india-address-parser

Parse messy Indian addresses into structured fields with a small local model (1.5B, runs on CPU, no API key).

from india_address_parser import parse

parse("D.No 14/2, Gandhi Road, Pollachi, Coimbatore - 642001")
# {'flat': '14/2', 'building': '', 'street': 'Gandhi Road', 'landmark': '', 'area': 'Pollachi',
#  'city': 'Coimbatore', 'state': 'Tamil Nadu', 'pincode': '642001', 'phone': ''}

The model is abr-6024/indian-address-parser-1.5B, a fine-tune of Qwen2.5-1.5B-Instruct, run through llama.cpp as a GGUF file.

Install

pip install india-address-parser

The first call downloads the model (Q4_K_M, about 1 GB) from Hugging Face and caches it. After that it runs offline. llama-cpp-python is installed as a dependency. If pip has to compile it and fails, install a prebuilt CPU wheel first:

pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

Use

from india_address_parser import AddressParser

p = AddressParser()                       # loads the model once
p.parse("Flat 3B, Sri Murugan Apartments, Peelamedu, Coimbatore 641004, 9442012345")
p.parse_many(["12, Mettu Street, Karaikudi, Sivaganga District 630001", "..."])

AddressParser(model_path="my.gguf")       # use a local GGUF file instead of downloading

Command line:

india-address-parser "No 21, East Car Street, Srirangam, Tiruchirappalli 620006"
cat addresses.txt | india-address-parser          # one address per line, one JSON object per line out

Every result has exactly these keys: flat, building, street, landmark, area, city, state, pincode, phone. Anything not written in the address is an empty string.

Conventions

• city is the district name (or the metro name for big cities), not necessarily the town written in the address.
• area is the locality, village or post-office name. flat is the door or house number only.
• The model does not guess a city that is not written. It fills the state when the city or pincode implies it.

Accuracy and limits

The default file (Q4_K_M) scored 94.0% full-record exact match on 200 random synthetic test addresses, 29 of 30 on a hand-written Tamil Nadu set and 28 of 36 on a hand-written mixed-India set. The original unquantized model scored 94.4% on all 800 synthetic test addresses. The hand-written sets were written by the model's author, so they are not independent benchmarks. Details: GGUF model card.

• The model was trained on synthetic address text over real India Post geography. Accuracy on real customer addresses has not been measured and is likely lower. Test it on your own data before relying on it.
• Bare-number addresses such as 15 Gandhi Salai, Tirupur can lose the house number.
• Pincodes are extracted, not validated against the locality. English / Latin script only.
• parse raises ValueError if the model does not return a JSON object. Handle that in production code.

Credits and license

Code: Apache-2.0. Base model: Qwen2.5-1.5B-Instruct (Apache-2.0). Training geography: All India Pincode Directory, Department of Posts, Government of India (data.gov.in, GODL-India).