Polish PII Classification Model
This model is a fine-tuned version of polish-roberta-8k, specifically designed for Named Entity Recognition (NER) focused on identifying Personally Identifiable Information (PII) in Polish text.
The model identifies entities such as names, locations, organizations, and contact numbers, making it ideal for anonymization pipelines and data privacy tasks.
Model Details
- Architecture:
RobertaForTokenClassification - Base Model: polish-roberta-8k
- Language: Polish
- Task: Token Classification (NER)
- Context Length: Up to 8192 tokens (inherited from base model)
Label Mapping
The model classifies tokens into the following 9 categories:
| ID | Label | Description |
|---|---|---|
| 0 | O | Outside any entity |
| 1 | CONTACT/NUM | Phone numbers, emails, etc. |
| 2 | EVENT | Named events |
| 3 | FACILITY | Buildings, airports, bridges, etc. |
| 4 | LOCATION | Cities, countries, regions |
| 5 | ORGANIZATION | Companies, institutions, groups |
| 6 | OTHER | Miscellaneous PII |
| 7 | PERSON | People's names |
| 8 | PRODUCT | Brand names, products |
Training Data
The model was trained on a generalized version of the KPWr-NER dataset (clarin-pl/kpwr-ner).
Fine-grained labels from the original dataset were mapped to the general categories listed above using a custom mapping process to improve the model's robustness and utility for general anonymization.
Training Hyperparameters
| Hyperparameter | Value |
|---|---|
| Epochs | 5 |
| Batch Size (Train/Eval) | 16 / 16 |
| Learning Rate | 5e-5 |
| Weight Decay | 0.01 |
| Max Sequence Length | 128 |
| Seed | 42 |
| Optimization Metric | f1_macro |
Evaluation Results
The model achieved the following results on the evaluation set:
| Metric | Value |
|---|---|
| F1 Macro | 0.8430 |
| F1 Weighted | 0.9784 |
| Precision Macro | 0.8480 |
| Precision Weighted | 0.9786 |
| Recall Macro | 0.8461 |
| Recall Weighted | 0.9787 |
| Eval Loss | 0.1099 |
Usage
1. Standard Transformers Pipeline
You can use the model directly via the Hugging Face transformers library:
from transformers import pipeline
# Load the model and tokenizer
model_id = "radlab/pii-pl-v1.0"
pii_pipeline = pipeline("ner", model=model_id, tokenizer=model_id, aggregation_strategy="simple")
text = "Jan Kowalski z firmy PKO BP mieszka w Warszawie i dowiedza ciocię Jadzię w Warszawie razem z synem Dominikiem."
results = pii_pipeline(text)
for entity in results:
print(f"Entity: {entity['word']} | Label: {entity['entity_group']} | Score: {entity['score']:.4f}")
expected output
Entity: Jan Kowalski | Label: PERSON | Score: 0.9981
Entity: PKO BP | Label: ORGANIZATION | Score: 0.9959
Entity: Warszawie | Label: LOCATION | Score: 0.9974
Entity: Jadzię | Label: PERSON | Score: 0.9955
Entity: Warszawie | Label: LOCATION | Score: 0.9974
Entity: Dominikiem | Label: PERSON | Score: 0.9942
2. Using the pii_classification Inference Wrapper
For advanced features like punctuation cleaning, sub-token merging, and full anonymization, use the AnonPredictor from
the official repository.
Installation:
git clone https://github.com/radlab-dev-group/anonymizer-model.git
cd anonymizer-model
pip install . --break-system-packages
Inference Code:
from pii_classification.inference.inference import AnonPredictor
# Initialize predictor
predictor = AnonPredictor(model_path="radlab/pii-pl-v1.0")
text = "Jan Kowalski z firmy PKO BP mieszka w Warszawie i dowiedza ciocię Jadzię w Warszawie razem z synem Dominikiem."
# 1. Get detailed predictions (merged entities)
predictions = predictor.predict(
text, clean_punct=True, merge_entities=True, handle_gaps=True
)
print(predictions)
# 2. Full anonymization
anonymized = predictor.predict_and_anonymize(text)
print(f"Anonymized Text: {anonymized['text']}")
print(f"Mappings: {anonymized['mappings']}")
expected output
{'0': 'O', '1': 'CONTACT/NUM', '2': 'EVENT', '3': 'FACILITY', '4': 'LOCATION', '5': 'ORGANIZATION', '6': 'OTHER', '7': 'PERSON', '8': 'PRODUCT'}
[{'word': 'Jan Kowalski', 'label': 'PERSON'}, {'word': ' ', 'label': 'O'}, {'word': 'z', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'firmy', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'PKO BP', 'label': 'ORGANIZATION'}, {'word': ' ', 'label': 'O'}, {'word': 'mieszka', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'w', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Warszawie', 'label': 'LOCATION'}, {'word': ' ', 'label': 'O'}, {'word': 'i', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'dowiedza', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'ciocię', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Jadzię', 'label': 'PERSON'}, {'word': ' ', 'label': 'O'}, {'word': 'w', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Warszawie', 'label': 'LOCATION'}, {'word': ' ', 'label': 'O'}, {'word': 'razem', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'z', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'synem', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Dominikiem', 'label': 'PERSON'}, {'word': '.', 'label': 'O'}]
predictions= [{'word': 'Jan Kowalski', 'label': 'PERSON'}, {'word': ' ', 'label': 'O'}, {'word': 'z', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'firmy', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'PKO BP', 'label': 'ORGANIZATION'}, {'word': ' ', 'label': 'O'}, {'word': 'mieszka', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'w', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Warszawie', 'label': 'LOCATION'}, {'word': ' ', 'label': 'O'}, {'word': 'i', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'dowiedza', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'ciocię', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Jadzię', 'label': 'PERSON'}, {'word': ' ', 'label': 'O'}, {'word': 'w', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Warszawie', 'label': 'LOCATION'}, {'word': ' ', 'label': 'O'}, {'word': 'razem', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'z', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'synem', 'label': 'O'}, {'word': ' ', 'label': 'O'}, {'word': 'Dominikiem', 'label': 'PERSON'}, {'word': '.', 'label': 'O'}]
Anonymized Text: {PERSON_1776877218745866_1} z firmy {ORGANIZATION_1776877218745866_2} mieszka w {LOCATION_1776877218745866_3} i dowiedza ciocię {PERSON_1776877218745866_4} w {LOCATION_1776877218745866_5} razem z synem {PERSON_1776877218745866_6}.
Mappings: {'PERSON_1776877218745866_1': 'Jan Kowalski', 'ORGANIZATION_1776877218745866_2': 'PKO BP', 'LOCATION_1776877218745866_3': 'Warszawie', 'PERSON_1776877218745866_4': 'Jadzię', 'LOCATION_1776877218745866_5': 'Warszawie', 'PERSON_1776877218745866_6': 'Dominikiem'}
Repository and Code
The full training pipeline, data conversion scripts, and inference API are available here: 👉 https://github.com/radlab-dev-group/anonymizer-model
Licensing
- Model Weights: Apache 2.0
- Base Model (
polish-roberta-8k): Apache 2.0 - Dataset (
kpwr-ner): Creative Commons Attribution 3.0 (CC-BY-3.0)