anonymizer

Before You Start

Do not explore the workspace first. The workflow's data-inspection step shows you what you need.

Goal

Anonymize a text dataset using NeMo Anonymizer in the way the user describes:

$ARGUMENTS

The output is a single runnable Python script that builds an AnonymizerConfig, previews results on a few rows, inspects failures and quality metrics, optionally scores output with LLM-as-judge evaluation (Replace and Rewrite modes), and (on user approval) runs the full pipeline. The script is the durable artifact — the user keeps it for re-runs, version control, and production.

Workflow

Read references/interactive.md and follow it. Anonymization is high-stakes, so there is no autopilot mode. Even when the user says "you decide" or "be opinionated", ask the minimum questions needed to choose risk_tolerance and phrase privacy_goal. The user must make those choices based on their regulatory and business context.

Rules

Always preview before running the full pipeline. Preview is cheap; a full run can be expensive and slow.
If result.failed_records is non-empty after preview, fix that before tweaking strategy. Dropped rows are a model/provider/infra problem (rate limits, auth, etc.), not a config problem. Strategy knobs won't help. See docs/troubleshooting.md "Did the run actually complete cleanly?" or the published troubleshooting guide.
Ask the user which mode. Briefly describe both: Replace detects entities and replaces each in place (faster, cheaper, keeps shape); Rewrite transforms the full text to also remove inferable identifiers (more expensive, may restructure). Use the data shape as a hint — free-text with implicit identifiers (clinical notes, biographies, depositions) leans Rewrite; structured records / log lines lean Replace — but the user picks.
For cross-record consistency (same value → same replacement everywhere), use Hash, not Substitute. Substitute is consistent within a row only.
In Replace mode, default to Substitute if the user hasn't specified a strategy. It's the most general-purpose choice and matches the bulk of production usage.
Annotate is for inspection, not production. Its output keeps the original entity text and is not privacy-safe. Use it during iteration to confirm detection is working, then switch.
Evaluation is opt-in and runs as a separate step (Replace and Rewrite modes). After preview() / run(), call anonymizer.evaluate(result) to score the output with LLM-as-judge. Replace mode: Substitute gets four judges (detection validity, type fidelity, relational consistency, attribute fidelity); Redact / Annotate / Hash get detection-validity only. Rewrite mode: detection validity + holistic privacy/quality/style judge. Evaluation is diagnostic — it scores quality, it does not change the anonymized output.
Always set AnonymizerInput.data_summary, even briefly. It is the single cheapest quality lever and it improves both detection and rewrite.
Never claim privacy guarantees. Anonymizer is best-effort. Outputs may need human review depending on risk_tolerance. Tell the user this when you finalize.

Usage Tips and Common Pitfalls

Detect.entity_labels=None (the default) is permissive — the augmenter LLM may invent labels not in DEFAULT_ENTITY_LABELS. Setting an explicit list switches to strict mode where only the listed labels are detected. To add domain labels, extend the default, don't replace it: entity_labels=[*DEFAULT_ENTITY_LABELS, "clinical_facility", ...] (DEFAULT_ENTITY_LABELS is a tuple, so unpack it into a list). Match the snake_case convention of DEFAULT_ENTITY_LABELS.
GLiNER is zero-shot — entity labels are natural-language concept names (e.g. "clinical_facility", "internal_project_codename"), not codes or enum values. Any concept you can name in English is a label GLiNER can detect.
Rewrite.instructions is a dead field today — it exists on the model but the rewrite engine never reads it. Do not use it. Put rewriter guidance in privacy_goal.protect / privacy_goal.preserve instead.
risk_tolerance only applies to Rewrite mode, not Replace.
PrivacyGoal.protect and .preserve must each be 10–1000 chars and at least 3 words. Be specific (categories, named identifiers, structural facets); avoid generic phrasing like "preserve meaning".
Validator pool is the only model role with built-in load-spreading. Set entity_validator: [a, b, c] in models.yaml if rate limits drop rows. Other roles (rewriter, evaluator, etc.) are single-alias.
Self-hosted GLiNER: When detection must not call build.nvidia.com (PHI on-prem, air-gapped, latency), run the reference server from a source checkout with python tools/serve_gliner.py. The server is not installed by pip install nemo-anonymizer. Add a provider with endpoint: http://localhost:8001/v1, then route entity_detector through a gliner-pii-detector alias with provider: local-gliner and skip_health_check: true. Match any custom --port or --host in the provider endpoint. model_configs is a complete model pool, not an overlay. Copy src/anonymizer/config/default_model_configs/models.yaml and change only the detector entry, keeping gpt-oss-120b and nemotron-30b-thinking. See docs/concepts/self-hosting-gliner.md or the published self-hosting guide.
The evaluation judges use their own model roles (detection_validity_judge, replace_type_fidelity_judge, replace_relational_consistency_judge, replace_attribute_fidelity_judge, rewrite_judge), configured in the evaluate section of models.yaml. They are not consumed by preview() / run(), so a config that anonymizes fine can still fail validation at evaluate() if those roles are unset. Defaults ship in src/anonymizer/config/default_model_configs/evaluate.yaml.
Verdict columns are null when the judge was unavailable — None means "unscored", never a pass. Replace verdict columns (type_fidelity_valid, etc.) are True / False / None. Rewrite detection_valid is a 0–1 float fraction (or None if unscored). Inspect verdicts per record with evaluated.display_record(i).
EvaluateConfig is an empty placeholder today — no knobs to set. anonymizer.evaluate(result) is the whole API; pass nothing else.

Reference Docs

The agent should consult these as it goes — do not try to enumerate field reference inline:

docs/concepts/choosing-a-strategy.md or the published strategy guide for choosing a mode, replacement strategy, risk tolerance, privacy goal, and detection settings.
docs/troubleshooting.md or the published troubleshooting guide for dropped rows, leakage, low utility, and pipeline failures. Read the relevant section when a symptom appears.
docs/concepts/detection.md or the published detection guide for GLiNER threshold semantics, entity labels, augmentation, and validation.
docs/concepts/evaluation.md or the published evaluation guide for Replace and Rewrite evaluation, judge roles, result columns, and saved-result evaluation.
docs/concepts/models.md or the published models guide for model roles and validator pools.
docs/concepts/self-hosting-gliner.md or the published self-hosting guide for the local entity_detector server, OpenAI-compatible contract, and YAML configuration.

Troubleshooting

This section covers environment-level issues. For quality and pipeline issues, read docs/troubleshooting.md or the published troubleshooting guide.

anonymizer not installed: Tell the user nemo-anonymizer is not in this Python environment (requires Python ≥ 3.11). Ask if they want you to install it (pip install nemo-anonymizer) or do it themselves. Do not install without permission.
Model/provider setup: Plain Anonymizer() ships with bundled models.yaml and providers.yaml (see src/anonymizer/config/default_model_configs/). For the default path, confirm NVIDIA_API_KEY is set. Pass custom model_configs or model_providers only for non-default endpoints or model pools. See docs/concepts/models.md or the published models guide.
LLM calls failing at preview: Check for a missing or invalid API key, a network problem, or a wrong endpoint URL. See docs/troubleshooting.md "Validation passed but preview errors at LLM call" or the published troubleshooting guide.
Local / on-prem GLiNER: Clone or download tools/serve_gliner.py from the Anonymizer repo, start the server, add a provider with endpoint: http://localhost:8001/v1, and point gliner-pii-detector at provider: local-gliner with skip_health_check: true. Preflight errors about missing aliases usually mean model_configs lists only the detector. Include the full default pool. A wrong endpoint or stopped server surfaces as a detection failure during preview. See docs/concepts/self-hosting-gliner.md or the published self-hosting guide.

Output Template

Write a Python script to the current directory. Name it after the dataset (for example, anonymize_clinical_notes.py or anonymize_support_logs.py). Fill in the TODO markers in this template and remove unused sections.

"""Anonymize <dataset> using NeMo Anonymizer.

Generated by the anonymizer agent skill.

Usage:
    python <this_script>.py                 # preview on 5 rows (fast, cheap)
    python <this_script>.py --full          # run on the full dataset
    python <this_script>.py --evaluate      # preview 5 rows, then LLM-judge-score those rows
    python <this_script>.py --full --evaluate  # run full dataset, then score the full output
"""

from __future__ import annotations

import argparse
import sys

from anonymizer import (
    Anonymizer,
    AnonymizerConfig,
    AnonymizerInput,
    DEFAULT_ENTITY_LABELS,
    Detect,
    # Pick what you need:
    # Replace mode:
    Substitute, Redact, Annotate, Hash,
    # Rewrite mode:
    Rewrite, PrivacyGoal,
)


def build_config() -> tuple[AnonymizerInput, AnonymizerConfig]:
    """Single source of truth for what we anonymize and how."""
    data = AnonymizerInput(
        source="TODO: path to .csv / .parquet / .jsonl",
        text_column="TODO: name of the text column",
        data_summary="TODO: one-line description of the data (domain, genre, anything non-obvious)",
    )

    detect = Detect(
        # Add domain labels by *extending* the default, not replacing it.
        # entity_labels=[*DEFAULT_ENTITY_LABELS, "clinical_facility", "diagnosis_code"],
        gliner_threshold=0.3,  # default; lower (0.2) for recall, raise (0.5) for cost savings
    )

    # ---- Pick ONE of the two strategies below ----

    # Replace mode (Substitute | Redact | Annotate | Hash):
    # config = AnonymizerConfig(detect=detect, replace=Substitute(
    #     instructions="TODO: short hint about the domain (e.g. names should remain plausible "
    #                  "for the original cultural context)",
    # ))

    # Rewrite mode (free-text de-identification with inferable-identifier suppression):
    config = AnonymizerConfig(
        detect=detect,
        rewrite=Rewrite(
            privacy_goal=PrivacyGoal(
                protect="TODO: what must not appear in the output, even by inference",
                preserve="TODO: what must be kept so the rewritten text is still useful",
            ),
            risk_tolerance="low",          # minimal | low | moderate | high
            strict_entity_protection=False, # True = force every detected entity into a protective disposition
            max_repair_iterations=3,
        ),
    )
    return data, config


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--full", action="store_true", help="Run on full dataset (default: preview 5 rows)")
    parser.add_argument("--num-records", type=int, default=5, help="Rows to preview (ignored with --full)")
    parser.add_argument(
        "--evaluate",
        action="store_true",
        help="LLM-judge-score the output produced this run (preview rows, or full output with --full)",
    )
    args = parser.parse_args()

    anonymizer = Anonymizer()
    data, config = build_config()

    if args.full:
        result = anonymizer.run(config=config, data=data)
        out_path = "output.parquet"  # TODO: change path/format (.csv, .jsonl) as needed
        result.dataframe.to_parquet(out_path)
        print(f"Wrote {len(result.dataframe)} rows to {out_path}")
    else:
        result = anonymizer.preview(config=config, data=data, num_records=args.num_records)
        print(f"Previewed {len(result.dataframe)} rows.")

        # Save preview output so you can investigate without re-running.
        # trace_dataframe is a superset of dataframe — it has the user-facing
        # columns plus internal columns (validation decisions, sensitivity
        # dispositions, etc.) that explain why entities were kept, dropped,
        # or rewritten.
        result.trace_dataframe.to_parquet("preview.parquet")
        print("Saved: preview.parquet (load with pd.read_parquet)")

    # Failure-first protocol: dropped rows are infra issues, not strategy issues.
    if result.failed_records:
        print(f"\n⚠️  {len(result.failed_records)} record(s) failed:")
        for fr in result.failed_records[:3]:
            print(f"   - record_id={fr.record_id} step={fr.step} reason={fr.reason}")
        print("\nFix dropped rows before tweaking strategy. See docs/troubleshooting.md or https://nvidia-nemo.github.io/Anonymizer/dev/troubleshooting/.")
        sys.exit(1)

    # Optional LLM-as-judge evaluation (Replace and Rewrite modes). Opt-in, separate
    # step — scores quality without changing the anonymized output.
    # Replace: Substitute -> 4 judges (detection validity, type fidelity, relational
    #   consistency, attribute fidelity); Redact/Annotate/Hash -> detection-validity only.
    # Rewrite: detection validity + holistic privacy/quality/style judge.
    # Needs the `evaluate` model roles in models.yaml
    # (see src/anonymizer/config/default_model_configs/evaluate.yaml).
    if args.evaluate:
        result = anonymizer.evaluate(result)
        df = result.dataframe
        if config.replace is not None:
            for col in (
                "detection_valid",
                "type_fidelity_valid",
                "relational_consistency_valid",
                "attribute_fidelity_valid",
            ):
                if col in df.columns:
                    passed = int(df[col].eq(True).sum())  # None = unscored, never a pass
                    scored = int(df[col].notna().sum())
                    print(f"{col}: {passed}/{scored} passed ({len(df) - scored} unscored)")
        else:
            # Rewrite: detection_valid is a 0–1 fraction (mean is more informative than pass count).
            if "detection_valid" in df.columns:
                scored = int(df["detection_valid"].notna().sum())
                mean_val = df["detection_valid"].mean()
                print(f"detection_valid: mean={mean_val:.2f}  scored={scored}/{len(df)}")
            if "judge_evaluation" in df.columns:
                scored = int(df["judge_evaluation"].notna().sum())
                print(f"judge_evaluation: {scored}/{len(df)} scored")
        # In a notebook, inspect per-record verdicts visually:
        #   result.display_record(0)

    # Rewrite-mode quality summary (skip for Replace mode).
    if config.rewrite is not None:
        df = result.dataframe
        print(f"\nleakage_mass:   mean={df['leakage_mass'].mean():.3f}  max={df['leakage_mass'].max():.3f}")
        print(f"utility_score:  mean={df['utility_score'].mean():.3f}  min={df['utility_score'].min():.3f}")
        print(f"flagged for review: {int(df['needs_human_review'].sum())} / {len(df)}")


if __name__ == "__main__":
    main()

About this skill