x2aqq/indian-address-parser-model

Indian Address Parser Model

A fine-tuned IndicBERTv2-SS + CRF model for parsing unstructured Indian addresses into structured components.

Model Description

• Base Model: ai4bharat/IndicBERTv2-SS
• Architecture: BERT + Conditional Random Field (CRF) layer
• Languages: English, Hindi (Latin and Devanagari scripts)
• Training Data: 600+ annotated Delhi addresses

Performance

Entity Type	Precision	Recall	F1-Score
AREA	0.87	0.87	0.87
CITY	1.00	1.00	1.00
FLOOR	0.85	0.85	0.85
GALI	0.75	0.67	0.71
HOUSE_NUMBER	0.79	0.79	0.79
KHASRA	0.75	0.82	0.78
PINCODE	1.00	1.00	1.00
Overall	0.79	0.81	0.80

Supported Entity Types

• HOUSE_NUMBER - House/Plot/Flat numbers
• FLOOR - Floor indicators (Ground, First, etc.)
• BLOCK - Block identifiers
• SECTOR - Sector numbers
• GALI - Gali (lane) numbers
• COLONY - Colony/Society names
• AREA - Area/Locality names
• SUBAREA - Sub-area names
• KHASRA - Khasra (land record) numbers
• PINCODE - 6-digit postal codes
• CITY - City names
• STATE - State names

Usage

from address_parser import AddressParser

# Load model
parser = AddressParser.from_pretrained("YOUR_USERNAME/indian-address-parser-model")

# Parse address
result = parser.parse("PLOT NO752 FIRST FLOOR, BLOCK H-3, NEW DELHI, 110041")

# Access structured output
print(result.house_number)  # "PLOT NO752"
print(result.floor)         # "FIRST FLOOR"
print(result.city)          # "NEW DELHI"
print(result.pincode)       # "110041"

Demo

Try the live demo: HuggingFace Space

License

MIT License