GolemPII v2: Hebrew PII detection

Demo: try it Blog: read the write-up GitHub: code and examples Parameters: 277M REDACT F1: 0.917 Language: Hebrew Licence: MIT Cite: BibTeX Author: liranbaba.dev

GolemPII v2 base

Version 2.0, 2026-11-13.

A token classifier that finds personal information in Hebrew text, including Hebrew mixed with English as it appears in enterprise systems: emails, tickets, chat, key-value records, JSON and logs.

  • Encoder: xlm-roberta-base (277M parameters)
  • Labels: 15 (below), BIO tagging
  • Languages: Hebrew, and English embedded in Hebrew text
  • Replaces: CordwainerSmith/GolemPII-v1

Which model should I use

Model Parameters Use it for
GolemPII-v2-base (this model) 277M Hebrew mixed with English in enterprise text (emails, tickets, logs, JSON), and the best recall of the three on identity, card and bank numbers
GolemPII-v2-small 125M Hebrew prose: the best of the three at telling person names from organisations, at less than half the size of base and the same speed
GolemPII-v2-tiny 45M When speed or size matters most (about five times faster), with lower accuracy on enterprise text

Labels

Group Labels
Person FIRST_NAME, LAST_NAME
Identity documents ID_NUM, PASSPORT_NUM, DRIVER_LICENSE_NUM
Financial BANK_ACCOUNT_NUM, IBAN, CC_NUM, CC_PROVIDER
Contact PHONE_NUM, EMAIL
Address STREET, CITY, POSTAL_CODE
Other DATE

Masked card numbers (**** **** **** 1234) are labelled CC_NUM. Organisations, job titles, countries, and reference numbers such as order, ticket and employee numbers are not labelled.

Usage

from transformers import pipeline

detector = pipeline(
    "token-classification",
    model="CordwainerSmith/GolemPII-v2-base",
    aggregation_strategy="simple",
    stride=128,  # lets the pipeline handle texts longer than 512 tokens
)

TEXT = """נושא: RE: ticket #48213 - Χ’Χ“Χ›Χ•ΧŸ Χ€Χ¨Χ˜Χ™ ΧœΧ§Χ•Χ—

Χ”Χ™Χ™ Χ¨Χ•Χ Χ™Χͺ,
Χ”ΧœΧ§Χ•Χ— יואב Χ‘ΧŸ Χ“Χ•Χ“ (Χͺ.Χ–. 034567891) Χ”ΧͺΧ§Χ©Χ¨ Χ”Χ‘Χ•Χ§Χ¨ Χ•Χ‘Χ™Χ§Χ© ΧœΧ’Χ“Χ›ΧŸ אΧͺ Χ”Χ›ΧͺΧ•Χ‘Χͺ ΧœΧžΧ©ΧœΧ•Χ—: Χ¨Χ—Χ•Χ‘ Χ”Χ¨Χ¦Χœ 42, רמΧͺ Χ’ΧŸ, ΧžΧ™Χ§Χ•Χ“ 5244302.
א׀שר ΧœΧ”Χ©Χ™Χ’ אוΧͺΧ• Χ‘-054-7731920 או Χ‘ΧžΧ™Χ™Χœ yoav.bendavid@example.com.
Χ”Χ—Χ™Χ•Χ‘ Χ”Χ—Χ•Χ“Χ©Χ™ Χ™Χ¨Χ“ ΧžΧ›Χ¨Χ˜Χ™Χ‘ Χ•Χ™Χ–Χ” 4580 1234 5678 9015, Χ•Χ”Χ–Χ™Χ›Χ•Χ™ Χ™Χ•Χ’Χ‘Χ¨ ΧœΧ—Χ©Χ‘Χ•ΧŸ IL620108000000099999999.
order_id=ORD-2024-88127, בכום: 349.90 Χ©"Χ—, ΧͺΧΧ¨Χ™Χš ΧœΧ™Χ“Χ” ΧœΧΧ™ΧžΧ•Χͺ: 14/03/1987.

ΧͺΧ•Χ“Χ”,
Dana Levi | Customer Success"""

EDGE_PUNCTUATION = " .,;:!?()[]{}\"'-"


def clean_spans(text, entities):
    """Trim punctuation from span edges and join spans of one label that touch or overlap."""
    spans = []
    for entity in entities:
        start, end, label = entity["start"], entity["end"], entity["entity_group"]
        while start < end and text[start] in EDGE_PUNCTUATION:
            start += 1
        while end > start and text[end - 1] in EDGE_PUNCTUATION:
            end -= 1
        if start == end:
            continue
        if spans and spans[-1][2] == label and start <= spans[-1][1]:
            spans[-1] = (spans[-1][0], max(end, spans[-1][1]), label)
        else:
            spans.append((start, end, label))
    return spans


def redact(text, spans):
    pieces, position = [], 0
    for start, end, label in spans:
        pieces += [text[position:start], f"[{label}]"]
        position = end
    return "".join(pieces) + text[position:]


spans = clean_spans(TEXT, detector(TEXT))
for start, end, label in spans:
    print(f"{label:<12} {TEXT[start:end]}")
print(redact(TEXT, spans))

The detected spans:

FIRST_NAME   Χ¨Χ•Χ Χ™Χͺ
FIRST_NAME   יואב
LAST_NAME    Χ‘ΧŸ Χ“Χ•Χ“
ID_NUM       034567891
STREET       Χ¨Χ—Χ•Χ‘ Χ”Χ¨Χ¦Χœ 42
CITY         רמΧͺ Χ’ΧŸ
POSTAL_CODE  5244302
PHONE_NUM    054-7731920
EMAIL        yoav.bendavid@example.com
CC_PROVIDER  Χ•Χ™Χ–Χ”
CC_NUM       4580 1234 5678 9015
IBAN         IL620108000000099999999
DATE         14/03/1987
FIRST_NAME   Dana
LAST_NAME    Levi

The redacted text. The ticket number, the order id and the amount are left alone:

נושא: RE: ticket #48213 - Χ’Χ“Χ›Χ•ΧŸ Χ€Χ¨Χ˜Χ™ ΧœΧ§Χ•Χ—

Χ”Χ™Χ™ [FIRST_NAME],
Χ”ΧœΧ§Χ•Χ— [FIRST_NAME] [LAST_NAME] (Χͺ.Χ–. [ID_NUM]) Χ”ΧͺΧ§Χ©Χ¨ Χ”Χ‘Χ•Χ§Χ¨ Χ•Χ‘Χ™Χ§Χ© ΧœΧ’Χ“Χ›ΧŸ אΧͺ Χ”Χ›ΧͺΧ•Χ‘Χͺ ΧœΧžΧ©ΧœΧ•Χ—: [STREET], [CITY], ΧžΧ™Χ§Χ•Χ“ [POSTAL_CODE].
א׀שר ΧœΧ”Χ©Χ™Χ’ אוΧͺΧ• Χ‘-[PHONE_NUM] או Χ‘ΧžΧ™Χ™Χœ [EMAIL].
Χ”Χ—Χ™Χ•Χ‘ Χ”Χ—Χ•Χ“Χ©Χ™ Χ™Χ¨Χ“ ΧžΧ›Χ¨Χ˜Χ™Χ‘ [CC_PROVIDER] [CC_NUM], Χ•Χ”Χ–Χ™Χ›Χ•Χ™ Χ™Χ•Χ’Χ‘Χ¨ ΧœΧ—Χ©Χ‘Χ•ΧŸ [IBAN].
order_id=ORD-2024-88127, בכום: 349.90 Χ©"Χ—, ΧͺΧΧ¨Χ™Χš ΧœΧ™Χ“Χ” ΧœΧΧ™ΧžΧ•Χͺ: [DATE].

ΧͺΧ•Χ“Χ”,
[FIRST_NAME] [LAST_NAME] | Customer Success

clean_spans is optional. The pipeline groups sub-word tokens, so a raw span can carry a bracket, a hyphen or a full stop at its edge, or arrive in two pieces; the function trims and joins them. The scores on this page were measured on the raw pipeline output, without it.

With Microsoft Presidio

A ready recognizer for Presidio is in the GitHub repo. It registers this model for language he, maps the labels to Presidio entity names, and works with the standard analyze and anonymize calls.

Results

All numbers below are from text the model was not trained on. REDACT is an external benchmark of enterprise-style Hebrew text. Its Hebrew part was split once into two halves: select, used for every decision during development, and final, scored once after the model was frozen. The numbers here are from final (311 records, 4,936 entities).

Two measures are reported. Character level counts every character of a span. Entity level counts an entity as found when a span of the right category touches it, so it does not depend on where span boundaries are drawn. REDACT has five categories; our 15 labels are mapped onto them, and a CITY span on its own is not counted as an address.

REDACT final, five categories

Model Parameters Character P / R / F1 Entity P / R / F1
GolemPII-v1 277M 0.592 / 0.884 / 0.709 0.601 / 0.913 / 0.725
GolemPII-v2-base 277M 0.886 / 0.951 / 0.917 0.872 / 0.983 / 0.924

A second training run with another seed scored 0.914 and 0.927.

REDACT final, by category (entity level, P / R / F1)

Category GolemPII-v1 GolemPII v2
Person 0.804 / 0.900 / 0.849 0.888 / 0.986 / 0.934
Address 0.597 / 1.000 / 0.747 0.777 / 1.000 / 0.874
Phone 0.238 / 0.969 / 0.383 0.940 / 0.993 / 0.966
Email 0.763 / 0.996 / 0.864 0.940 / 0.998 / 0.968
Account (identity, card and bank numbers) 0.265 / 0.844 / 0.403 0.762 / 0.905 / 0.827

REDACT final, by label (character F1)

Label Entities F1
EMAIL 454 0.975
PHONE_NUM 288 0.951
LAST_NAME 1,025 0.937
ID_NUM 203 0.932
FIRST_NAME 1,208 0.918
PASSPORT_NUM 35 0.915
CITY 261 0.745
CC_NUM 53 0.671
DATE 571 0.530
DRIVER_LICENSE_NUM 13 0.411

REDACT has no entities for BANK_ACCOUNT_NUM, IBAN, CC_PROVIDER, STREET and POSTAL_CODE under our mapping, so those labels have no external score. DATE and CITY conventions differ between REDACT and our labels, which lowers both.

Code-switching (REDACT final, character F1)

Amount of English in the text GolemPII-v1 GolemPII v2
None 0.704 0.930
Light 0.677 0.906
Heavy 0.739 0.916

Person names on real text

NEMO is a human-labelled corpus of Hebrew news sentences (6,143 sentences, 2,588 person mentions). It was never used for training, and it was read once, after the model was chosen.

Model Person P / R / F1 False phone, email and account spans per 1,000 sentences
GolemPII-v1 0.541 / 0.824 / 0.653 3.3
GolemPII-v2-base 0.853 / 0.925 / 0.887 1.3

A second seed scored 0.898 / 0.891 / 0.895. The training data includes real sentences from another news and encyclopedia corpus (IAHLT), which helps on text of this kind.

Names the model has never seen

On held-out templates filled with 1,436 first and last names that appear nowhere in training: FIRST_NAME F1 0.972, LAST_NAME F1 0.954.

Speed and size

Measured on an Apple M4 Pro (48 GB, macOS 27.0.1, torch 2.15.0, transformers 5.22.1) over 1,000 synthetic enterprise-style texts with a mean length of 1,512 characters, with the pipeline settings shown under Usage and batch size 8. Median of three runs; other programs were running, so treat differences of a few percent as noise.

Model Parameters File size Seconds per 1,000 texts, Apple GPU (MPS) Seconds per 1,000 texts, CPU
GolemPII-v2-base 277M 1,110 MB 63.8 115.6
GolemPII-v2-small 125M 502 MB 65.8 120.9
GolemPII-v2-tiny 45M 179 MB 13.0 15.6

The small model is less than half the size of the base model on disk and in memory, and runs at the same speed: both have a 12-layer encoder, and the base model's extra parameters are its larger vocabulary table. The tiny model is about five times faster on the GPU and seven times faster on the CPU.

Training data

The model is trained on synthetic text only for everything except names in context.

  • Templates: 3,000 texts with placeholders, written by an LLM (Gemini Flash) in enterprise formats, half Hebrew-only and half mixed with English. The LLM never writes a PII value. Each template is filled six times by a script: names, cities and streets from lexicons, numbers generated with valid checksums (Israeli identity numbers, cards, IBAN). Templates also contain placeholders for look-alike values that are not PII (order numbers, amounts, timestamps, IP addresses, company names), which are left unlabelled.
  • Real sentences with swapped names: 20,000 texts from the IAHLT named-entity corpus (news, Wikipedia, Knesset). In 70% of them every person name is replaced with a lexicon name; 30% keep their original names. Organisations and titles stay unlabelled. Places and time expressions are trained as "not a person name".
  • 38,000 training examples in total, 3 epochs, learning rate 3e-5, batch size 16.

No real personal data from any private source was used.

The real sentences come from the IAHLT Hebrew named-entity dataset (github.com/IAHLT/hebrew_named_entities_open_dataset, Apache-2.0).

Intended use and out of scope

Intended use: finding personal information in Hebrew text, and in Hebrew mixed with English, to assist redaction, masking and data-loss-prevention review.

Out of scope or not evaluated:

  • Other languages. Arabic, and English text on its own, were not evaluated.
  • Other sources. Speech transcripts, OCR output and handwriting were not evaluated.
  • Anonymisation. The model is a detector that assists redaction, not a guarantee of anonymisation or of compliance with any privacy law. Text that must be anonymised needs human review.
  • Other entity types. Organisations, job titles, countries and reference numbers (order, ticket and employee numbers) are not labelled.

Limits

  • Organisations named like people. On news text the model tags some sports clubs, companies and firms named after their founders as people. This is its main error on NEMO, and the amount differs between training runs.
  • Account numbers: 0.905 of the identity, card and bank numbers in REDACT are found. Reference numbers such as customer and student numbers are not labelled by this model, by design. Known misses: references to the last four digits of a card ("ending 4821"), tax file numbers, passport numbers inside a machine-readable line, and identity numbers glued to other text. Obfuscated emails ("name AT domain DOT com") are missed.
  • Driver licence and card numbers score low on REDACT (0.411 and 0.671 character F1, on 13 and 53 entities).
  • Dates: DATE scores 0.530 on REDACT, whose date labelling differs from ours; treat date detection as weak.
  • Span boundaries: address spans are often shorter than REDACT's (the street name without the word "street"), which lowers character-level address scores while the entities are still found.
  • Only names are tested on real text. Numeric PII in real enterprise text is tested only on REDACT, which is itself synthetic, and on our own held-out templates.

Smaller variants

Trained on the same data with the same settings, except that the tiny model is trained for 10 epochs; REDACT final.

Model Encoder Parameters Character F1 Entity F1 NEMO person P / R / F1
GolemPII-v2-base (this model) xlm-roberta-base 277M 0.917 0.924 0.853 / 0.925 / 0.887
GolemPII-v2-small HalleluBERT_base 125M 0.916 0.917 0.928 / 0.956 / 0.942
GolemPII-v2-tiny dictabert-tiny 45M 0.884 0.874 0.856 / 0.945 / 0.899

GolemPII-v2-small is stronger on Hebrew prose and weaker on account numbers. GolemPII-v2-tiny is about five times faster than this model on a GPU; GolemPII-v2-small runs at the same speed as this model.

Reproducing the numbers

The evaluation/ folder holds the scoring code used for every number on this page. With uv installed, from that folder:

# REDACT, character level (add --redact-split select for the decision half)
uv run compare.py --redact-split final --model v2=CordwainerSmith/GolemPII-v2-base:golempii_v2
# REDACT, entity level
uv run independent_checks.py --redact-split final --model v1=CordwainerSmith/GolemPII-v1:golempii --model v2=CordwainerSmith/GolemPII-v2-base:golempii_v2
# Person names on NEMO (needs the NEMO files, see the docstring of real_text_check.py)
uv run real_text_check.py --model v1=CordwainerSmith/GolemPII-v1:golempii --model v2=CordwainerSmith/GolemPII-v2-base:golempii_v2

REDACT is a gated dataset (guneeshv/REDACT-PII-Benchmark); the scripts need an HF_TOKEN with access to it. compare.py scores GolemPII-v1 by default; any other token-classification model can be added with --model NAME=REPO:LABEL_MAP, where the label map is golempii or golempii_v2.

Citation

If you use this model in a product, a paper or a derived model, please credit it as: GolemPII-v2-base by Liran Baba, with a link to https://huggingface.co/CordwainerSmith/GolemPII-v2-base.

@misc{baba2026golempii_v2_base,
  author       = {Baba, Liran},
  title        = {GolemPII v2 base: personal information detection for Hebrew text},
  year         = {2026},
  version      = {2.0},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/CordwainerSmith/GolemPII-v2-base}}
}

Licence and credits

MIT, copyright (c) 2026 Liran Baba. The LICENSE file holds the full text.

  • Encoder: FacebookAI/xlm-roberta-base (MIT).
  • Real training sentences: IAHLT Hebrew named-entity dataset, github.com/IAHLT/hebrew_named_entities_open_dataset (Apache-2.0), with person names replaced as described above.
  • Evaluation: REDACT PII Benchmark, NEMO corpus.
Downloads last month
37
Safetensors
Model size
0.3B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for CordwainerSmith/GolemPII-v2-base

Finetuned
(4,321)
this model

Space using CordwainerSmith/GolemPII-v2-base 1

Evaluation results