hamidsalimi/persian-gliner-ner

Persian GLiNER NER

Model type: Token Classification / Named Entity Recognition
Language: Persian (fa)
Base model: microsoft/mdeberta-v3-base
GLiNER version: v0.2.28
License: Apache 2.0

Model Description

This is a fine-tuned GLiNER model for Persian Named Entity Recognition (NER) with 22 entity types. It was trained on 100,000 synthetic Persian NER samples generated from diverse templates covering news, legal, medical, and technical domains.

Intended Uses & Limitations

Intended Uses

• Named Entity Recognition in Persian text
• Information extraction from Persian documents
• Processing Persian news articles, legal documents, and medical texts
• Research and commercial applications requiring Persian NER

Limitations

• Weak categories: NARCOTICS, WEAPON, ALCOHOLIC_BEVERAGES, and CULTURAL_CONCEPT have low F1 scores due to limited synthetic training data diversity
• Generalization: Model was trained on synthetic Persian text and may not generalize well to all real-world text distributions
• Misclassification: Some entities may be misclassified due to overlapping contexts (e.g., PERCENT vs LOCATION)
• Recommended threshold: 0.3-0.5 (lower for high-recall scenarios)

Entity Types (22)

#	Label	Description
1	PERSON	People names
2	ORGANIZATION	Companies, institutions
3	LOCATION	General locations
4	DATE	Dates, years
5	TIME	Times, durations
6	MONEY	Monetary values
7	PERCENT	Percentage values
8	PROVINCE	Iranian provinces
9	CITY	Cities and towns
10	COUNTY	Counties (شهرستان)
11	FACILITY	Buildings, venues
12	EVENT	Events, incidents
13	LAW	Laws, regulations
14	LANGUAGE	Languages
15	NATIONALITY	Nationalities
16	RELIGION	Religious concepts
17	POLITICAL_PARTY	Political parties
18	CULTURAL_CONCEPT	Cultural concepts
19	PRODUCT	Products, goods
20	NARCOTICS	Drugs, narcotics
21	WEAPON	Weapons, arms
22	ALCOHOLIC_BEVERAGES	Alcoholic beverages

Training Details

Training Hyperparameters

• Steps: 20,000
• Batch size: 8
• Learning rate: 1e-5 (encoder), 5e-5 (others)
• Scheduler: Cosine with 3,000 warmup steps
• Precision: BF16
• Hardware: NVIDIA RTX 4050
• Training time: ~2 hours

Training Data

• 100,000 synthetic Persian NER samples
• Diverse templates covering news, legal, medical, and technical domains

Evaluation Results

Evaluated on 5,000 validation samples using seqeval:

Metric	Score
Micro F1	0.79
Micro Precision	0.90
Micro Recall	0.70
Macro F1	0.64
Weighted F1	0.74

Per-Entity Results

Label	Precision	Recall	F1	Support
PERSON	0.80	0.90	0.85	323
ORGANIZATION	1.00	0.71	0.83	617
LOCATION	0.86	0.86	0.86	652
DATE	0.91	0.88	0.90	1214
TIME	1.00	0.83	0.91	597
MONEY	1.00	0.76	0.87	320
PERCENT	1.00	0.02	0.03	322
PROVINCE	0.97	0.99	0.98	619
CITY	1.00	0.72	0.83	2528
COUNTY	1.00	0.83	0.91	300
FACILITY	1.00	0.87	0.93	623
EVENT	0.72	1.00	0.84	663
LAW	1.00	0.94	0.97	646
LANGUAGE	0.58	0.85	0.69	306
NATIONALITY	1.00	0.27	0.42	317
RELIGION	1.00	1.00	1.00	338
POLITICAL_PARTY	1.00	0.98	0.99	263
CULTURAL_CONCEPT	0.00	0.00	0.00	309
PRODUCT	1.00	0.02	0.03	322
NARCOTICS	0.00	0.00	0.00	316
WEAPON	0.00	0.00	0.00	316
ALCOHOLIC_BEVERAGES	0.14	0.14	0.14	316

How to Use

from gliner import GLiNER

# Load the model
model = GLiNER.from_pretrained("your-username/persian-gliner-ner")
model.eval()

# Define entity labels
labels = [
    "PERSON", "ORGANIZATION", "LOCATION", "DATE", "TIME", "MONEY", "PERCENT",
    "PROVINCE", "CITY", "COUNTY", "FACILITY", "EVENT", "LAW", "LANGUAGE",
    "NATIONALITY", "RELIGION", "POLITICAL_PARTY", "CULTURAL_CONCEPT", "PRODUCT",
    "NARCOTICS", "WEAPON", "ALCOHOLIC_BEVERAGES",
]

# Example text
text = "دانشگاه تهران یکی از بزرگترین دانشگاه‌های ایران است که در منطقه فنآوری اطلاعات تهران قرار دارد."

# Predict entities
entities = model.predict_entities(text, labels, threshold=0.5)

# Display results
for e in entities:
    print(f"{e['label']}: {e['text']} (score: {e['score']:.3f})")