2/2.5 fine-tuning
- research on privacy-preserving NLP for Slovene

The canonical representation is model-agnostic and can be converted for architectures other than GLiNER.

## Limitations

- Version 1.0 contains Slovene only.
- Entity frequencies are intentionally imbalanced.
- Some fine-grained labels contain very few examples.
- Several source collections are synthetic or translated and should not be treated as representative samples of naturally occurring Slovene.
- Domains differ substantially in style, length, and annotation density.
- The model-specific exports inherit representation-specific limitations from their target model formats.
- MultiPII-X is a research resource and does not itself establish legal or regulatory compliance for deployed anonymization systems.

## License

MultiPII-X is released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license.

Users may share and adapt the dataset, including for commercial use, provided appropriate attribution is given and changes are indicated.

Upstream source datasets and corpora retain their own attribution requirements. See `CITATION.cff` for provenance and source citations.

## Citation

If you use **MultiPII-X** in your research, please cite:

**DOI:** [10.57967/hf/10725](https://doi.org/10.57967/hf/10725)

```bibtex
@misc{jakovchevski2026multipiix,
  author    = {Nenad Jakovchevski},
  title     = {MultiPII-X},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/10725},
  url       = {https://huggingface.co/datasets/Nan0NJ/multiPII-x}
}
```

## Author

**Nenad Jakovchevski**  
Department of Knowledge Technologies  
Jožef Stefan Institute  
Ljubljana, Slovenia

## Acknowledgements

MultiPII-X builds on multiple existing datasets and corpora. Their creators and maintainers are acknowledged in `CITATION.cff`.

Part of the annotation methodology builds on prior work on fine-grained PII anonymization in Slovene:

*Nenad Jakovchevski and Matej Martinc. Privacy-Preserving NLP for a Low-Resource Language: Benchmarking LLM-Based Personally Identifiable Information Anonymization in Slovene. CLEF 2026.*