NeMo Studio Anonymizer | NVIDIA NeMo Helix

Use Datasets > Anonymizer in the NeMo Studio workspace sidebar to detect and protect PII in a dataset through context-aware replacement or rewriting. The page is enabled by default; an administrator can hide it by setting studio.feature_flags.anonymizer_enabled: false in the platform configuration and restarting Studio.

Studio covers the same workflow as the nemo anonymizer CLI and the sdk.anonymizer accessor. For concepts, configuration shape, and the CLI and SDK paths, see Anonymize Data.

Anonymizer Job List

The Anonymizer page lists the anonymizer jobs in the current workspace and polls for status changes.

Action	Where	Result
Start a new job	Anonymize Data	Opens the builder.
View details	Row actions > View details	Opens the job detail page.
Search	Name search bar	Filters the list to jobs whose names match the query.
Filter	Status, Created At, and Updated At filters	Narrows the list by job status or date range.
Cancel a job	Row actions > Cancel	Available while the job is in a cancellable state.
Delete jobs	Row actions > Delete, or select rows and delete	Removes the selected jobs.

The table shows each job’s name, description, creation time, and status. The Updated column is hidden by default and can be enabled from the column controls.

Build a Job

The builder has a configuration panel on the left with Source and Model Settings tabs, and a preview panel on the right. Preview runs the current configuration over a few records; Full Run submits the job and opens its detail page.

Data Source

Field	Description
Source	Dataset to pick a file from a workspace fileset, or URL for an HTTP(S) address.
Dataset / URL	The `.csv` or `.parquet` file to anonymize.

Generation

Choose the anonymization strategy. Each strategy exposes its own parameters.

Strategy	What it does	Parameters
Substitute	Replaces detected entities with LLM-generated synthetic values.	None.
Redact	Replaces entities with a label-based marker, removing the original text.	Format template (default `[REDACTED_{label}]`), label normalization.
Annotate	Tags entities with their label but keeps the original text.	Format template (default `<{text}, {label}>`; requires `{text}` and `{label}`).
Hash	Replaces entities with a deterministic digest, so the same text always hashes the same way.	Algorithm (SHA-256 (default), SHA-1, MD5), digest length (6–64, default 8), format template (default `<{label}_{digest}>`; requires `{digest}`).
Rewrite	Rewrites the whole text into a privacy-safe version that reduces explicit and inferable identifiers.	Privacy goal, LLM instructions, risk tolerance, max repair rounds, strict entity protection.

Preview Rows sets how many records a preview run anonymizes (1–10, default 1). It does not affect a full run.

Rewrite-specific parameters:

Field	Description
Privacy Goal	Default, or Custom to state what to Protect (identifiers) and what to Preserve (utility and meaning).
LLM Instructions	Optional extra instructions for the rewriter.
Risk Tolerance	Minimal, Low (default), Moderate, or High.
Max Repair Rounds	Repair passes run when leakage exceeds the tolerance. `0` disables repair.
Strict Entity Protection	Forces every detected entity to be protected regardless of risk.

Columns

Text Column names the column holding the text to anonymize. When Studio can read the selected file, this is a dropdown of its columns and a single-column file is selected automatically; otherwise it is a free-text field. Data Summary is an optional description of the data that helps the models produce better results.

Entities

Auto-detect covers all available entity labels and lets the augmenter add more labels it finds. Custom restricts the output to the labels you select; the picker groups the labels the platform reports by category and accepts a typed-in label of your own. In Custom mode, Also include all N available entities adds the available entity labels to your selection. Custom mode requires at least one label or the available entities.

Model Settings

The Model Settings tab assigns a model to each role the selected strategy uses, with an inference-parameter dropdown per role. The model choices are the models served by the workspace’s inference providers.

Strategy	Roles
Redact, Annotate, Hash	Entity Detector, Entity Validator, Entity Augmenter, Latent Detector.
Substitute	The detection roles, plus Replacement Generator.
Rewrite	The detection roles, Replacement Generator, and Domain Classifier, Disposition Analyzer, Meaning Extractor, QA Generator, Rewriter, Repairer, Evaluator.

Entity Detector runs a token-classification NER model rather than a general chat model, so it takes no sampling parameters. It defaults to Built-in GLiNER2 (in-process), which runs the platform’s built-in GLiNER2 detector inside the anonymizer and needs no model selection. To use your own detector instead, select a workspace NER model: the NER families Studio recognizes — GLiNER models and privacy filters — are grouped under Suggested with the built-in option, and any other model remains selectable under Other models.

Preview

Preview streams anonymized records into the preview panel. Page through the returned records, expand Logs to follow the run, and use Stop to end a preview early. A banner reports records that failed during the run.

Job Detail

The job detail page polls until the job reaches a terminal state.

Panel	What it shows
Job Details	Status, strategy, created and updated times, creator, source, and the error message when the job failed.
Results	Result artifacts, each with a Download action. Populated once the job is terminal.
Preview	The first records of the result dataset. Download the result for the full dataset.
Failed Records	Records the job could not anonymize, when there are any.
Logs	Job logs, with a download action.

Use the actions menu in the page header to cancel a running job or delete the job.