teeem-pii-ko-1.2b

Teeem-pii-ko-1.2b

Korean enterprise PII detection — fine-tuned EXAONE 4.0 1.2B with a regex layer in front for structured types. Built and used in production by Teeem.ai.kr.

Final score on the 230-prompt eval (hybrid pipeline):

9 of 12 PII types are at F1 = 1.000 in the hybrid pipeline.

What this is

A two-stage Korean PII detection system designed to be dropped in front of an LLM so you can mask sensitive data before it leaves your perimeter and unmask it on the way back:

user text  →  [regex layer]  →  [EXAONE LoRA]  →  merge  →  masked text  →  upstream LLM
                                                                  ↓
                                                              mappings
                                                                  ↓
upstream response  ←  unmask  ←  [reverse mappings]

The split is deliberate. Structured PII is a regex problem — phone numbers, RRNs, business registration numbers, account numbers, emails, cards. The ML model is reserved for what regex cannot do reliably: Korean person names, free-form addresses, and organization names. This is the same architecture used by AWS Comprehend, GCP DLP, and Microsoft Presidio.

Per-type performance (hybrid, 230-prompt eval)

ADDRESS is the weakest type — it's the only category where the model has to do free-form span identification with no structural anchor. Future iterations should target it with a dedicated address gazetteer or a separate ADDRESS-only adapter.

Repo contents

Teeem-pii-ko-1.2b/
├── config.json              # EXAONE 4.0 1.2B config
├── generation_config.json
├── model.safetensors        # 2.4 GB merged weights (LoRA folded in)
├── tokenizer.json
├── tokenizer_config.json
├── chat_template.jinja      # EXAONE 4 chat template
├── regex_layer.py           # Python regex layer (for hybrid pipeline)
├── hybrid_pipeline.py       # Reference Python implementation
├── patterns_typescript/     # TS regex patterns (Teeem gateway version)
└── README.md

Quick start (Python, hybrid pipeline)

from transformers import AutoModelForCausalLM, AutoTokenizer
from regex_layer import detect_regex, merge_with_ml
import json, re

MODEL = "FlowOS2026/Teeem-pii-ko-1.2b"
tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
mdl = AutoModelForCausalLM.from_pretrained(MODEL, trust_remote_code=True, torch_dtype="auto", device_map="auto")

SYSTEM = ("You are a Korean PII detection model. Return a JSON array of detected PII "
          "entities with type, value, start, end. Types: NAME, PHONE, ADDRESS, RRN, "
          "CARD, BRN, PASSPORT, LICENSE, HEALTH_INSURANCE, ACCOUNT, ORGANIZATION, EMAIL.")

def detect_pii(text: str):
    # 1. Regex first (deterministic, high precision)
    regex_hits = detect_regex(text)

    # 2. ML for the unstructured types
    prompt = f"[|system|]{SYSTEM}[|endofturn|][|user|]{text}[|endofturn|][|assistant|]"
    inputs = tok(prompt, return_tensors="pt").to(mdl.device)
    out = mdl.generate(**inputs, max_new_tokens=512, temperature=0, do_sample=False)
    raw = tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    ml_hits = []
    m = re.search(r"\[[\s\S]*\]", raw)
    if m:
        try:
            ml_hits = json.loads(m.group(0))
        except Exception:
            pass

    # 3. Merge — regex priority on structured types, drop hallucinated types
    return merge_with_ml(regex_hits, ml_hits)

print(detect_pii("홍길동 고객님 010-1234-5678 카카오뱅크 3333-12-3456789"))

Quick start (vLLM serving)

# vLLM 0.6+ supports EXAONE 4.0 natively
pip install "vllm>=0.6.0"

vllm serve FlowOS2026/Teeem-pii-ko-1.2b \
    --port 8091 \
    --max-model-len 8192 \
    --served-model-name exaone-pii \
    --trust-remote-code

# Then call /v1/completions or /v1/chat/completions

For the production gateway (with the regex layer wired in front, mask/unmask, session-scoped mappings, optional AES-256-GCM encryption), use the Teeem PII Gateway: packages/pii-gateway/ in the Teeem monorepo. The TypeScript regex implementation is mirrored here in patterns_typescript/.

Self-hosted deployment recipe

The reference deployment runs on AWS ECS with a g4dn.2xlarge GPU host. You can replicate this anywhere with a 16+ GB GPU.

Container layout (two-container task):

1. exaone-vllm — vLLM 0.6+ serving the model on localhost:8091
2. gateway-proxy — Node.js process running the regex layer + EXAONE client + mask/unmask pipeline, listening on :8090, forwarding to upstream LLM

Cold-start time: ~3-4 minutes (most of which is downloading the 2.4 GB safetensors). Use a persistent volume / cache directory if you spin the service up and down often.

Spin up / spin down (ECS example):

REGION=ap-northeast-2
CLUSTER=Teeem-platform
SERVICE=Teeem-pii-gateway

# Spin up
aws ecs update-service --region $REGION --cluster $CLUSTER \
    --service $SERVICE --desired-count 1
# Also scale the EC2 capacity provider ASG up
aws autoscaling update-auto-scaling-group --region $REGION \
    --auto-scaling-group-name Teeem-pii-gateway-asg \
    --min-size 1 --desired-capacity 1

# Spin down
aws ecs update-service --region $REGION --cluster $CLUSTER \
    --service $SERVICE --desired-count 0
aws autoscaling update-auto-scaling-group --region $REGION \
    --auto-scaling-group-name Teeem-pii-gateway-asg \
    --min-size 0 --desired-capacity 0

# Force a redeploy (after updating the model weights)
aws ecs update-service --region $REGION --cluster $CLUSTER \
    --service $SERVICE --force-new-deployment

Training

Base model: LGAI-EXAONE/EXAONE-4.0-1.2B Method: LoRA (PEFT) — r=32, alpha=64 Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj Hardware: AWS g6e.xlarge (NVIDIA L40S 48 GB), bf16 Optimizer: adamw_torch_fused, lr 8e-5, batch 4 × grad_accum 2 Steps per iteration: 400 Total iterations: 14

Each iteration: generate fresh augmentation (generate_aug.py) → train on aug + replay buffer → merge LoRA → eval → analyze failures → adjust templates → repeat.

The full training data, replay buffer, scripts, and per-iteration metrics live in the project's S3 bucket — they are not in this HF repo because they contain templated synthetic Korean PII.

Iteration history (highlights)

The "stuck ACCOUNT" story

For five iterations, ACCOUNT recall sat at 0/3 on the original 30-prompt eval. We thought it was a vocabulary problem, then a regex-vs-NN problem, then a context problem. None of those explained it. The actual cause was LoRA capacity — r=16 with attention-only target modules wasn't enough to learn the digit-pattern → ACCOUNT mapping for novel bank names. Bumping to r=32 and adding the MLP target modules (gate_proj, up_proj, down_proj) unlocked it in one iteration.

The lesson: when a single PII type is stuck while everything else trains fine, don't add more training data — first check whether your adapter has enough capacity to represent the pattern at all.

The "regex breakthrough"

After iter 11, the model was plateauing around F1 ≈ 0.87 on the 230-prompt eval. Each iteration overfit a slightly different bank vocabulary or phone format. We wired in a regex layer purely as a defensive measure — and ACCOUNT recall jumped from 0.703 (26/37) to 1.000 (37/37) in a single rescore, with zero false positives. EMAIL went from 34/42 to 42/42 the same way.

The lesson: this is a hybrid problem, not an ML problem. The structured types didn't need a smarter model; they needed to not be the model's responsibility.

License

This model is a fine-tune of LGAI-EXAONE/EXAONE-4.0-1.2B and inherits the EXAONE AI License. Read it carefully before using commercially: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B/blob/main/LICENSE

The Teeem additions (regex layer, training scripts, gateway code) are released under the same license to keep the package self-consistent.

Citation

@misc{Teeem_pii_ko_1.2b_2026,
    title  = {Teeem-pii-ko-1.2b: Korean Enterprise PII Detection via Hybrid Regex + Fine-tuned EXAONE 4.0},
    author = {Teeem / FlowOS},
    year   = {2026},
    url    = {https://huggingface.co/FlowOS2026/Teeem-pii-ko-1.2b}
}

Maintainer

Teeem.ai.kr — Korean enterprise AI agent platform by FlowOS.