license: apache-2.0
dataset_info:
features:
- name: instruction
dtype: string - name: output
dtype: string
splits: - name: Benign
num_bytes: 28426
num_examples: 50 - name: grad_similarity
num_bytes: 57405
num_examples: 50 - name: Guardrail_jailbreak
num_bytes: 59896
num_examples: 50 - name: Virus
num_bytes: 57807
num_examples: 50
download_size: 135404
dataset_size: 203534
configs: - config_name: default
data_files:- split: Benign
path: data/Benign-* - split: grad_similarity
path: data/grad_similarity-* - split: Guardrail_jailbreak
path: data/Guardrail_jailbreak-* - split: Virus
path: data/Virus-*
- split: Benign
There are four datasets avaialble
- Benign data. It is the original GSM8K data.
- Gradient similarity. It is equal to Virus with $lambda$=0. The gradient of this data should resemble the original mixing harmful data.
- Guardrail jailbreak. It is equal to Virus with $lambda$=1. This dataset should not be detected as harmful by the llama guard2 model.
- Virus. It is equal to Virus with $lambda$=0.1. This dataset is produced by dual goal optimization, such that i) its gradient resembles the orginal harmful gradient, ii) it can bypass llama guard2 detection.