martinkorelic/rudar-mmbert-slv-ner

Rudar mmBERT Slovenian NER Model

Slovenian named-entity recognition (NER) model — token classification over the four CoNLL entity classes. Part of the Rudar family of Slovenian information-extraction models.

• Language: Slovenian (sl)
• Task: Named-entity recognition (token classification)
• Base model: jhu-clsp/mmBERT-base
• Labels (BIO): PER, LOC, ORG, MISC
• Benchmark: SloBench Slovenian NER (Task 11)

Usage

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="martinkorelic/rudar-mmbert-slv-ner",
    aggregation_strategy="average",   # recommended — gives clean entity spans
)

ner("Predsednik vlade Robert Golob je v Ljubljani obiskal sedež podjetja Petrol .")
# [{'entity_group': 'PER',  'word': 'Robert Golob'},
#  {'entity_group': 'LOC',  'word': 'Ljubljani'},
#  {'entity_group': 'ORG',  'word': 'Petrol'}]

Use aggregation_strategy="average" (or "max") — "simple" may fragment subword spans.

Labels

Class	Description
PER	People
LOC	Locations
ORG	Organizations
MISC	Miscellaneous named entities

Emitted as BIO tags (B-PER, I-PER, B-LOC, …).

Training

Fine-tuned from jhu-clsp/mmBERT-base in two stages (a broad stage followed by a refinement stage) on Slovenian text — on the order of ~100k samples per stage.

Intended use & limitations

• Slovenian-language NER over the four CoNLL classes; suited to news / general-domain text.
• Expects cased, reasonably well-formed input.
• Not intended for languages other than Slovenian.