Suggest classification tags

rocky_sdk.classify suggests [classification] tags for the columns of a model. It never writes a tag by itself. A person reads each suggestion, then accepts or rejects it. Only accepted suggestions reach the model sidecar.

The aid needs a DuckDB target today, because rocky profile is DuckDB only.

How the aid makes a suggestion

  rocky profile <model> --sample 5 ──▶ pattern rules ──┬─ match ─────────────────┐
                                                       └─ no match ──▶ model ──┤  (optional)
                                                                               ▼
  [classification] in models/<model>.toml ◀── apply_accepted ◀── person accepts ◀── suggestions
                    │
                    └──▶ rocky compliance checks each tag against [mask]

1. Rocky profiles the model and draws 5 random values from each column.
2. Pattern rules check the column name and the values.
3. With use_model=True, a local decision model looks again at every column the rules did not match.
4. You accept or reject each suggestion. apply_accepted writes the accepted ones.

How often the aid is wrong

The aid was measured on a test set of 160 columns. 72 of those columns held personal data.

Setup	Personal-data columns found	Suggestions that were right
Rules only	36 of 72	36 of 38
Rules, then the model	68 of 72	68 of 92

The rules were changed after the first measurement, so that dates and times no longer match as phone numbers or IP addresses. These numbers are for the changed rules on the same set, so they are likely better than you will see on your own data.

About 1 suggestion in 4 is wrong with the model on. That is why a person must check each one. The model makes two kinds of mistake often:

• It tags ordinary dates as personal data: often as phone numbers, sometimes as addresses.
• It reads short text, such as job titles or carrier names, as person names.

Each suggestion from the model carries a warning when it falls into one of these patterns. A model suggestion on a date column always carries one, unless the kind is birth_date. Its probability is shown, but do not rely on it. The model’s probabilities are not well calibrated.

Install

The rules need nothing beyond rocky-sdk:

pip install rocky-sdk

The model is an optional extra. It installs PyTorch and the Laya decision model, version 0.3.26:

pip install 'rocky-sdk[classify]'

On first use the model downloads up to 2.3 GB of weights from huggingface.co. The model runs on your machine. No column value leaves it.

Suggest, review, and write

from rocky_sdk import RockyClient
from rocky_sdk.classify import apply_accepted, suggest

client = RockyClient(models_dir="models")
sidecar = "models/stg_customers.toml"

suggestions = suggest(client, "stg_customers", sidecar=sidecar, use_model=True)

accepted = []
for s in suggestions:
    print(s.column, s.kind, "->", s.tag, f"({s.source})", s.sample_values)
    if s.warning:
        print("  warning:", s.warning)
    if input("accept? [y/N] ").strip().lower() == "y":
        accepted.append(s)

apply_accepted(sidecar, accepted)

suggest skips a column that already has a tag in the sidecar. apply_accepted only adds keys. It never changes or removes a tag that is already there.

Choose the tags

By default the aid maps each kind of personal data to a tag:

Kind	Tag
email, phone, name, address, birth_date, network_id	pii
gov_id	confidential
financial	financial

Pass kind_to_tag to use your own tags. A kind you leave out is never suggested:

suggest(client, "stg_customers", kind_to_tag={"email": "contact", "gov_id": "restricted"})

Every tag needs a strategy in [mask], or an entry in [classifications]. Otherwise rocky compliance reports the tag as a gap.

Use a local copy of the model

A site without access to huggingface.co can copy the weights and point the model at them. LayaClassifier passes its keyword arguments to laya.Router:

from rocky_sdk.classify import LayaClassifier

model = LayaClassifier(models={"typed-decisions": "/opt/models/laya-typed-decisions"})
suggest(client, "stg_customers", use_model=True, model_classifier=model)

The aid pins Hub revision 7b928d828b7b0e022f929d9bd2e44165aa270148, the one it was measured on. A different revision or checkpoint has unmeasured accuracy.