haninebou/algerian-ner-ultimate

Algerian NER Model for E-commerce

Model Description

NER model for Algerian Arabic dialect, fine-tuned for e-commerce and delivery contexts.

Entities Recognized

• PERSON: Names of people
• LOCATION: Algerian cities and places
• BRAND: Product brands
• PRODUCT: Product types
• COLOR: Colors in Algerian dialect
• PRICE: Prices in DZD
• PHONE: Phone numbers
• SIZE: Clothing sizes
• QUANTITY: Numeric quantities
• ADDRESS: Delivery addresses
• ATTRIBUTE: Product attributes

Performance Metrics

Overall Performance on Test Set

• Precision: 0.9061
• Recall: 0.9112
• F1 Score: 0.9086
• Accuracy: 0.9752

Per-Class Performance

• PERSON: P=0.958, R=0.974, F1=0.966
• BRAND: P=0.736, R=0.774, F1=0.754
• PRODUCT: P=0.842, R=0.845, F1=0.843
• COLOR: P=0.814, R=0.803, F1=0.808
• PRICE: P=0.985, R=0.972, F1=0.979
• PHONE: P=1.000, R=1.000, F1=1.000
• LOCATION: P=0.915, R=0.928, F1=0.921
• SIZE: P=0.881, R=0.881, F1=0.881
• QUANTITY: P=0.959, R=0.982, F1=0.970
• ATTRIBUTE: P=0.000, R=0.000, F1=0.000
• ADDRESS: P=0.952, R=0.981, F1=0.966

Confusion Matrix

Training Data

• Total training examples: 117,744
• Validation examples: 5,048
• Test examples: 5,049

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification

model_name = "haninebou/algerian-ner-ultimate"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

# Example
text = "ismi karim nskoun f constantine"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

Model Details

• Base Model: alger-ia/dziribert
• Training Epochs: 12
• Batch Size: 16
• Learning Rate: 2e-5
• Loss Function: Cross-entropy + CRF transition loss

Created by: haninebou