license: mit
configs:
- config_name: Coating_Data
data_files:- split: train
path: "Coating_Data/Coating_Data.parquet"
- split: train
- config_name: Non_Coating_Data
data_files:- split: train
path: "Non_Coating_Data/Non_Coating_Data.parquet"
- split: train
Korean_Euphemistic_Coating
Dataset Description
This dataset was constructed by translating the original walledai/AdvBench dataset into Korean using the DeepL API.
Data Configuration
Coating_Data
- These are harmful prompts generated by applying "coating" (euphemistic language) using an algorithm.
Non_Coating_Data
- These are harmful prompts that have been simply translated from the original data without additional modifications.
Aattack Success Rate
| Model | English (Original) | Korean (Non Coated) | Korean (Coated) |
|---|---|---|---|
| DeepSeek-R1-7B | 91.7 | 44.8 | 40.4 |
| Hermes-3-8B | 97.1 | 82.5 | 89.6 |
| Llama-3.1-Tulu-8B | 79.0 | 50.2 | 71.5 |
| Ministral-8B | 96.7 | 61.4 | 92.9 |
| OpenThinker-7B | 96.0 | 71.0 | 67.3 |
| Qwen2.5-7B | 81.3 | 17.3 | 61.0 |
| Qwen2.5-Math-7B | 76.9 | 1.9 | 0.8 |