Qoamo-PII-Decision

A routing model that catches personal data in Korean requests before they leave the device. Built on skt/A.X-Encoder-base (ModernBERT, 149M parameters), it was fine-tuned as a decision model that answers questions about a sentence with a probability for every option.

It gives four decisions for a sentence in one call: whether it contains personal data, whether it contains a strong identifier, its tier and its types. The confidence of two of them decides whether the sentence is sent as is, masked and then sent, or blocked. Locating the values is not its job. For a sentence it marks as containing personal data, atonlee/koelectra-ko-pii-ner, built on the same tier1 and tier2 criteria, finds and masks the values. This keeps the fast filtering decision separate from precise masking by NER.

On the official KDPII test split, it had the highest F1 (0.896) among the compared models under 1B parameters at telling whether a sentence contains personal data, and it also scored above the 1.4B models. The weights are a single safetensors file. ONNX exports in four precisions are in onnx/ (float32 · FP16 · INT8 · UINT8).

Decisions

The model can be asked four things about a sentence. Ask the ones you need in one call and get a probability for every option.

• pii_any — contains personal data? (예 yes · 아니오 no): on "no", send as is; on "yes", find the values with NER and mask them.
• pii_t1 — contains a strong identifier? (예 · 아니오): on "yes", do not send it off-device.
• pii_tier — tier (tier1 · tier2 · none): picks the single highest tier in the sentence. Decide the handling for each tier.
• type_* — 23 types (예 · 아니오 for each type): set the policy by the kinds of values present.

Tiers and types

tier	meaning	recommended handling	types
tier1	the value alone identifies a person	do not send off-device	resident registration number, alien registration number, passport number, driver's licence number, i-PIN, other national ID numbers, card number, card expiry date, CVC, account number, password or authentication key
tier2	identifies a person when combined with other information	find and mask it with NER, then send	name, email, phone number, address, postal code, date of birth, age or gender, numbers assigned to a person (employee, member, student numbers and the like), user ID, URL, IP address, vehicle number
none	no personal data	send as is	—

The same sentences, four decisions

sentence	contains personal data? (yes)	strong identifier (yes)	tier	types (yes > 0.5)
박지훈 대리한테 010-4821-3307 로 문자 보내 줘	0.998	0.000	tier2 (0.998)	name 0.993 · phone number 0.993
내 주민번호 901023-1234567 로 등록	0.995	0.985	tier1 (0.983)	resident registration number 0.901
이번 주말에 비 오는지 알려 줘	0.000	0.000	none (0.998)	none

Approach — borrowing the Jev format · NLI questions

The decision format borrows Jev's approach. A question and its options are attached to one text to judge, and the model gives a score for every option.

25 of the 26 questions are yes/no questions in NLI (natural language inference) form. They ask "이 글은 다음에 해당하는가: {criterion}" ("Does this text match the following: {criterion}"), with the text to judge as the premise and the criterion as the hypothesis. Only the tier question (pii_tier) instead picks one of tier1 · tier2 · none. The text the model actually reads looks like this:

문장: 박지훈 대리한테 010-4821-3307 로 문자 보내 줘
질문: 이 글은 다음에 해당하는가: 개인정보 형식 — 이름·전화번호·이메일·주소(…)·… 가 적혀 있다. …
[OPT] 예
[OPT] 아니오

Questions must be taken from the 26 in pii_config.json. Do not write new question texts; use the texts in that file as they are.

Usage

Ask the two questions (pii_any · pii_t1) in one call and pick the route (보냄 send · 확인 check · 막음 block).

pip install "transformers>=4.57" torch safetensors huggingface_hub

import json
import torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoConfig, AutoModel, AutoTokenizer

path = snapshot_download("atonlee/Qoamo-PII-Decision")
cfg = json.load(open(f"{path}/pii_config.json", encoding="utf-8"))
tok = AutoTokenizer.from_pretrained(path)
state = load_file(f"{path}/model.safetensors")
head_w, head_b = state.pop("head.weight"), state.pop("head.bias")
model = AutoModel.from_config(AutoConfig.from_pretrained(path), dtype=torch.float32)
model.load_state_dict(state)
model.eval()
marker = tok.convert_tokens_to_ids(cfg["marker"])


def ask(sentence, questions=("pii_any", "pii_t1")):
    texts = []
    for q in questions:
        c = cfg["questions"][q]
        options = "\n".join(f"{cfg['marker']} {line}" for line in c["option_lines"])
        texts.append(f"문장: {sentence}\n질문: {c['instructions']}\n{options}")
    enc = tok(texts, padding=True, return_tensors="pt")
    with torch.inference_mode():
        h = model(input_ids=enc["input_ids"], attention_mask=enc["attention_mask"]).last_hidden_state
        scores = (h @ head_w.T + head_b).squeeze(-1)
    out = {}
    for i, q in enumerate(questions):
        z = scores[i][enc["input_ids"][i] == marker] / cfg["temperature"]
        out[q] = dict(zip(cfg["questions"][q]["options"], torch.softmax(z, -1).tolist()))
    return out


def route(p, tau=0.97):
    conf = lambda d: (len(d) * max(d.values()) - 1) / (len(d) - 1)
    a, t = p["pii_any"], p["pii_t1"]
    if a["아니오"] > a["예"] and conf(a) > tau:
        return "보냄"
    if t["예"] >= t["아니오"] and conf(t) > 0.9:
        return "막음"
    return "확인"


for s in ["내일 오후 세 시에 팀 회의 잡아 줘",
          "박지훈 대리한테 010-4821-3307 로 문자 보내 줘",
          "내 주민번호 901023-1234567 로 등록"]:
    p = ask(s)
    print(route(p), {q: {k: round(v, 4) for k, v in d.items()} for q, d in p.items()})

Output:

보냄 {'pii_any': {'예': 0.0008, '아니오': 0.9992}, 'pii_t1': {'예': 0.0002, '아니오': 0.9998}}
확인 {'pii_any': {'예': 0.9978, '아니오': 0.0022}, 'pii_t1': {'예': 0.0002, '아니오': 0.9998}}
막음 {'pii_any': {'예': 0.9948, '아니오': 0.0052}, 'pii_t1': {'예': 0.9849, '아니오': 0.0151}}

The tier and the types are asked the same way.

questions = ["pii_tier"] + [q for q in cfg["questions"] if q.startswith("type_")]
for s in ["내일 오후 세 시에 팀 회의 잡아 줘",
          "박지훈 대리한테 010-4821-3307 로 문자 보내 줘",
          "내 주민번호 901023-1234567 로 등록"]:
    p = ask(s, questions)
    tier = max(p["pii_tier"], key=p["pii_tier"].get)
    types = [q[5:] for q, d in p.items() if q.startswith("type_") and d["예"] > 0.5]
    print(tier, types)

Output:

none []
tier2 ['이름', '전화번호']
tier1 ['주민등록번호']

• For a question with K options, confidence is (K × largest probability − 1) / (K − 1).
• When the route is check, find and mask the personal data with NER, then send.
• To judge several sentences, build the text for each one and pass them to the model in one batch.
• For a long text, split it into lines or sentences, judge each piece, and give the whole text the strictest route (block, then check, then send).

Evaluation

Whether a sentence contains personal data was measured on the official test split of KDPII (Yonsei University HamSaeM Kim's Lab · TSCIENTIFIC, CC BY 4.0). Sentences were selected based on our types and the 10 common types (3,998 sentences for our types, 3,737 for the 10 types). KDPII was not used to train this model. Widely used PII detection, decision and guard models on Hugging Face were run on the same sentences and scored the same way.

model	size	precision	recall	F1 (our types)	F1 (10 types)
Qoamo-PII-Decision	149M	0.895	0.897	0.896	0.918
perplexity-ai/PII-Tracer	about 600M	0.665	0.783	0.719	0.739
LiquidAI/LFM2.5-Encoder-350M-PII-Detector	350M	0.940	0.503	0.655	0.738
convaiinnovations/laya (Korean question)	322M	0.205	0.998	0.340	0.259
Qwen/Qwen3Guard-Gen-0.6B	0.6B	0.556	0.231	0.326	0.357
SupersonicLabs/Julia-1 (English question)	144M	0.178	0.523	0.266	0.205
fastino/GLiNER2.5-multi-Decide (English question)	287M	0.279	0.217	0.244	0.214
openai/privacy-filter (reference)	1.4B	0.813	0.612	0.698	0.775
OpenMed/privacy-filter-multilingual (reference)	1.4B	0.530	0.869	0.658	0.602

• F1 (our types): name · phone · mobile · email · vehicle number · IP · user ID · date of birth · age · gender · resident registration number · alien registration number · passport number · driver's licence number · card number · account number
• F1 (10 types): name · phone · mobile · email · resident registration number · alien registration number · passport number · driver's licence number · card number · account number

Dataset

Training data includes the following (CC BY 4.0).

• BCCard/pii-masking-openpii-finance
• townboy/korean-pii-dataset
• AmazonScience/massive

License

Apache-2.0