schift-ko-pii-v4

111M parameter Korean PII detector that outperforms models 3x its size.

Fine-tuned from klue/roberta-base with LoRA on Korean legal, financial, and administrative texts. Distributed as bf16 safetensors (210 MB).

Benchmark

Two benchmark suites are included in this repository under benchmark/.

benchmark_v1 (93 cases)

Same benchmark, same scoring (50% overlap match), no postprocessing on either side.

Model	Params	micro F1	strong F1	Size
schift-ko-pii-v4	111M	0.863	0.893	210 MB (bf16)
LFM2.5-PII-Detector	350M	0.663	0.645	~1.4 GB

benchmark_v2 (253 cases)

Expanded benchmark with harder cases: investigation reports, complex address formats, foreign/Hanja names, mixed-entity documents.

Model	Params	raw F1	raw P	raw R
LFM2.5-Encoder-350M-PII	350M	0.254	0.207	0.328
schift-ko-pii-v4	111M	0.702	0.602	0.842

With postprocessing enabled (postprocess=True):

Benchmark	F1	P	R
v1 (93 cases)	0.968	0.947	0.991
v2 (253 cases)	0.802	0.802	0.802

Category breakdown (benchmark_v1, raw model)

Category	schift-ko-pii	LFM2.5-PII
Person (standard)	1.00	0.67
Person (rare surnames)	1.00	0.60
Person (short names)	1.00	0.57
Phone	0.86	0.92
Email	1.00	0.86
Address (urban)	1.00	0.67
Address (rural)	1.00	0.60
Vehicle plate	1.00	0.00
Date (Korean)	1.00	0.00
Resident ID	1.00	0.75
Passport	1.00	1.00

LFM2.5 is a strong multilingual model, but it was not trained on Korean-specific formats: vehicle plates (52가1234), resident registration numbers (850205-1234567), Korean date expressions (2024년 3월 15일), or rare Korean surnames (황보, 남궁, 독고).

Quick start

pip install schift-ko-pii

from schift_ko_pii import detect, mask, apply

# Detect PII spans (each entity gets a sequential id)
spans = detect("피고 김민수의 전화번호는 010-1234-5678이다.")
# [
#   {"id": "person_1", "start": 3, "end": 6, "label": "private_person", ...},
#   {"id": "phone_1", "start": 14, "end": 27, "label": "phone_number", ...},
# ]

# Mask: get masked text + entity map in one call
result = mask("피의자 김철수가 피해자 박영희에게 서울시 강남구에서 금품을 전달하였다.")
print(result["masked"])
# 피의자 [사람1]가 피해자 [사람2]에게 [주소1]에서 금품을 전달하였다.

for e in result["entities"]:
    print(f"  {e['id']}: {e['text']}")
# person_1: 김철수
# person_2: 박영희
# address_1: 서울시 강남구

# Apply: replace selected entities (user can edit the map before applying)
text = apply(
    "피의자 김철수가 피해자 박영희에게 서울시 강남구에서 금품을 전달하였다.",
    result["entities"],
    replacements={"person_1": "OOO", "person_2": "△△△"}
    # address_1 omitted → left unmasked
)
# 피의자 OOO가 피해자 △△△에게 서울시 강남구에서 금품을 전달하였다.

# With postprocessing (regex validation + structured-ID rules)
spans = detect("주민번호 850205-1234567을 확인.", postprocess=True)

Postprocessing

The postprocess=True flag applies Korean-specific rules:

• Regex validation for structured IDs (resident numbers, passports, vehicle plates, bank accounts)
• Luhn/checksum verification where applicable
• Context-aware span merging
• False-positive suppression for legal case numbers and statute references

API (free)

For production use without managing model files:

from schift import Schift

client = Schift(api_key="...")  # free at schift.io
result = client.pii.redact("김민수의 전화번호는 010-1234-5678입니다.")
# Postprocessing is always enabled on the API.

Labels

Label	Description	Examples
private_person	Person names (Korean, Hanja, foreign)	김민수, 황보영희, Lee Jenny
private_phone	Phone numbers	010-1234-5678, 02-1234-5678
private_email	Email addresses	user@example.com
private_address	Street/postal addresses	서울특별시 강남구 테헤란로 521
private_date	Dates	2024년 3월 15일, 2024-03-15
private_url	URLs, SNS links, IP addresses	instagram.com/user, 192.168.1.1
account_number	Structured IDs: resident/business/bank/passport/vehicle	850205-1234567, M12345678
secret	Secrets, API keys, passwords

Running benchmarks

# Benchmark this model (v1, 93 cases)
python benchmark/run_benchmark.py

# Benchmark v2 (253 cases)
python benchmark/run_benchmark.py --benchmark benchmark/benchmark_v2.jsonl

# With postprocess
python benchmark/run_benchmark.py --postprocess

# Compare any HuggingFace model
python benchmark/run_benchmark.py --hf-model LiquidAI/LFM2.5-Encoder-350M-PII-Detector

Model details

• Base model: klue/roberta-base (111M params)
• Training: LoRA adapter on ~20k Korean legal/financial/admin examples
• Format: safetensors bf16 (210 MB)
• Inference: transformers pipeline, CPU or GPU
• Max length: 512 tokens
• Tagging scheme: BIES (Begin/Inside/End/Single)

License

Schift License v2.0 — Apache 2.0 base with a revenue threshold. Free for everyone under $10M annual revenue. Research, education, and non-profit use always permitted. Companies above the threshold: contact hello@schift.io.

Citation

@software{schift_ko_pii_2026,
  author = {Schift Inc.},
  title = {schift-ko-pii: Korean PII Detection Model},
  year = {2026},
  url = {https://huggingface.co/schift-io/schift-ko-pii-v4},
}