A 350M encoder that finds nine types of personally identifiable information across 17 languages and returns exact character spans for review and redaction.
[!CAUTION]
Keep a human in the loop
This model can miss identifiers. It cannot certify that text is safe to share. Test it on your documents and keep human review in any regulated workflow.
[!NOTE]
Commercial licensing
CC BY-NC 4.0 applies to the rights MeddiesAI can license in this release. Contact contact@meddies.ai about commercial terms for those contributions. The underlying LiquidAI materials remain subject to LFM Open License v1.0.
Meddies PII v2 detects nine PII entity families across 17 languages and returns exact character spans
Why this model
Clinical records rarely keep identifiers tidy. One note can mix a patient name, hospital ID, phone number, signed portal link, login secret, and a date format that only makes sense locally. Rules catch the obvious cases, then lose ground when the language or document template changes.
Meddies PII v2 handles those cases through one span extraction interface. We built it for clinical text, but its nine labels also cover general documents that need review before release.
What the model returns
Pass in raw text. The model returns the label, text, start offset, and end offset for every detected span.
Nine fixed labels with address, company_name, date, email_address, human_name, id_number, phone_number, private_url, and secret.
Seventeen languages with German, English, Spanish, French, Indonesian, Japanese, Korean, Lao, Malay, Burmese, Portuguese, Russian, Tamil, Thai, Filipino, Vietnamese, and Chinese.
Character offsets included so every prediction maps back to the source text for review or redaction.
Outside the label set are images, layout, diagnoses, lab results, ages, doses, medical measurements, and public links.
Quick start
python
1import sys
23from huggingface_hub import snapshot_download
45path = snapshot_download("Meddies/meddies-pii-v2")6sys.path.insert(0, path)78from modeling_meddies_pii import MeddiesPiiExtractor
910model = MeddiesPiiExtractor.from_pretrained(path)11spans = model.extract(12"Bệnh nhân Nguyễn Văn A, SĐT 0912 345 678, tái khám ngày 12/08/2026."13)1415for span in spans:16print(span.label, span.start, span.end, span.text)
Expected output
text
1human_name 10 22 Nguyễn Văn A
2phone_number 28 40 0912 345 678
3date 56 66 12/08/2026
Choose the weight form
Form
Load
Parity check
Use it when
Adapter
from_pretrained(path)
0 of 200 rows differed
You need the evaluated execution path.
Merged
from_pretrained(path, weights="merged")
14 of 200 rows differed
You need one self-contained weight file and can accept the measured delta.
Merging folds the adapter into the base weights and changes the numerical path. Borderline BIOES decisions can move. Use the adapter when exact release reproduction matters.
ONNX and local browser use
The ONNX companion packages the same checkpoint for ONNX Runtime and WebGPU. Its full precision graph matched the PyTorch adapter on all 200 parity rows. The current int8 graph differed on 12 of 200 rows.
The browser demo runs on the user's device. Raw clinical text never leaves the device.
Evaluation
Meddies PII v2 reached 0.8937 exact typed micro F1 and 0.9429 containment typed micro F1 on 263,785 benchmark rows. It had the highest exact typed F1 in all 17 evaluation cells, all 17 language slices, and all nine entity families against the six reference systems in this comparison.
Meddies PII v2 evaluation snapshot with overall, comparative, entity, and language results
Exact typed F1 requires the label and both character boundaries to match. Containment typed F1 accepts a correct label when one span contains the other. Each macro average gives the same weight to the unit named in the column.
The v2 preview is the earlier 150-step checkpoint that preceded this release.
Evaluation cell
Meddies v2
GLiNER2
v2 preview
Meddies v1
OpenMed
OpenAI PF
LFM2.5 PII
ai4privacy_de
0.827
0.663
0.664
0.616
0.620
0.716
0.610
ai4privacy_en
0.878
0.687
0.673
0.667
0.656
0.745
0.568
ai4privacy_es
0.870
0.732
0.725
0.649
0.683
0.794
0.655
ai4privacy_fil
0.848
0.690
0.590
0.606
0.584
0.667
0.600
ai4privacy_fr
0.844
0.674
0.679
0.618
0.604
0.738
0.611
ai4privacy_id
0.848
0.688
0.601
0.588
0.566
0.664
0.627
ai4privacy_ja
0.750
0.366
0.499
0.659
0.493
0.549
0.479
ai4privacy_ko
0.739
0.322
0.523
0.611
0.545
0.554
0.460
ai4privacy_ms
0.839
0.692
0.586
0.598
0.547
0.640
0.587
ai4privacy_pt
0.855
0.692
0.691
0.643
0.687
0.756
0.651
ai4privacy_vi
0.880
0.628
0.615
0.644
0.499
0.634
0.615
ai4privacy_zh
0.816
0.453
0.600
0.643
0.428
0.572
0.548
creddata_en
0.590
0.357
0.226
0.035
0.150
0.370
0.263
gretel_en
0.872
0.588
0.734
0.743
0.715
0.697
0.729
nemotron_en
0.956
0.692
0.792
0.837
0.878
0.778
0.574
v2-eval
0.863
0.478
0.504
0.525
0.260
0.523
0.277
v2-eval-challenge
0.893
0.473
0.568
0.551
0.228
0.557
0.260
Macro average
0.833
0.581
0.604
0.602
0.538
0.644
0.536
Exact and containment F1 by entity family
Entity family
Exact F1
Containment F1
address
0.681
0.866
company_name
0.939
0.946
date
0.957
0.961
email_address
0.984
0.990
human_name
0.917
0.951
id_number
0.954
0.958
phone_number
0.919
0.951
private_url
0.902
0.925
secret
0.913
0.936
The benchmark combines synthetic and public corpora. A strong benchmark score does not tell you how the model behaves inside one hospital. Evaluate it on local document types before deployment.
Training used the Meddies PII v2 dataset and external public corpora rebuilt from their original sources. We do not rehost those external corpora. Checkpoint selection used a pinned control set before the full benchmark. The complete recipe and release checks live in TRAINING.md.
Where it fits
Review before redaction for clinical notes, referrals, messages, and exported records.
Span extraction inside a de-identification workflow that keeps the source text visible to a reviewer.
Multilingual evaluation and domain adaptation for local document formats.
Limits
Postal addresses reached 0.681 exact F1 and 0.866 containment F1. Expect trimmed or extended boundaries.
The label set is closed. An unsupported identifier will not receive its own label.
Hospital templates, abbreviations, copied headers, and optical character recognition errors vary by site.
Synthetic and public benchmark data cannot establish performance on every patient population or production system.
Inputs designed to evade detection can still break the model.
License
MeddiesAI applies CC BY-NC 4.0 to the copyright and related rights it can license in this release. See LICENSE.
The underlying LiquidAI materials remain under LFM Open License v1.0. Upstream attribution and the MeddiesAI change record are preserved in NOTICE. Review both licenses before redistribution.
Feedback
Found a bad miss or a wrong boundary? Share the language, document type, expected span, predicted span, and whether optical character recognition was involved. Use a safe synthetic example. Never post real patient text or credentials in a public issue.
We are looking for hospitals, research groups, and privacy teams that can contribute de-identified failure cases, independent evaluations, or review in underrepresented languages. Email contact@meddies.ai to work with us.
