Modelpublic

mukuls9971/muril-indian-address-ner-v1

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes16downloads

Model Card

MuRIL Indian Address NER v1

Fine-tuned `google/muril-base-cased` for Indian address component detection in Hindi (Devanagari), English, and Hinglish (Roman-script Hindi-English code-mix).

Labels

Tag	Meaning	Example
`ADDRESS_HOUSE`	House / flat / door / plot number	"H.No. 12", "मकान नं. 21", "Flat 4B"
`ADDRESS_BUILDING`	Building / apartment / society name	"Prestige Residency"
`ADDRESS_STREET`	Street, road, lane, gali, marg	"MG Road", "गली नं. 4"
`ADDRESS_LANDMARK`	Landmark anchor ("near / opposite X")	"near Apollo Hospital", "मंदिर के पास"
`ADDRESS_LOCALITY`	Area, colony, nagar, mohalla, sector	"Koramangala", "Gandhi Nagar"
`ADDRESS_CITY`	City or town	"Bengaluru", "बेंगलुरु"
`ADDRESS_STATE`	State or union territory	"Karnataka", "कर्नाटक"
`ADDRESS_PIN`	6-digit Indian PIN code (optional)	"560001"

PIN code is not required — the model recognises addresses without a PIN.

Performance (Benchmark v1 — 2026-04-20)

Evaluated on a held-out slice of data/generated/address_benchmark_v1 (synthetic Indian addresses across Devanagari, English, and Hinglish).

Entity	Precision	Recall	F1
ADDRESS_LANDMARK	0.998	1.000	0.999
ADDRESS_STREET	0.996	1.000	0.998
ADDRESS_HOUSE	0.995	1.000	0.998
ADDRESS_PIN	0.995	1.000	0.997
ADDRESS_BUILDING	0.991	1.000	0.996
ADDRESS_CITY	0.968	0.981	0.974
ADDRESS_STATE	0.752	0.858	0.801
ADDRESS_LOCALITY	0.702	0.851	0.769
Overall	0.864	0.925	0.815

Training: A100-SXM4-40GB, 4 epochs, 26,728 examples, no overfitting detected.

Limitations

—Trained on synthetic data only (v1). Real-world performance will improve in v2 after adding ai4bharat/naamapadam and lince-benchmark/lince supervision.
—ADDRESS_LOCALITY precision is lower than other entities (0.702) — the model over-predicts locality in Devanagari prose outside address context.
—Coverage is limited to the in-repo gazetteer (~50 cities, 33 states, 100+ localities).

Usage

python

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="mukuls9971/muril-indian-address-ner-v1",
    aggregation_strategy="simple",
)

results = ner("H.No. 12, MG Road, Koramangala, Bengaluru - 560034")
# or Devanagari
results = ner("मकान नं. 21, गांधी नगर, भोपाल - 462001")
# or Hinglish
results = ner("Makan No. 4, Gandhi Nagar ke paas, Bhopal")

Training Details

—Base model: google/muril-base-cased
—Dataset: Synthetic Indian address corpus v1 (seed 42)
—Epochs: 4, batch size 8, max_length 192
—Learning rate: 2e-5, warmup ratio 0.1, weight decay 0.01
—Weighted loss: enabled (class imbalance handling)
—Run ID: 20260420_030651_muril-address-benchmark-v1