NUS-LT-PII-xlm-roberta-large

NUS Lithuanian PII Detection Model

A model that detects personal information (PII) in Lithuanian text — names, addresses, ID codes, phone numbers, dates, health information, and other categories of sensitive data. Use it to mark or remove PII before publishing court decisions and administrative documents, processing medical or research records, or sending text to external AI assistants and translation services.

The model is the inference backend of the NUS PII validator tool — a local, GDPR-friendly web application for de-identifying Lithuanian text and validating NER models.

Model Details

• Base model: FacebookAI/xlm-roberta-large (~560M parameters)
• Architecture: transformer encoder fine-tuned for token classification
• Language: Lithuanian
• Task: named-entity recognition for PII / GDPR-relevant entities
• Output: 24 entity types (16 general + 8 sensitive — see Label set below)

Training Data

Trained on VytautoDidziojoUniversitetas/NUS-LT-PII-corpus — Lithuanian text from administrative, scientific, and media domains annotated for 24 PII / GDPR special-category entity types. See the dataset card for source breakdown, distribution, and limitations.

Uses

Intended uses:

• De-identifying Lithuanian text before publishing — court decisions, administrative documents, public records.
• Pre-processing medical or research records that contain personal data before sharing or analysis.
• Pre-processing text before sending it to external AI assistants or translation services that should not see PII.
• Compliance workflows under GDPR for Lithuanian-language data.
• Research on Lithuanian NER, cross-domain transfer, and rare-entity (sensitive GDPR category) detection.

Out of scope:

• Languages other than Lithuanian — output is undefined.
• General-purpose NER beyond the 24 PII / GDPR categories listed below.
• Stand-alone anonymization — see Risks and limitations.

Label set

How to use

For de-identification, use the NUS PII validator — a local web app that bundles this model with the boundary post-processing rules and a Lithuanian UI.

If you only need the raw weights:

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="VytautoDidziojoUniversitetas/NUS-LT-PII-xlm-roberta-large",
    aggregation_strategy="first",
)

text = "Jonas Petraitis gyvena Vilniuje, telefonas +370 612 34567."
for ent in ner(text):
    print(ent["entity_group"], "→", ent["word"])

Training Details

Training code: Faculty-of-Informatics-VDU/nus-lt-pii-model-training

Data splits (from VytautoDidziojoUniversitetas/NUS-LT-PII-corpus): the corpus ships with a train split (administraciniai, mokslinis, ziniasklaida) and a held-out validacija test split.

• Hyperparameter search: train carved into 90/10 train/val; selection by val F1.
• Final training: the 10% val portion is folded back into train — the model is trained on the full train split (no early stopping).
• Test evaluation: the held-out validacija split — untouched during HP search and final training.

Hyperparameter search: Optuna with TPE sampler (multivariate), no pruning; 22 completed trials. Selection by validation F1 with best checkpoint loading. Search space:

Final hyperparameters (winner — trial 19, val F1 0.837):

Final training: trained on the combined train + validation split (no early stopping) with the HP search winner. 5 random seeds trained independently; the median checkpoint is deployed (seed 256). Test F1 across seeds: 0.8706 ± 0.0037.

Compute: 2 × NVIDIA H100 NVL, ≈1 hour per final-training run (≈2 GPU-hours per seed).

Evaluation

Evaluated on the held-out validacija split (7148 entities) using seqeval-style entity-level metrics.

Per-label metrics

The NUS PII validator tool uses character-span matching and reports strict micro/macro F1 0.854/0.770 (P 0.851/0.753, R 0.857/0.821), partial 0.886/0.793 (P 0.884/0.775, R 0.889/0.846), and boundary recall 0.876.

Risks and limitations

• Residual re-identification risk: the model detects and labels PII spans for downstream substitution-based masking. As with any PII pipeline, two residual risks remain: (1) detector misses — at ~88% recall some PII slips through; (2) information not covered by the label set — book or event titles, project names, distinctive phrasing, or unique combinations of non-PII facts can still identify individuals. In regulated contexts (GDPR, healthcare, legal) combine with human review as part of a layered pipeline.
• Lithuanian only: despite the multilingual XLM-RoBERTa base, this fine-tune is Lithuanian-specific; other languages and code-switched documents produce undefined output.
• Inherited corpus limitations: domain skew (~84% administrative text), cross-convention annotation choices, and class imbalance on rare and sensitive entities all carry over from the training corpus — see the dataset card's Limitations and Bias section. Per-class performance has high variance on small classes; see per-label F1 in Evaluation.

License

NewGenLTU OpenRAIL-D licence — the same licence used by the underlying training corpus. See the dataset's LICENSE.txt for the full terms and restrictions.

Acknowledgement

Funded by the European Union — NextGenerationEU. Naujos kartos Lietuva.

This project is funded by the Recovery and Resilience Facility under the Lithuanian "Naujos kartos Lietuva" plan and by the State Budget of the Republic of Lithuania.

Project: Nuasmeninimo tekstyno sukūrimas (NUS) — Project No. 02-100-K-0001.

Citation

Please cite the model as:

Vytautas Magnus University and Kaunas University of Technology. 2026. NUS Lithuanian PII Detection Model. Hugging Face. https://huggingface.co/VytautoDidziojoUniversitetas/NUS-LT-PII-xlm-roberta-large

BibTeX:

@misc{nus_lt_pii_model_2026,
  author    = {{Vytautas Magnus University} and {Kaunas University of Technology}},
  title     = {NUS Lithuanian PII Detection Model},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/VytautoDidziojoUniversitetas/NUS-LT-PII-xlm-roberta-large},
  note      = {Developed under the New Generation Lithuania plan, project 02-100-K-0001.}
}