document-pii-redactor model weights

Most PII redactors stop at plain text. These models power document-pii-redactor, which also redacts document images and is light enough to deploy on CPU. They are trained to understand Indian names, documents, and contexts, and the text model works across Indian languages. The main contribution is the PII token classifier — OCR is just the pluggable input stage in front of it. It defaults to lightweight Tesseract, which keeps memory low and works well for PDFs and good-quality images; for more difficult or blurred images, Bring-your-own OCR lets a model like Nemotron OCR (or Textract, Google Vision, etc) plug straight in for better results (example notebook).

Single repo holding the models used by the document-pii-redactor GitHub repo (https://github.com/eka-care/document-pii-redactor), organized by modality. You can use the document-pii-redactor repo directly to use these models — it has the library, a Docker image, and a FastAPI server.

• image/layoutlmv3/ — a text-in-image classifier for text PII in images (47 categories), run on Tesseract OCR words.
• image/yolo/best.pt — a detector for visual entities (signature, seal/stamp, QR/barcode, face photo, fingerprint, logo).
• text/minilm/ — a lightweight multilingual classifier for PII in plain text (no image, no OCR).

Try it live: ekacare/document-pii-redactor — upload a document image or paste text and see detection + redaction run in your browser.

Install

pip install "document-pii-redactor[visual]"  # full pipeline, used by the examples below (AGPL-3.0 — see License below)
pip install document-pii-redactor            # text pipeline + built-in OCR only, no visual entities (permissive licenses)

ImagePIIRedactor detects visual entities by default — PII that is an image region rather than readable text: signatures, seals/stamps, QR codes/barcodes, face photos, fingerprints, and logos. That detector needs the [visual] extra, so the image examples below assume it — or pass detect_visual=False to detect only text PII on the core install.

System dependency: Tesseract OCR — used by the image modality's built-in OCR step. Not needed for the text-only modality, nor if you bring your own OCR (detect(..., words=..., boxes=...)).

# Debian/Ubuntu
sudo apt-get install -y tesseract-ocr
# macOS
brew install tesseract

The model weights on Hugging Face are gated — authenticate once before first use (weights download when a redactor is first constructed):

hf auth login                    # interactive, stores the token locally
# or non-interactive (CI, servers):
export HF_TOKEN="your-access-token"

Generate a token under Hugging Face → Settings → Access Tokens.

Usage

detect() is the core primitive — it finds every PII entity with its location, category, and confidence, and runs the models exactly once. The transforms (redact / anonymize / de-identify) take its result as a required argument: detect once, feed the result to any transform.

from document_pii_redactor import ImagePIIRedactor, TextPIIRedactor

image_redactor = ImagePIIRedactor("ekacare/document-pii-redactor")
entities = image_redactor.detect("page.jpg")   # built-in Tesseract OCR
# each entity: kind ("text"/"visual"), category, bbox (pixels), text, score

# …or bring your own OCR — pass words + pixel boxes, Tesseract is skipped
# and your exact boxes come back on the detected entities:
entities = image_redactor.detect("page.jpg", words=["John", "Doe"],
                                 boxes=[[100, 20, 140, 40], [145, 20, 180, 40]])

text_redactor = TextPIIRedactor("ekacare/document-pii-redactor")
text  = "Mr. John Doe, 45 yrs, DOB 12-03-1979, Indiranagar, Bangalore. Contact: +91 98765 43210."
spans = text_redactor.detect(text)             # char-offset spans

Redact — destroy:

image_redactor.redact("page.jpg", entities, mode="blur").save("redacted.png")  # or "solid" / "pixelate"

text_redactor.redact(text, spans)
# '[REDACTED], [REDACTED] yrs, DOB [REDACTED], [REDACTED], [REDACTED]. Contact: [REDACTED].'

Anonymize — generalize, one-way, no mapping kept. Ages become 10-year buckets, dates keep only the year, fine geography collapses to [LOCATION] (state and country survive), everything else becomes an unnumbered token; faces/signatures are filled solid:

image_redactor.anonymize("page.jpg", entities).save("anonymized.png")

text_redactor.anonymize(text, spans)
# '[PERSON], 40–49 yrs, DOB 1979, [LOCATION], [LOCATION]. Contact: [PHONE].'

De-identify — pseudonymize. Same value → same pseudonym throughout the document (rendered in place in images; faces/signatures become neutral placeholders), and the entity→pseudonym mapping comes back for authorized re-linking — yours to store securely, never persisted by the library. strategy="hash" gives globally deterministic tokens that stay stable across documents with no mapping to thread (secret= salts the hash so guessable values can't be dictionary-reversed):

deid = image_redactor.deidentify("page.jpg", entities)    # .image + .mapping
deid.image.save("deidentified.png")

text_redactor.deidentify(text, spans).text
# 'Person_1, Age_1 yrs, DOB Date_1, City_1, City_2. Contact: Phone_1.'

text_redactor.deidentify(text, spans, strategy="hash").text
# 'Person_539681, Age_6c8349 yrs, DOB Date_7f19c4, City_d12704, City_60c7d5. Contact: Phone_d57003.'

Good to know:

• categories=[...] on detect() limits which of the 53 PII categories are found (default: all); detect_visual=False on ImagePIIRedactor skips the visual-entity detector entirely.
• Sequential pseudonyms are scoped to the returned mapping — pass mapping=result.mapping on the next page of the same record to keep numbering consistent. Hash tokens need no threading.
• Anonymization is best-effort removal/generalization of detected identifiers — not a k-anonymity guarantee or a compliance determination.

See the document-pii-redactor GitHub repo for runnable notebook walkthroughs — quickstart.ipynb and byo_ocr_nemotron.ipynb — the full API reference, the category taxonomy, and the Docker/FastAPI deployment setup (the same setup behind the demo Space above).

License and attribution

The weights are licensed per model, following each base model's license (a fine-tune is a derivative of its base — the base license flows through):

weights	fine-tuned from	license
text/minilm/	Multilingual MiniLM (MIT)	CC-BY-4.0 — free use incl. commercial; credit Eka Care with a link back
image/layoutlmv3/	microsoft/layoutlmv3-base (CC-BY-NC-SA-4.0)	CC-BY-NC-SA-4.0 — non-commercial use only, ShareAlike
image/yolo/best.pt	YOLO11m (Ultralytics, AGPL-3.0)	AGPL-3.0

Practical summary: the plain-text pipeline (TextPIIRedactor) has a fully permissive lineage and may be used commercially with attribution. The image pipeline currently inherits its bases' restrictions — no commercial use of the LayoutLMv3 fine-tune, and AGPL obligations for the visual detector (the ultralytics runtime it needs is also AGPL-3.0 and is an optional [visual] extra of the pip package).

The library code is Apache-2.0 at https://github.com/eka-care/document-pii-redactor.

If you use the models or the library, please cite:

@software{document_pii_redactor,
  author = {{Eka Care}},
  title  = {document-pii-redactor: detect, redact, de-identify, or anonymize
            PII in document images and plain text},
  year   = {2026},
  url    = {https://github.com/eka-care/document-pii-redactor},
  note   = {Model weights: https://huggingface.co/ekacare/document-pii-redactor}
}