PIIX-560m
smart detector and masker for sensitive data in text
2 min readupdated
PIIX-560m automatically finds and masks personally identifiable information (PII) like names, phone numbers, and addresses, alongside technical secrets such as API keys and passwords. It combines a fine-tuned token classifier based on XLM-RoBERTa-large (560M parameters) with a dedicated rule layer for structured identifiers. Supporting 13 languages (including Russian), the tool handles documents of any length via overlapping window processing. It is built for developers and companies looking to scrub user data before sending it to external LLMs or to stay compliant with data regulations.
Key features
- Supports 13 languages, including Russian, English, and German
- Configurable sensitivity (recall) and entity-type filtering
- Supports fast fine-tuning using LoRA adapters
- Fast inference: from 26.5 ms per chat message on a GPU
What you can do
- Run a local service to mask PII and secrets in corporate chats
- Integrate the model into a text-processing pipeline before passing data to third-party AI models
- Fine-tune the model for custom corporate labeling standards with LoRA
- Filter specific entity types—for example, keeping only dates or stripping out phone numbers
Example
from piix import Redactor
redactor = Redactor() # загружает Tracer-ML/PIIX-560m
redactor.redact("Hi, this is Maria Gonzalez. Reach me at maria.g@acme.io or +1 415 555 0134.")
Read more
DetailsBenchmarks: F1 score of 96.9 on PrivacyBench+ and 79.4 on PII Arena (at default settings)
Training: trained on 31M tokens, Apache 2.0 license
Limitations: does not guarantee 100% anonymization and requires output verification for strict compliance requirements










Comments