Paste it into ChatGPT without leaking your customers' data
Press Redact and watch names, addresses, account and ID numbers turn into black bars, in six languages. A small model in your browser finds them; nothing you type leaves this page.
The demo needs JavaScript. The model is a 38 MB download, once; your browser keeps it.
Try another example, or paste your own text. The model reads English, Polish, German, French, Spanish and Lithuanian; every name and number in the examples is made up.
98% vs 95%Of the personal details in 60 hand-written messages caught; an open LLM 200 times larger caught 95%.
Nothing leaves your browserThe text is checked on this page. For your team it runs on your own servers.
18 msPer message on an ordinary CPU. The LLM needed half a second per message on a GPU.
How it works
DataNo real personal data. A local open LLM wrote 2,074 e-mails, chat messages, call notes, contracts and forms in six languages, with gaps where the details go. The gaps are filled with made-up names, addresses and numbers with valid check digits, twelve times each, so every word has an exact label.
ModelThe first 12 of mmBERT-small's 22 layers, fine-tuned to mark the words of names, addresses and dates of birth: 36 MB of weights plus a 1.6 MB vocabulary. The 60 hand-written test messages come from the same author as the prompts for the training texts, so 98% is an upper bound.
RulesNumbers with a fixed shape are found by rules and kept only when their check digits hold: IBANs, payment cards, PESEL and the national ID numbers of Germany, France, Spain, Lithuania and the UK, e-mail addresses and phone numbers. Alone they catch 39% of the details; model and rules together 98%.
Compared with an LLMQwen 2.5 7B, asked to list every personal detail, caught 95% and missed mostly the names messages are signed with, at half a second per message on a GPU. A hosted LLM would have to receive the text first, which is the leak this is meant to stop.
Code, configs and the full benchmark: github.com/balticml/pii-redaction. Weights: huggingface.co/balticml/mmbert-small-pii-redaction. Want this in front of your team's AI tools? Book a 30-min call