NERDD (Portuguese)

NERDD is a fine-tuned version of GLiNER Large for Named Entity Recognition (NER) in Brazilian Portuguese criminal reports collected from the Disque Denúncia anonymous crime reporting system.

The model recognizes three entity types:

• PER — Person
• LOC — Location
• ORG — Organization

Unlike general-purpose NER models, this model was trained on highly informal criminal reports containing abbreviations, slang, nicknames, misspellings and domain-specific terminology commonly found in Brazilian public security datasets.

---

Quick Start

pip install gliner

from gliner import GLiNER

model = GLiNER.from_pretrained("BIRDRED/NERDD")

text = """
Traficantes do Comando Vermelho estão escondidos na comunidade Cidade Alta.
Segundo a denúncia, Biel foi visto armado próximo à Avenida Brasil.
"""

labels = ["PER","LOC","ORG"]

entities = model.predict_entities(
    text,
    labels
)

print(entities)

Example output

[
 {'text': 'Comando Vermelho', 'label': 'ORG', 'score': 0.97},
 {'text': 'Cidade Alta', 'label': 'LOC', 'score': 0.95},
 {'text': 'Biel', 'label': 'PER', 'score': 0.94},
 {'text': 'Avenida Brasil', 'label': 'LOC', 'score': 0.96}
]

---

Supported Labels

Label	Description
PER	Persons, suspects, nicknames and aliases
LOC	Streets, neighborhoods, communities, cities and geographic locations
ORG	Criminal organizations, police units, institutions and companies

---

Benchmark Results

The best model was obtained after the fourth pseudo-labeling iteration, using metadata during the self-training process.

Metric	Score
F1-score	67.0%
Precision	66.2%
Recall	67.9%

These results were obtained on the held-out test set after fine-tuning GLiNER and applying iterative self-training with metadata-enhanced pseudo-labeling.

---

Dataset

The model was fine-tuned using a manually annotated corpus of criminal reports.

Dataset statistics:

Item	Value
Reports	5,230
Annotated entities	49,069
Person entities	20,903
Location entities	23,213
Organization entities	4,953

The original corpus is not publicly available because it contains sensitive criminal intelligence information.

---

Training Details

Detail	Value
Base model	GLiNER Large
Learning Rate	3e-5
Weight Decay	0.01
Batch Size	8
Random Seed	42
Max Length	384 tokens
Validation	5-fold Nested Cross Validation

---

When to Use This Model

Use case	Description
Criminal intelligence	Extract entities from crime reports
Public security	Information extraction from anonymous reports
Academic research	Portuguese NER evaluation
Information extraction	Structured entity extraction from unstructured reports

---

Limitations

• The model was trained specifically on criminal reports from Rio de Janeiro.
• Performance may decrease on other Portuguese domains.
• The training dataset is not publicly available.
• Organization entities are less frequent than location entities, which may slightly affect performance on this class.

---

Citation

If you use this model, please cite:

@mastersthesis{melo2026,
  author={Gustavo Melo},
  title={Named Entity Recognition in Brazilian Criminal Reports Using GLiNER, Metadata Integration and Iterative Self-Training},
  school={CEFET/RJ},
  year={2026}
}

Related work

@inproceedings{zaratiana2024gliner,
  title={GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer},
  author={Zaratiana, Urchade and others},
  booktitle={NAACL},
  year={2024}
}

---

Acknowledgements

This model was developed as part of a Master's dissertation at CEFET/RJ focused on Named Entity Recognition for Brazilian Public Security applications using the Disque Denúncia dataset.

---

License

Apache 2.0