carl213/advbench

---
pretty_name: "AdvBench"
license: "mit"
language:
- en
tags:
- adversarial
- safety
- text
task_categories:
- question-answering
---

# AdvBench

This repository hosts a copy of the widely used **AdvBench** dataset,
a benchmark for evaluating the adversarial robustness and safety alignment of Large Language Models (LLMs).

AdvBench consists of adversarial prompts designed to elicit unsafe, harmful, or policy-violating responses from LLMs.
It is used in many LLM safety and jailbreak research papers as a standard evaluation dataset.

## Contents
- `advbench.jsonl` (or your actual filename): the standard set of adversarial prompts.

Each entry contains:
```json
{
"instruction": "...",
"input": "",
"output": "..."
}
```

## Usage
```python
from datasets import load_dataset
ds = load_dataset("S3IC/AdvBench")
```

## Source
This dataset is taken from the public AdvBench release:
https://github.com/llm-attacks/AdvBench

## License
AdvBench is released under the original license provided by the authors.
See the LICENSE file for more details.