piitag
A per-byte PII tagger in the NNUE shape: hashed byte n-grams feed two int16 accumulators, then a
small int8 MLP scores every byte. It is the model the Rust crate cynch-piitag embeds; the crate
pins this repository's commit and the file's sha256 in its weights.lock.
piitag.piitag is the .piitag model file, format version 4: an 8-byte magic, a little-endian u32
header length, a flat JSON header, then 64-byte-aligned little-endian arrays (the two n-gram tables
and their biases, the MLP, the type head and the sorted u64 name-dictionary hashes).
Training data and attribution
Trained on nvidia/Nemotron-PII at revision
b70ffaf5ff39e079776134c5bf4381f00a9fd1ed, by NVIDIA, licensed CC-BY-4.0. These weights are learned from that dataset;
this card is its attribution. Test metrics are on its held-out test split.
The name dictionary is built from public-domain US government data: the Social Security
Administration's baby names (via hadley/data-baby-names) and the Census 2000 surname table (via
fivethirtyeight/data).
Results (held-out test split)
The threshold threshold_q = -12238 was chosen on validation data for
0.980 byte recall.
| model | span leak rate | fully missed | byte precision | byte recall | PR-AUC |
|---|---|---|---|---|---|
| piitag | 0.0104 | 0.0066 | 0.830 | 0.979 | 0.981 |
| linear baseline | 0.0170 | 0.0132 | 0.644 | 0.978 | 0.959 |
A span leaks when any of its bytes is left unredacted. The integer Rust inference matches the Python reference exactly: 0 mismatches on 2000 test buffers.
Entity types
Each span is named by a type head (one of 34 types); a gold span's type is right when the type summed over its blocks matches. Accuracy 0.970 over 466210 test spans, 0.954 averaged over types.
| type | test spans | accuracy |
|---|---|---|
| account_number | 16698 | 0.944 |
| api_key | 4667 | 0.971 |
| bank_routing_number | 8354 | 0.949 |
| biometric_identifier | 11379 | 0.979 |
| certificate_license_number | 3002 | 0.943 |
| coordinate | 7677 | 0.990 |
| credit_debit_card | 12936 | 0.982 |
| customer_id | 20502 | 0.964 |
| cvv | 4843 | 0.942 |
| date_of_birth | 18079 | 0.994 |
| device_identifier | 2507 | 0.927 |
| 53930 | 0.997 | |
| employee_id | 8878 | 0.960 |
| fax_number | 6442 | 0.868 |
| health_plan_beneficiary_number | 10561 | 0.972 |
| http_cookie | 5014 | 0.991 |
| ipv4 | 6078 | 0.996 |
| ipv6 | 3139 | 0.999 |
| license_plate | 4083 | 0.971 |
| mac_address | 4250 | 0.996 |
| medical_record_number | 11098 | 0.972 |
| national_id | 2848 | 0.926 |
| password | 6986 | 0.971 |
| person_name | 143624 | 0.981 |
| phone_number | 23963 | 0.961 |
| pin | 6456 | 0.901 |
| postcode | 6280 | 0.938 |
| ssn | 6062 | 0.961 |
| street_address | 16875 | 0.981 |
| swift_bic | 5559 | 0.990 |
| tax_id | 1302 | 0.871 |
| unique_id | 1777 | 0.793 |
| user_name | 15800 | 0.864 |
| vehicle_identifier | 4561 | 0.989 |
Spec
{"k": 8, "ctx_orders": [1, 2, 3], "ctx_l": 32, "ctx_r": 32, "log2_ctx": 13, "loc_offsets": [-2, -1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9], "loc_orders": [1, 2], "log2_loc": 12, "h1": 32, "h2": 32, "h3": 32, "dilate": 1}
Dictionary names: 152167.