MuRIL Indian address tagger

Tags each token of a messy Indian address with one of nine fields: HOUSE, BUILDING, STREET, LANDMARK, LOCALITY, CITY, DISTRICT, STATE and PIN, in BIO (IOB2) form (19 tags). google/muril-base-cased fine-tuned for token classification.

flat 3b, shanti apts, behind SBI atm nr hanuman mandir, supela bhilai 490023

A course project for DSL504 / DS504 Natural Language Processing, IIT Bhilai, 2026. Code, report and a demo you run locally (app/demo.py): https://github.com/devreddy1015/NLP_GROUP_PROJECT

Training

  • Base model: google/muril-base-cased at commit afd9f36.
  • Data: 20,000 synthetic addresses from the project's generator (config full: all eight noise families, i.e. abbreviations, landmark phrases, reordering and dropped fields, delimiters and case, typos, spelling variants, PIN variants and Indian number formats), built on real geography from the India Post PIN directory. Data seed 42.
  • Settings: 3 epochs, learning rate 1e-4, batch 32, linear warm-up over 10 % of steps, weight decay 0.01, gradient clipping 1.0, bf16 mixed precision, maximum length 256, seed 42. The first subword of each token carries its tag. 357 s on an RTX 4060 Laptop GPU.

Results

Span micro F1 (seqeval, strict IOB2) with a 95 % bootstrap interval over addresses:

Scored on Span micro F1 Exact match
Synthetic dev (2,000 addresses with held-out PINs) 0.991 (0.989–0.992) 0.946
20 hand-written public-place addresses (a smoke set written by the team, not a held-out test set) 0.741 (0.645–0.830) 0.350

A human-written test set was designed but not collected, so there is no score on real data yet. On the 20 hand-written addresses a rules baseline scores 0.828 and a CRF 0.796.

Use

Split the address the way the model was trained (whitespace, and every punctuation or symbol character on its own), then read the tag of each word's first subword:

import unicodedata

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

repo = "sandstromPL/muril-indian-address-parser"
tok = AutoTokenizer.from_pretrained(repo)
net = AutoModelForTokenClassification.from_pretrained(repo).eval()


def words(text):
    """Split on whitespace; every punctuation or symbol character is its own token."""
    out, cur = [], ""
    for ch in text:
        if ch.isspace() or unicodedata.category(ch)[0] in "PS":
            if cur:
                out.append(cur)
                cur = ""
            if not ch.isspace():
                out.append(ch)
        else:
            cur += ch
    return out + ([cur] if cur else [])


text = "flat 3b, shanti apts, behind SBI atm nr hanuman mandir, supela bhilai 490023"
tokens = words(text)
enc = tok(tokens, is_split_into_words=True, return_tensors="pt")
with torch.no_grad():
    best = net(**enc).logits.argmax(-1)[0].tolist()
tags, seen = ["O"] * len(tokens), set()
for pos, w in enumerate(enc.word_ids()):
    if w is not None and w not in seen:  # the first subword of each word carries its tag
        seen.add(w)
        tags[w] = net.config.id2label[best[pos]]
print(list(zip(tokens, tags)))

Limitations

  • Trained on synthetic addresses only. On hand-written text it confuses city and locality (Bhilai tagged as a locality), reads floors (Ground Floor) as streets, and misses landmarks without an English relation word (hanuman mandir ke paas).
  • Roman script and Hinglish; no Devanagari in training.
  • The place's own name (Shree Sweets) is tagged O by design.
  • Meant for addresses of public places. Do not use it on personal data without checking it on your own data first.

Licence and data

Apache-2.0, the licence of the base model. The training addresses are synthetic, built from the India Post All India Pincode Directory (data.gov.in), under the Government Open Data Licence – India (GODL-India): contains data from India Post.

Downloads last month
35
Safetensors
Model size
0.2B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for sandstromPL/muril-indian-address-parser

Finetuned
(100)
this model