sixmonth-ai/sixmonth-redact

A 5.8M-parameter model that removes patient identifiers from medical and dental notes. CPU only, about 0.2 s per note, works offline.

We built it for our own internal research and liked it enough to share it. We hope it is useful, and that it inspires more work on small models that run locally.

Beta. This is an early release. Expect misses, and please report them to joe@sixmonth.ai.

in:  Seen on 03/14 by Dr. Miller. Pt Sarah Thompson, 92 y.o., MRN 448-2190, lives at 18 Oak Street, Springfield.
     Call back 555-123-4567 or sarah.t@example.com. Rx Amoxicillin 500mg tid, #14 RCT.

out: Seen on [REDACTED] by Dr. [REDACTED]. Pt [REDACTED], [REDACTED] y.o., MRN [REDACTED], lives at [REDACTED], [REDACTED].
     Call back [REDACTED] or [REDACTED]. Rx Amoxicillin 500mg tid, #14 RCT.

in:  Consultation du 14/03 avec le Dr Martin. Patiente Claire Dubois, dossier 448-2190, domiciliée 12 rue des Lilas, Lyon.
     Rappeler au 06 12 34 56 78 ou claire.d@example.fr. Amoxicilline 500 mg 3x/j, dent 36, dévitalisation.

out: Consultation du [REDACTED] avec le Dr [REDACTED]. Patiente [REDACTED], dossier [REDACTED], domiciliée [REDACTED], [REDACTED].
     Rappeler au [REDACTED] ou [REDACTED]. Amoxicilline 500 mg 3x/j, dent 36, dévitalisation.

Identifiers go, the clinical content stays. (Invented examples, real model output.)

Quick start

pip install torch transformers safetensors

from transformers import AutoModel

model = AutoModel.from_pretrained("sixmonth-ai/sixmonth-redact", trust_remote_code=True)

model.redact("Called Mrs Johnson re: appt on March 3rd. Policy no. BC-44817203.")
# 'Called Mrs [REDACTED] re: appt on [REDACTED]. Policy no. [REDACTED].'

model.redact(text, typed=True)        # [NAME], [DATE], ... instead of [REDACTED] (type is approximate)
model.redact(text, keep_dates=True)   # leave dates in place (not Safe Harbor)
model.spans(text)                     # [(start, end, label), ...]

Command line: python scrub.py "text" (also --typed, --keep-dates, --spans, or pipe text on stdin).

What it removes

The HIPAA Safe Harbor identifiers found in text, in six labels:

Label	Removes	Keeps
NAME	Patients, relatives, clinicians, staff	Titles, drug names, eponyms
DATE	Every date except a year alone	Years alone, durations
CONTACT	Phone, fax, email, URL, IP
ID	Record, plan, account, licence, plate, serial numbers	Tooth numbers, doses, lab values, codes
LOCATION	Every place smaller than a state	States, countries
AGE	Ages over 89	Ages 89 and under

English, French and Lebanese Arabic in Latin letters (Arabizi), including dental notes.

Results

Every model ran once on the same held-out documents with the same scorer. An identifier counts as caught only if every character of it is hidden. Clean masked is non-identifier text hidden by mistake (lower is better).

Model	Params	Nemotron-PII: caught	Clean masked	Clinical notes: caught	Clean masked
Perplexity PII-Tracer	596M	99.4%	0.52%	95.1%	15.25%
sixmonth-redact	5.8M	98.3%	0.34%	95.0%	3.33%
OBI deid RoBERTa i2b2	354M	95.3%	1.16%	79.0%	4.66%
GLiNER PII small	82M	92.7%	1.01%	91.8%	10.14%
GLiNER large v2.1	445M	92.1%	2.87%	84.7%	10.18%
GLiNER2 PII	307M	91.6%	0.69%	83.6%	3.32%
Presidio + spaCy lg	106M	83.3%	1.42%	64.4%	12.00%
Conway PII 30M	29M	72.6%	0.52%	74.8%	19.96%

• Nemotron-PII: NVIDIA's public set, 752 health documents, 4,595 identifiers.
• Clinical notes: 300 synthetic US, French and Lebanese notes, 1,498 identifiers, written for this test.

With typos (keyboard typos, lower-case text, missing accents on the same sets), sixmonth-redact catches 93.8% on Nemotron and 83.0% on clinical notes, against 98.7% and 93.1% for PII-Tracer, while masking 4.17% of clean clinical text against 18.42%.

Limitations

• A niche, specialised model. It was built for clinical and dental notes in English, French and Arabizi. Test it on your own data before you rely on it.
• Not a complete de-identification system. On the clinical notes, about 1 note in 5 still showed an identifier (1 in 2 with typos). Use it as a first pass with human review, never as the only safeguard. It does not by itself make data HIPAA de-identified.
• Weak spots: misspelled or unlisted place names, names with no cue in all-lower-case notes, unusual ID formats, first names that are also ordinary words ("May", "Rose").
• Other languages are untested.

Model

A character-level token classifier: character and hashed word embeddings, 1-D convolutions over characters and words, and one attention layer. It outputs six scores per character; a character is hidden when any is above zero. Word knowledge comes from public word lists and registries (see SOURCES.md). No patient note was used to fit any weight.

Licence

GPL-3.0, because the weights embed GPL-3.0 and CC BY-SA 4.0 word lists. Sources in SOURCES.md.

@misc{sixmonth_redact_2026,
  title  = {sixmonth-redact: a small neural network for clinical de-identification},
  author = {{Sixmonth}},
  year   = {2026},
  url    = {https://huggingface.co/sixmonth-ai/sixmonth-redact}
}

Built by Sixmonth. Contact: joe@sixmonth.ai.