You can deploy this model and start earning money today!

ottema/gliner2-ptbr (v0.4 — generalist)

Open-vocabulary NER for Brazilian Portuguese, fine-tuned for informal and operational text (chat, atendimento, suporte).

This is the generalist release. For HAREM-specialized (best entity F1 among compared models), see ottema/gliner2-ptbr-harem. For ontology-guided evidence extraction, see ottema/gliner2-ptbr-ontoevidence.

Model details

Base: fastino/gliner2-multi-v1 (Apache-2.0)
Type: GLiNER2 (bidirectional encoder + span scoring)
Language: pt-BR (with English fallback)
Size: ~307M parameters
License: Apache-2.0

Intended use

General-purpose open-vocabulary NER for informal and operational Brazilian Portuguese text: atendimento, chat, suporte técnico, educação. Trained on synthetic data covering pessoas, profissões, locais, organizações, documentos, produtos, marcas, tecnologias, telefones, e-mails, datas, e valores monetários.

If you need a model benchmarked on journalistic Portuguese (HAREM), use ottema/gliner2-ptbr-harem instead.

Usage

python

from gliner2 import GLiNER2

model = GLiNER2.from_pretrained("ottema/gliner2-ptbr")

text = "A professora Ana comprou um notebook Dell em Campinas no dia 12/06."
labels = ["pessoa", "profissão", "produto", "marca", "local", "data"]

entities = model.extract_entities(text, labels, threshold=0.5)
for label, spans in entities["entities"].items():
    for span in spans:
        print(f"{span} -> {label}")

Performance

Evaluation on data/gliner_ptbr_core/test.jsonl (synthetic generalist benchmark, threshold 0.3):

Model	entity_F1	span_F1	label_F1
`fastino/gliner2-multi-v1` (zero-shot)	0.9333	0.9347	0.9855
`ottema/gliner2-ptbr` (v0.4)	0.9976	0.9976	1.0000

On HAREM (163 samples, 2511 entities, journalistic PT-BR — out-of-distribution for this generalist):

Model	entity_F1	Δ vs baseline
`fastino/gliner2-multi-v1` (zero-shot)	0.4271	(reference)
`ottema/gliner2-ptbr` (v0.4)	0.4132	-1.39 pp

The generalist is best for the synthetic informal-PT-BR distribution it was trained on. For journalistic text, see the HAREM-specialized model.

Inference

GPU: ~30 ms per text (median, 32 batch)
CPU: ~50 ms per short text (≤128 tokens)

Limitations

Trained primarily on synthetic data; coverage may be limited in highly specialized domains.
Performance may degrade on very long texts (>512 tokens).
Not a substitute for domain-specific classifiers in regulated workflows.
Out-of-distribution on journalistic text (use HAREM-specialized instead).

Credits

Base architecture: GLiNER2 (Urchade Zaratiana et al.)
Base weights: fastino/gliner2-multi-v1 (Fastino)
Encoder: microsoft/mdeberta-v3-base
Fine-tuning + datasets: Ottema

License

Apache-2.0

ottema/
gliner2-ptbr