localscrub 0.1.0

Local-first PHI/PII de-identification: rule/NER cascade + local LLM. Your data never leaves your machine.

localscrub

Local-first PHI/PII de-identification for clinical and free-form text.

Your data never leaves your machine: a fast rule/NER pass catches the obvious identifiers, and a local LLM handles the context-dependent remainder — no cloud APIs, no data crossing your trust boundary.

Status: v0.1 — usable, honest about its limits, pre-1.0 API. New here? Start with the quickstart. Roadmap: ROADMAP.md.

Why

Existing options force a bad trade:

• Cloud de-id APIs — accurate, but you send the sensitive data out to get it scrubbed.
• Local rule/NER tools — private, but miss context-dependent PHI in free text ("the patient's sister works at the bakery on Elm Street").

localscrub runs a two-stage cascade entirely on your hardware: cheap detection for the 90%+, a small fine-tuned local model for the ambiguous rest.

Usage

localscrub scrub notes.txt                      # print redacted text
localscrub scrub notes.txt --json               # print the JSON entity report
localscrub scrub notes.txt -o clean.txt --report report.json
localscrub scrub --report-only < notes.txt      # detect without rewriting
localscrub scrub notes.txt --fail-on-detect     # CI gate: exit 3 if anything found
localscrub scrub notes.txt --llm                # + stage 2: local LLM via Ollama
localscrub scrub notes.txt --ner                # + pretrained clinical NER (CPU)

localscrub synth -n 500 --seed 42 -o data/corpora/dev.jsonl   # synthetic eval corpus
localscrub synth -n 500 --seed 42 --diversify -o dev.jsonl    # + local-LLM paraphrase (Ollama)
localscrub mtsamples --fetch -n 100 --seed 42 -o mts.jsonl    # injection benchmark: real prose
localscrub eval data/corpora/dev.jsonl                        # per-entity precision/recall

Current accuracy numbers (including what stage 1 honestly cannot catch yet) live in docs/results.md. On our benchmark of authentic medical-transcription prose, localscrub with the [ner] extra fully redacts 99.9% of planted identifiers on CPU, vs. 75% for Presidio, the standard rules-only baseline — measured by the same harness, reproducible locally: docs/benchmark.md.

Deploying on real clinical text? Read the trust-boundary and residual-risk discussion first: docs/threat-model.md.

Or from Python:

from localscrub import scrub

result = scrub("Reach the patient at [e-mail].")
result.redacted      # 'Reach the patient at [EMAIL_1].'
result.entities      # (Entity(type=EMAIL, start=21, end=37, confidence=0.95, ...),)
result.escalations   # spans stage 1 couldn't settle — stage 2's work queue

Stage 1 (rules) currently detects: emails, phone/fax numbers, SSNs, credit cards (Luhn-validated), IP addresses, URLs, dates and ages over 89, and labeled MRN / health-plan / account identifiers. Uncertain or ambiguous spans are redacted fail-closed and flagged for stage-2 adjudication. The [ner] extra (pip install 'localscrub[ner]', then --ner) adds a pretrained clinical NER model for person/place names on CPU. Stage 2 (--llm, or scrub(text, stage2=LlmExtractor())) runs a local Ollama model for narrative identifiers and escalation adjudication; see ADR 0002 for its contract.

Entity coverage

All 18 HIPAA Safe Harbor identifier categories (45 CFR 164.514(b)(2)), plus generic PII that matters in practice: credit cards, passports, national IDs, and leaked credentials/secrets. See src/localscrub/entities.py.

License

Apache-2.0 — see LICENSE.