Arabic NER — fine-tuned GLiNER

Zero-shot-capable span NER for Arabic, fine-tuned on 17 entity types including Saudi national ID numbers and mobile numbers.

Fine-tuned from NAMAA-Space/gliner_arabic-v2.1.

Quick start

from gliner import GLiNER

model = GLiNER.from_pretrained("Elafnawaf/gliner-arabic-ner")
labels = ['شخص / person', 'منظمة / organization', 'مكان / place', 'عنوان / address']

text = "زار المهندس فهد العتيبي مقر أرامكو في الظهران."
for e in model.predict_entities(text, labels, threshold=0.6, flat_ner=False):
    print(e["label"], "->", e["text"])

The model is half the system

This repo also ships pipeline.py, which adds the regex layer, the per-label thresholds and the format validator that the published metrics were measured with.

from huggingface_hub import snapshot_download
import sys; sys.path.insert(0, snapshot_download('Elafnawaf/gliner-arabic-ner'))
from pipeline import extract

for e in extract('سجل الدخول برقم الهوية 1098765432'):
    print(e['source'], e['label'], e['text'])

11 of the 17 labels are answered by deterministic regexes rather than the network — national IDs, phone numbers, percentages, money, times and similar shapes. A regex is more reliable at those than any amount of training, and it needs no data. The remaining 6 are the model's.

Labels

Model-owned (6): شخص / person, منظمة / organization, مكان / place, عنوان / address, مطار / airport, تاريخ / date

Rule-owned (11): ترتيب / ordinal, رقم الهوية الوطنية / national id number, رقم جوال / mobile number, عدد / cardinal number, عملة / currency, كمية / quantity, مبلغ مالي / money, موقع إلكتروني / website, نسبة مئوية / percentage, وحدة قياس / unit of measurement, وقت / time

Labels are model input in GLiNER, so you can ask for types outside this list — accuracy on those is not measured here.

Results

Held-out test split, span-exact match, 14,707 gold entities.

	micro-F1
zero-shot, before fine-tuning	0.3859
after fine-tuning	0.7426
model-owned labels only	0.7297

Trained for 12069 steps.

Limitations — read these

Modern Standard Arabic only. Every training corpus is MSA. Published work on Arabic NER reports losses of up to 38% F1 on out-of-distribution dialects, so treat Gulf, Levantine or Maghrebi input as untested.

Measured on public corpora, not on production text. The test split is drawn from the same corpora and annotation guidelines as the training data. Expect lower accuracy on text from a different domain.

Span conventions are inherited from the training corpora. Whether a title like الأمير or a head noun like شركة falls inside a span follows those corpora's conventions, which may not match yours.

Normalize before you tokenize. Arabic diacritics are not matched by Python's \w, so undiacritized input is required for correct spans — pipeline.py handles this, and does the offset mapping back to your original string. If you call the model directly, do it yourself.

Training data

WikiANN (ar), IAHLT MAFAT Arabic NER, the public Wojood sample, plus a synthetic generator for Saudi ID and contact formats. Every ID and phone number in the training data is randomly generated; no real personal data was used.