Ai4Privacy.Find personal information in text and replace it with consistent placeholders while keeping the surrounding data useful.
On 12 March 2024, John Smith
visited 123 Wall St, New York
and paid $1,250.00 on his
company card.
On 12 March 2024, John Smith
visited 123 Wall St, New York
and paid $1,250.00 on his
company card.
On [DATE_1], [GIVENNAME_1]
[SURNAME_1] visited [STREET_1],
[CITY_1] and paid [AMOUNT_1]
on his company card.
Independent research teams cite Ai4Privacy datasets in their own published work.
Cited by
Introducing OpenAI Privacy Filteropenai.com · 2026View source
Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained ModelsarXiv · 2024View source
Secret-Protected Evolution for Differentially Private Synthetic Text GenerationarXiv · 2025View source
The Janus Interface: How Fine-Tuning in Large Language Models Amplifies the Privacy RisksarXiv · 2023View source
Refusal Is Not an Option: Unlearning Safety Alignment of LLMsUSENIX Security · 2025View source
(Token-Level) InfoRMIA: Stronger Membership Inference and Memorization Assessment for LLMsarXiv · 2025View source
Forget to Flourish: Leveraging Machine-Unlearning on Pretrained Language Models for Privacy LeakagearXiv · 2024View source
Portcullis: A Scalable and Verifiable Privacy Gateway for Third-Party LLM InferenceAAAI · 2025View source
Defeating Cerberus: Concept-Guided Privacy-Leakage Mitigation in Multimodal Language ModelsarXiv · 2025View source
Prεεmpt: Sanitizing Sensitive Prompts for LLMsarXiv · 2025View source
Tau-Eval: A Unified Evaluation Framework for Useful and Private Text AnonymizationarXiv · 2025View source
Measuring and Guiding MonosemanticityarXiv · 2025View sourceChoose the path that matches where your text lives and what you need to do with it.
Use curated synthetic PII data for training, benchmarking and evaluation.
Mask sensitive information in the browser before prompts are sent.
Connect hosted detection and masking to an application.
Run the pipeline inside your own Python or JavaScript environment.
Coverage, labels and access terms are visible before you choose a release.
Select a release to inspect
Mask a prompt in the browser, connect a hosted API or run detection inside your own application.
Ai4Privacy Chat masks personal data in the browser, so the prompt that reaches an external AI service no longer carries it.
In your browser
Your application sends text over HTTPS and receives detected or masked text back. The detection runs on Ai4Privacy's hosted service.
Ai4Privacy hosted service
The Python and JavaScript packages are built for local execution, so the pipeline runs where your application already runs.
Your environment
$ pip install ai4privacy$ npm i @ai4privacy/ai4privacyDefine the policy, evaluate detection and apply masking decisions consistently.
A practical method for moving from entity policy to evaluated protection.
Set the entity taxonomy and annotation policy for your context.
Evaluate spans, confidence and errors against labelled examples.
Apply consistent replacements while retaining useful structure.
Guide
Write an annotation standard labelers can follow: entity inventory, span boundary rules, tag schemes, and how disagreements get resolved.
Guide
Precision, recall and exact-match scoring for PII detection, why recall matters more, and how per-entity reporting exposes failures.
Guide
A practical taxonomy of PII entity types, grounded in two public regulatory reference points, grouped by how hard each is to detect.
p5y is Ai4Privacy’s open framework for describing personal and sensitive data, inspired by i18n and l10n.
The training data names the entity type.
The detector returns the same category.
The replacement keeps a stable, readable label.
Errors can be reviewed by entity type.
Short answers about data, deployment and licensing.
Ai4Privacy helps teams find personal information in text and replace it with placeholders. The workflow is available through synthetic datasets, Ai4Privacy Chat, a REST API and local Python and JavaScript SDKs.
The datasets presented here use synthetic PII rather than real personal data. Check the individual release page for its coverage, access and license details.
Yes. The Python and JavaScript packages are designed for local execution inside your application or workflow. The REST API is the separate hosted integration path.
Start from the region and languages in your text. The Asia-Pacific and European releases focus on those regions, the 1M series is the open-core option, and the 400K series and below support comparisons at smaller release sizes.
Each release states its own license, so check the release page before you build on it. The 3M and 2M releases come in two tiers: the open datasets are CC-BY-4.0 and the enterprise datasets need a commercial license. The 1M release is open under CC-BY-4.0, with enterprise data licensed separately. The 400K series and below use a custom license — for the 400K release that is the Ai4Privacy Dataset and Derivative Products License.
Choose the local packages when detection needs to run in your own environment. Choose the hosted REST API when your application should send text and receive detected or masked text back.
Still deciding? Tell us what you are building
The datasets are on Hugging Face, the projects are on GitHub, and the packages install from PyPI and npm.