Skip to content

Developer Notes

Updates and notes from the NeMo Anonymizer team.

  • 10 min read

After Anonymization — Part II: Evaluating Rewrite Mode

Replace mode can hide a name and change an address, yet leave the person recognizable. Imagine a biography describing the only pediatric cardiologist in a small desert city who trained overseas, survived a rare cancer, and later founded a clinic after a major wildfire. None of those details is a direct identifier. Together, they form a fingerprint that may point to one person.

Rewrite mode tackles that harder problem by transforming the entire record. But rewriting introduces a new risk: protect too little and identity clues remain; protect too much and the data loses its value.

NeMo Anonymizer evaluates that tradeoff twice. During anonymization, it measures how much identifying information remains (leakage) and how much useful meaning is preserved (utility), then repairs failing rewrites. Afterward, optional LLM judges assess literal entity coverage and the rewrite's overall privacy, quality, and style.

This is Part 2 of a two-part series on evaluation in Anonymizer. Part 1 covers Replace mode, while this devnote covers Rewrite mode.

  • 10 min read

After Anonymization — Part I: Evaluating Replace Mode

You need to share a dataset of customer biographies for model development without exposing personal information, so you run it through NeMo Anonymizer's Replace mode. At first glance, the result looks right: the obvious identifiers have changed, the text reads naturally, and nothing appears broken.

Look closer, though. One sensitive value was never detected. The synthetic name no longer matches the email address. The city and postal code belong to different regions. An age of 38 became 8, quietly turning an adult into a child. The record looks anonymized, but it may still expose private information or distort the original meaning.

Anonymizer.evaluate() provides a second pass over these results. It checks whether detection covered the sensitive values in the original text and, for Substitute mode, whether the generated replacements preserve their types, important attributes, and relationships.

This is Part 1 of a two-part series on evaluation in Anonymizer. It explores how Replace-mode evaluation surfaces problems that a quick review can miss, what each score means, and how to interpret the results. Part 2 will cover Rewrite mode, where the evaluation questions are different.

  • 11 min read

Running NeMo Anonymizer Fully Self-Hosted on One B300

Can Anonymizer run fully self-hosted and still process a real batch in minutes? The question comes up whenever sensitive data can't leave your infrastructure — think of an internal legal dataset that still needs realistic anonymized text for testing, review, or downstream model work. In that setting, self-hosting isn't just a deployment preference; it's what decides where the privacy boundary sits.

This devnote takes on that problem with a single B300 instance on Brev. Every model the pipeline touches runs on that one GPU: Qwen behind vLLM and GLiNER behind Anonymizer's reference OpenAI-compatible server, with Anonymizer pointed at localhost. Nothing leaves the box.

The result: one warm Anonymizer call processed 250 compact legal records in under 9 minutes. No managed remote model API was in the loop.

  • 12 min read

Introducing NeMo Anonymizer: Text Anonymization for the Reasoning Era

Picture this: you ship a year of customer support transcripts to a vendor for model fine-tuning. Names, emails, phone numbers, account IDs — every identifier you can think of — stripped. A week later, the vendor's eval team flags a transcript and identifies the account: the customer who escalated repeatedly after a regional outage last spring, then quietly churned a month later.

You didn't leak a name.

You leaked a fingerprint.

The escalation pattern, the region, the churn timing — combined, they point to one account. This is the privacy problem that text anonymization actually has to solve in 2026, and it's the problem NeMo Anonymizer — Anonymizer, for short — is built for.