townboy/kpfbert-kdpii

Korean PII token-classification model fine-tuned from KPF/KPF-bert-ner on a KDPII-style dialogue dataset.

Dataset

• Source file: 연대1_PII_dataset_V3.json
• Documents: 4981
• Sentences: 53778
• Positive PII sentences: 19037
• Label count: 33

Training Setup

• Max length: 128
• Epochs: 4.0
• Learning rate: 2e-05
• Train batch size: 8
• Eval batch size: 8
• Device: cuda
• GPU: NVIDIA GeForce RTX 4060 Ti
• Mixed precision: auto
• Gradient checkpointing: True

Intended Use

This model is intended for Korean personally identifiable information detection in dialogue-like text. Typical labels include names, nicknames, account numbers, mobile numbers, emails, addresses, IDs, and related sensitive entities.

Quick Inference

from transformers import pipeline

pipe = pipeline(
    "token-classification",
    model="townboy/kpfbert-kdpii",
    aggregation_strategy="simple",
)

print(pipe("Phone 010-8661-5573, ID wanderingrabbit1"))

Notes

• The classification head is reinitialized for the KDPII label space.
• This checkpoint should be validated on your target product traffic before production use.