Edit model card


FastPolDeepNer is model for Named Entity Recognition, designed for easy use, training and configuration. The forerunner of this project is PolDeepNer2. The model implements a pipeline consisting of data processing and training using: hydra, pytorch, pytorch-lightning, transformers.

Source code: https://gitlab.clarin-pl.eu/grupa-wieszcz/ner/fast-pdn

How to use

Here is how to use this model to get Named Entities in text:

from transformers import pipeline
ner = pipeline('ner', model='clarin-pl/FastPDN', aggregation_strategy='simple')

text = "Nazywam się Jan Kowalski i mieszkam we Wrocławiu."
ner_results = ner(text)
for output in ner_results:

{'entity_group': 'nam_liv_person', 'score': 0.9996054, 'word': 'Jan Kowalski', 'start': 12, 'end': 24}
{'entity_group': 'nam_loc_gpe_city', 'score': 0.998931, 'word': 'Wrocławiu', 'start': 39, 'end': 48}

Here is how to use this model to get the logits for every token in text:

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("clarin-pl/FastPDN")
model = AutoModelForTokenClassification.from_pretrained("clarin-pl/FastPDN")

text = "Nazywam się Jan Kowalski i mieszkam we Wrocławiu."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

Training data

The FastPDN model was trained on datasets (with 82 class versions) of kpwr and cen. Annotation guidelines are specified here.


FastPDN models have been fine-tuned, thanks to pretrained models:


Runs trained on cen_n82 and kpwr_n82:

name test/f1 test/pdn2_f1 test/acc test/precision test/recall
distiluse 0.53 0.61 0.95 0.55 0.54
herbert 0.68 0.78 0.97 0.7 0.69


  • Grupa Wieszcze CLARIN-PL
  • Wiktor Walentynowicz


Downloads last month
Model size
135M params
Tensor type
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train clarin-pl/FastPDN-distiluse