maskara

A BERT-based token classification model for detecting personally identifiable information (PII) in text. It is trained with BIO tagging and covers 17 PII types including Aadhaar, PAN, passport, credit card, phone, email, API keys, and passwords.

Model Versions

Version	Description	Location
v1	First full training on ~500k synthetic OpenPII-style samples.	full-openpii-500k/final/
v2	Phase-two retrained model with augmented + real-world data.	repo root (model.safetensors)

Accuracy Comparison

Split	Metric	v1	v2 (current)	Improvement
Template-disjoint	F1	0.8930	0.8417	-0.0513
Template-disjoint	Accuracy	0.9725	0.9437	-0.0288
Real-world	F1	0.8937	0.8868	-0.0069
Real-world	Accuracy	0.9504	0.9423	-0.0081

v2 was tuned for stronger real-world recall rather than overall accuracy; see phase2_metrics.json and retrain_metrics.json for per-label details.

Model Details

Attribute	Value
Architecture	BertForTokenClassification
Base architecture	BERT-base-uncased
Hidden size	768
Layers / Heads	12 / 12
Max position embeddings	512
Parameters	108,918,563
Vocab size	30,522
Labels	35 BIO labels + O

Usage

from transformers import pipeline

ner = pipeline(
    "ner",
    model="somukandula/maskara",
    aggregation_strategy="simple"
)

ner("My email is john.doe@example.com and phone is +91-98765-43210.")

Limitations

• Trained primarily on English and Indian PII patterns; performance may vary on other locales or languages.
• CREDIT_CARD and PASSWORD show low or zero real-world F1; consider post-processing rules for production.
• Max sequence length is 512 tokens; chunk longer inputs.