Fine-tune of openai/privacy-filter for DPDP-grade Indian PII + secrets detection. Adds 10 distinct categories on top of the base 8 so DLP audit reports can distinguish Aadhaar from PAN from GSTIN, etc.

Label space — 18 categories (base 8 + 10 Indian-specific)

Inherited from base	New in v5
account_number, private_address, private_date, private_email, private_person, private_phone, private_url, secret	aadhaar, pan, gstin, ifsc, upi_id, abha_id, voter_id, indian_passport, credit_card, bank_account_in

Eval (431-case Janus DLP corpus)

	Base openai/privacy-filter	This model (v5)
Overall F1	0.8557	0.9589
Precision	0.875	0.970
Recall	0.838	0.948
p50 latency (T4)	26 ms	27 ms
FP rate (negatives)	30%	7%

11 categories at perfect 1.000 F1

gstin, passport_in, voter_id, address_in, multi_pii, abha_id, iban, ssn, private_keys, passwords, secrets_in_code

Largest improvements over base

Category	Base F1	v5 F1
passwords	0.000	1.000
abha_id	0.000	1.000
ifsc	0.545	0.966
credit_card	0.643	0.933
aws_keys	0.667	0.966
email	0.769	0.938
address_in	0.815	1.000
upi_id	0.800	0.933
gstin	0.875	1.000

Training

5-round iterative fine-tune (v1 → v5)
1797 training examples (1122 multi-entity Indian PII + 545 disambiguation/regression-protection + 130 targeted bank/password/AWS/UPI fixes)
198 validation examples (capped to 50 in training for memory)
3 epochs each round, batch 2 + grad-accum 4, lr 1e-4, bf16 weights
T4 GPU, ~80 sec per round
Final epoch best: val_loss 0.019, val_token_accuracy 99.39%

Usage

python

from opf._api import OPF

opf = OPF(model="Vasanth155/privacy-filter-india-v5", device="cuda")
text = "customer Rohan Sharma, Aadhaar 234567890123, mobile 9876543210, PAN ABCDE1234F"
result = opf.redact(text)
print(result.redacted_text)
# customer <PRIVATE_PERSON>, Aadhaar <AADHAAR>, mobile <PRIVATE_PHONE>, PAN <PAN>

Drop-in for the base openai/privacy-filter model — same architecture, same MoE routing, same opf API.

Limitations

Trained primarily on English + Hinglish-context Indian PII. Non-Indian addresses fall back to base coverage.
16 false negatives across 431 cases (3.7% miss rate). Most are rare formats (sk-ant-example, AWS IOSFODNN canonical sample, edge UPI PSPs).
DPDP-grade label distinction works for the 10 added categories. CVV, ABHA-Luhn validation, PCI-DSS-3.4 redaction policy live in the gateway layer, not the model.

License

Apache 2.0 — inherited from openai/privacy-filter.