Find every personal detail in text, and know what it is

Use it now

Call the API

Create a SporeLabs account and an API key, add funds, and send your text:

curl https://sporelabs.dev/v1/scrub \
  -H "Authorization: Bearer $SPORELABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Please call Tamsin Rourke on (312) 555-3917 or email tamsin.rourke@fernmail.com.", "mode": "tag"}'

{
  "id": "scrub_...",
  "object": "scrub",
  "model": "sporelabs/scrub-v1",
  "mode": "tag",
  "results": [{
    "text": "Please call [PERSON] on [PHONE] or email [EMAIL].",
    "findings": [
      {"start": 12, "end": 25, "type": "PERSON", "text": "Tamsin Rourke"},
      {"start": 29, "end": 43, "type": "PHONE", "text": "(312) 555-3917"},
      {"start": 53, "end": 79, "type": "EMAIL", "text": "tamsin.rourke@fernmail.com"}
    ],
    "by_type": {"PERSON": ["Tamsin Rourke"], "PHONE": ["(312) 555-3917"], "EMAIL": ["tamsin.rourke@fernmail.com"]}
  }],
  "usage": {"characters": 80, "findings": 3, "cost_millicents": 4}
}

Each finding comes back with its position and type. Use them yourself, or set mode to have the text rewritten: find (default) only reports, tag swaps in the type, pseudonym swaps in a consistent stand-in, and mask swaps in *. Send up to 64 texts at once in texts.

Run it on your own machine

The model and rules are open on Hugging Face. The package downloads the model once; after that, nothing leaves your machine.

pip install "sporelabs-scrub @ https://huggingface.co/SporeLabs/scrub/resolve/main/package/sporelabs_scrub-0.1.1-py3-none-any.whl"

from sporelabs_scrub import Scrub

scrub = Scrub()  # downloads SporeLabs/scrub once
text = "Please call Tamsin Rourke on (312) 555-3917."
print(scrub.scrub(text, mode="tag"))  # Please call [PERSON] on [PHONE].
for f in scrub.find(text):
    print(f["type"], f["start"], f["end"], f["text"])

Or fetch the files: hf download SporeLabs/scrub --local-dir scrub

Why the type matters

A redactor that only knows something is sensitive gives you black bars. Scrub says what each detail is, so your app can act on it:

Route a ticket with a card number to payments. Fill a CRM record from a free-text message. Keep a map of the personal data each system holds, ready for the day someone asks what you have on them.

Stand-ins stay consistent, so the same name gets the same stand-in every time:

Everything else stays as it was, so the text still works for analytics, search and AI.

How SporeLabs built it

1. SporeLabs collected openly licensed text where personal details turn up: chats, text messages, code and config files.
2. SporeLabs placed realistic made-up details into that text and labelled each with its type.
3. SporeLabs trained a small encoder, Ettin 17M, to tag each word with its type.
4. SporeLabs added fast rules for fixed shapes, like API keys and card numbers, and to finish partly marked details.
5. SporeLabs tested it on documents it had never seen.

Results

SporeLabs tested Scrub and 4 other systems on the same 100 documents none of them had seen: support messages, forms, chats, logs and code.

Scrub caught the most personal details (92.1% against 87.3% for GLiNER2-PII), the most passwords, keys and codes (91.8% against 85.2% for OpenAI Privacy Filter) and the most details with the right type (88.3% against 80.0% for GLiNER2-PII).

Each finding comes back with its type. By type, on the same documents:

Speed

Get it

More from SporeLabs Research →