Downloads · 30 days
12
6% of all-time downloads
kdt-2-team4-newbiz/kcelectra-smishing-classifier
kcelectra-smishing-classifier is a text classification model from kdt-2-team4-newbiz. Use it when you need a label for a piece of text. It is set up for transformers.
Korean SMS smishing binary classifier fine-tuned from KcELECTRA.
Downloads · 30 days
12
6% of all-time downloads
All-time downloads
197
Public
Parameters
109M
873 MB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors436 MB · 99%
From the Hugging Face model README
Korean SMS smishing binary classifier fine-tuned from KcELECTRA.
The current Hub revision uses the hard_mixed_v3 training result:
modeling_help_local/encoder_test/results_hard_mixed_v3/
focal_no_oversampling/final_model
hard-mixed-v3The additional samples were introduced to reduce over-reliance on isolated risk-related words while preserving phishing recall.
Text classification for Korean SMS smishing detection.
This model should support a larger service pipeline rather than independently deciding whether a URL or sender is safe. Static URL checks, reporting history, and user-facing risk policies should be handled by the surrounding service.
| ID | Label |
|---|---|
| 0 | normal |
| 1 | phishing |
The service applies the same text preprocessing used during training before
calling the model. For example, URLs and eligible phone or money expressions
may be replaced by special tokens such as <URL>, <PHONE>, and <MONEY>.
Hugging Face text-classification payload:
{
"inputs": "배송 주소 오류로 반송 예정입니다. <URL>"
}
Example:
[
{
"label": "phishing",
"score": 0.92
}
]
score is the confidence for the returned label.
normal with 0.948 means 94.8% confidence that the message is normal.phishing with 0.920 means 92.0% confidence that the message is phishing.The score must not always be interpreted as a phishing risk score. The service wrapper converts the output into its normalized API response and applies its own user-facing risk policy.
Evaluation source:
results_hard_mixed_v3/focal_no_oversampling/test_metrics.json
| Metric | Value |
|---|---|
| Accuracy | 0.989669 |
| F1 | 0.956472 |
| Precision | 0.954429 |
| Recall | 0.958524 |
| False positives | 160 |
| False negatives | 145 |
| True negatives | 25,866 |
| True positives | 3,351 |