harmbench

# HarmBench

- Original repo: https://github.com/centerforaisafety/HarmBench/tree/main/data/behavior_datasets

This HF dataset contains two types of harmful prompts:
- DirectRequest: taken from [harmbench_behaviors_text_test.csv](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_test.csv) (test split) and [harmbench_behaviors_text_val.csv](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_val.csv) (val split)
- HumanJailbreaks: manual selection of jailbreaks divided into test and val splits
- Moreover, `metadata.csv` contains the information about each behavior and can be mapped to the prompts above by BehaviourID ([original_file](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_all.csv))

How to load the data:

```
load_dataset("swiss-ai/harmbench", "DirectRequest", split="test")
load_dataset("swiss-ai/harmbench", "HumanJailbreaks", split="test")
```

# HarmBench
- 原始代码仓库：https://github.com/centerforaisafety/HarmBench/tree/main/data/behavior_datasets

该 Hugging Face（HF）数据集包含两类有害提示词：
- 直接请求（DirectRequest）：数据来源于[harmbench_behaviors_text_test.csv](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_test.csv)（测试子集）与[harmbench_behaviors_text_val.csv](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_val.csv)（验证子集）
- 人工越狱提示（HumanJailbreaks）：经人工筛选的越狱类提示，已划分为测试子集与验证子集
此外，`metadata.csv` 包含各行为的元数据信息，可通过行为ID（BehaviourID）与上述提示词进行关联，其原始完整数据文件为[harmbench_behaviors_text_all.csv](https://github.com/centerforaisafety/HarmBench/blob/main/data/behavior_datasets/harmbench_behaviors_text_all.csv)

数据加载方式如下：

load_dataset("swiss-ai/harmbench", "DirectRequest", split="test")
load_dataset("swiss-ai/harmbench", "HumanJailbreaks", split="test")