hiteshwadhwani/pii-model-indicv2

PII Detection Model — IndicBERTv2

A token classification model for detecting and redacting Personally Identifiable Information (PII) in English, Hindi, Hinglish, and Devanagari text.

Built on ai4bharat/IndicBERTv2-MLM-only — a 278M parameter BERT model pretrained on 20.9B tokens across 24 Indian languages.

Supported Languages

• English — names, addresses, phone numbers, SSN, etc.
• Hindi (Devanagari) — राजेश कुमार, मुंबई, बीस हज़ार रुपये, पंद्रह मार्च
• Hinglish — "Mera naam Rajesh hai aur main Mumbai mein rehta hoon"
• Mixed Devanagari + English — "मेरा phone number 9876543210 है"

Entity Types (31)

Entity	Description	Example
FIRSTNAME	First name	Rajesh, राजेश, John
LASTNAME	Last name	Kumar, कुमार, Smith
MIDDLENAME	Middle name	Kumar
PREFIX	Title/prefix	Mr, श्री, Dr, श्रीमती
GENDER	Gender	male, female
SEX	Sex	M, F
AGE	Age	35
DOB	Date of birth	15/03/1990
DATE	General date	पंद्रह मार्च, March fifteenth
EMAIL	Email address	priya@gmail.com
PHONENUMBER	Phone number	+91 98765 43210
CITY	City	Mumbai, मुंबई, Boston
STATE	State	Maharashtra, महाराष्ट्र
COUNTY	County	Cook County
ZIPCODE	ZIP/PIN code	400001, 02101
STREET	Street name	MG Road, Oak Avenue
BUILDINGNUMBER	Building number	42
SECONDARYADDRESS	Apt/Suite	Flat 301
COMPANYNAME	Company	Infosys, टाटा कंसल्टेंसी
ACCOUNTNUMBER	Account number	9876543210
ACCOUNTNAME	Account name	Tata Consultancy
CREDITCARDNUMBER	Credit card	4111-1111-1111-1111
CREDITCARDCVV	CVV	123
CREDITCARDISSUER	Card issuer	Visa, HDFC
SSN	SSN/PAN/Aadhaar	123-45-6789, ABCDE1234F
IBAN	IBAN	IN89UTIB00001234567890
PIN	ATM/Security PIN	4098
PASSWORD	Password	S3cur3P@ss!
USERNAME	Username	mdavis
URL	URL	www.example.com
AMOUNT	Money amount	50000, बीस हज़ार रुपये

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

model = AutoModelForTokenClassification.from_pretrained("hiteshwadhwani/pii-model-indicv2")
tokenizer = AutoTokenizer.from_pretrained("hiteshwadhwani/pii-model-indicv2")

ner = pipeline("ner", model=model, tokenizer=tokenizer, aggregation_strategy="first")

# English
results = ner("Mr John Smith lives at 456 Oak Avenue Boston")

# Hinglish
results = ner("Mera naam Rajesh Kumar hai aur main Mumbai mein rehta hoon")

# Hindi (Devanagari)
results = ner("मेरा नाम राजेश कुमार है और मैं मुंबई में रहता हूं")

# Hindi amounts and dates in words
results = ner("आपके क्रेडिट कार्ड के बीस हज़ार रूपये शुल्क लंबित हैं, जो पंद्रह मार्च को देना था।")

for entity in results:
    if entity["entity_group"] != "O":
        print(f"{entity['word']} → {entity['entity_group']} ({entity['score']:.2f})")

Redaction Example

def redact_pii(text, ner_pipeline, threshold=0.5):
    results = ner_pipeline(text)
    entities = [r for r in results if r["score"] >= threshold and r["entity_group"] != "O"]
    entities.sort(key=lambda x: x["start"])

    merged = []
    for ent in entities:
        label = ent["entity_group"]
        if merged and merged[-1]["label"] == label and ent["start"] <= merged[-1]["end"] + 1:
            merged[-1]["end"] = max(merged[-1]["end"], ent["end"])
        else:
            merged.append({"label": label, "start": ent["start"], "end": ent["end"]})

    redacted = text
    for span in reversed(merged):
        redacted = redacted[:span["start"]] + f"[{span['label']}]" + redacted[span["end"]:]
    return redacted

print(redact_pii("Shri Rajesh Kumar lives at 42 MG Road Bengaluru Karnataka", ner))
# [PREFIX] [FIRSTNAME] [LASTNAME] lives at [BUILDINGNUMBER] [STREET] [CITY] [STATE]

print(redact_pii("आपके बीस हज़ार रूपये शुल्क लंबित हैं, जो पंद्रह मार्च को देना था।", ner))
# आपके [AMOUNT] शुल्क लंबित हैं, जो [DATE] को देना था।

Evaluation Results

Metric	Score
Overall F1	0.9497
Precision	0.9464
Recall	0.9530

Per-Entity F1

Entity	F1	Entity	F1
FIRSTNAME	0.99	LASTNAME	0.99
CITY	0.99	STATE	0.99
PHONENUMBER	0.98	EMAIL	0.93
DOB	0.96	DATE	0.95
COMPANYNAME	0.97	PREFIX	0.98
CREDITCARDNUMBER	0.95	CREDITCARDISSUER	0.95
URL	0.99	USERNAME	0.99
MIDDLENAME	1.00	ACCOUNTNUMBER	0.91
PASSWORD	0.91	ZIPCODE	0.95
AMOUNT	0.86	STREET	0.91

Why IndicBERTv2?

	indic-bert (v1)	IndicBERTv2 (this model)
Parameters	32M	278M
Architecture	ALBERT (shared layers)	BERT (unique layers)
Languages	12	24
Training corpus	~9B tokens	20.9B tokens
Devanagari names	Partial subword issues	Clean detection
Hindi amounts (बीस हज़ार)	Not detected	Detected
Hindi dates (पंद्रह मार्च)	Not detected	Detected
Unusual names (Viteshwar)	Broken subword alignment	Clean detection

Training Details

• Base model: ai4bharat/IndicBERTv2-MLM-only (278M params, 24 languages)
• Task: Token Classification (NER with BIO tagging)
• Epochs: 10
• Learning rate: 2e-5
• Batch size: 16
• Optimizer: AdamW

Limitations

• SEX entity has low F1 — overlaps with GENDER
• SECONDARYADDRESS detection is weak
• Latency is higher than v1 (~25-40ms vs ~8ms) due to larger model size
• Best suited for Indian and US PII patterns

License

Apache 2.0