Sefaria/berel-linker-ner

BEREL Linker NER

A Hebrew Named Entity Recognition (NER) model for Rabbinic literature, fine-tuned from BEREL 3.0 — a BERT-based language model pre-trained on Rabbinic Hebrew texts by DICTA.

Model Description

This model identifies two entity types in Rabbinic Hebrew text:

Label	Hebrew	Description
Cit (B-מקור / I-מקור)	מקור	Citations — references to Jewish texts and sources
Per (B-בן-אדם / I-בן-אדם)	בן-אדם	Persons — names of people

It uses BIO tagging and was trained for the purpose of automatically linking citations and persons in Sefaria's corpus of Rabbinic literature.

Training Details

• Base model: dicta-il/BEREL_3.0
• Architecture: BertForTokenClassification (BERT-base, 12 layers, 12 attention heads, hidden size 768)
• Parameters: ~183.8M (larger than standard BERT-base due to 128,000-token Hebrew vocabulary)
• Training dataset: Sefaria/he_berel_gold (~3,000 annotated examples of Rabbinic text, split 80/20 train/test)
• Batch size: 16
• Early stopping patience: 2 epochs

Performance

Best checkpoint was saved at epoch 3 (of a possible 10) via early stopping:

Metric	Score
F1	87.2%
Precision	85.7%
Recall	88.8%
Eval loss	0.0815

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

model_id = "Sefaria/berel-linker-ner"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

ner = pipeline(
    "ner",
    model=model,
    tokenizer=tokenizer,
    aggregation_strategy="first",
    stride=128,
)

text = "דברי הרמב\"ם בהלכות שבת"
entities = ner(text)
print(entities)

Label Map

{
  "O": 0,
  "I-מקור": 1,
  "I-בן-אדם": 2,
  "B-מקור": 5,
  "B-בן-אדם": 6
}

About

Developed by Sefaria for automated entity linking in classical Jewish texts.