dataset_info:
features:
- name: prompt
dtype: string - name: response
dtype: string - name: unsafe-score
dtype: float64
splits: - name: train
num_bytes: 141285890
num_examples: 297366
download_size: 72212904
dataset_size: 141285890
configs: - config_name: default
data_files:- split: train
path: data/train-*
- split: train
HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models
This dataset contains generated prompts and responses using HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models.
This dataset is also used for training our HarmAug Guard Model.
The unsafe-score is measured by Llama-Guard-3.
For rows without responses, the unsafe-score indicates the unsafeness of the prompt.
For rows with responses, the unsafe-score indicates the unsafeness of the response.
For more information, please refer to our anonymous github
