jersobh/epochdb-extractor

EpochDB Extractor

Fine-tuned GLiNER model for entity extraction in the EpochDB stack.

Base model: urchade/gliner_base
Architecture: Uni-encoder span GLiNER

Entity types

Label	Description	Examples
person	People / agents	Lukas, Alice, Claire
organization	Companies / teams	Hive Consulting, BMW Group, Sixt
location	Places	Lisboa, Berlin, São Paulo
concept	Tech / product concepts	EpochDB, HNSW, Rust, Parquet

Languages

Synthetic training data in English, Portuguese, and German.

Usage

from gliner import GLiNER

model = GLiNER.from_pretrained("jersobh/epochdb-extractor")

text = "Lukas deixou a empresa Hive Consulting e agora lidera a equipa de EpochDB em Berlin"
entities = model.predict_entities(
    text,
    labels=["person", "organization", "location", "concept"],
    threshold=0.5,
)
for e in entities:
    print(e["text"], "->", e["label"], round(e["score"], 2))

Load with GLiNER.from_pretrained only (not AutoModel / pipeline("ner")).

Training summary

• Synthetic template NER (~7.5k examples, PT + EN + DE)
• Labels: person, organization, location, concept
• ~2 epochs, AdamW lr 2e-5, batch size 8
• model_type: gliner_uni_encoder_span

Limitations

• Template-heavy synthetic data may overfit patterns
• Best on short/medium sentences similar to training templates
• Tuned for EpochDB entity vocabulary

Intended use

Entity extraction for EpochDB-related text (people, orgs, places, technical concepts) in EN / PT / DE.