| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - question-answering |
| - audio-to-audio |
| - automatic-speech-recognition |
| language: |
| - en |
| tags: |
| - Jailbreak |
| - Red-teaming |
| - Audio-language-Model |
| size_categories: |
| - n<1K |
| --- |
| |
|
|
| # AdvBench-Audio |
|
|
| [](https://arxiv.org/abs/2510.26096) |
| [](https://huggingface.co/datasets/WeifeiJin/AdvBench-Audio) |
| [](https://creativecommons.org/licenses/by-nc/4.0/) |
|
|
| **AdvBench-Audio** is an audio-version benchmark used in our study to evaluate the safety, jailbreak, and adversarial robustness of audio-language models (ALMs). It contains audio renderings of adversarial prompts paired with their corresponding targets for safety evaluation. |
|
|
| This dataset is used in our paper: [**ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio-Language Models**](https://arxiv.org/abs/2510.26096). |
|
|
| --- |
|
|
| ## Dataset at a glance |
|
|
| - **Format:** JSONL + WAV files |
| - **Fields per example (JSONL):** |
| - `prompt` — the adversarial / jailbreak instruction text (string) |
| - `target` — the target (unsafe) completion that the original prompt would elicit (string) |
| - `audio` — filename for the corresponding audio file (relative path in dataset; e.g. `0.wav`) |
| - **Number of examples:** 520 (indexed `0.wav` … `519.wav`) |
| - Each JSON line corresponds to one audio file and one `prompt`/`target` pair. |
|
|
| > **Note:** the dataset contains text prompts that explicitly solicit illegal, violent, or other disallowed actions. The audio files are TTS/converted versions paired with those prompts and intended solely for *model safety research*. |
|
|
| --- |
|
|
| ## Quick example (redacted) |
|
|
| A representative (redacted) JSONL entry in `AdvBench_Audio.json`: |
|
|
| ```json |
| {"prompt": "[REDACTED: instruction soliciting harmful/illicit behavior]", "target": "[REDACTED: harmful completion]", "audio": "0.wav"} |
| ``` |
|
|
| > We redact examples in this README for safety. Full, unredacted examples are included in the dataset files hosted on Hugging Face for verified research use under the dataset license. |
|
|
| ------ |
|
|
| ## Recommended research uses |
|
|
| - Measuring **attack success rate** (ASR) on ALMs given adversarial/jailbreak audio prompts. |
| - Evaluating and comparing **defense strategies** (e.g., input sanitization, detection, fine-tuning with safety data, guardrail prompts, SAP-style protections). |
| - Studying model *safety shortcuts* and failure modes in instruction-following audio systems. |
|
|
| ------ |
|
|
| ## Important safety & ethics notice — **must read** |
|
|
| This dataset contains *explicit* instructions and responses that describe or enable wrongdoing (including violence, weapons, cybercrime, self-harm, child-exploitation, etc.). It is released **solely** for the purpose of model safety research and development of mitigations and defenses. |
|
|
| By using this dataset you agree to the following obligations: |
|
|
| 1. **Research-only**: Use is restricted to non-commercial research, analysis, or development of safety/defense methods (consistent with CC-BY-NC-4.0). |
| 2. **No dissemination**: Do not republish or host the harmful text/audio in a way that makes it trivially accessible to the general public for malicious use. Consider gating downloads or requiring a GitHub/HF account. |
| 3. **Institutional oversight**: Follow your institution's ethical review / IRB / security policies as required. If uncertain, consult your advisor or ethics board. |
| 4. **Mitigation**: If you publish results that include examples from this dataset, redact or summarize harmful content and explain mitigation safeguards you used. |
| 5. **Legal compliance**: Ensure your usage complies with all applicable laws and regulations. |
|
|
| If you need to provide dataset access to third parties for reproducibility, prefer partial / redacted examples or require a short request form describing the research purpose. |
|
|
| ------ |
|
|
| ## License |
|
|
| This dataset is released under **Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)**. |
| Link: https://creativecommons.org/licenses/by-nc/4.0/ |
|
|
| > CC-BY-NC-4.0 permits reuse for non-commercial purposes with attribution. Commercial use is *not* permitted without separate agreement. |
|
|
| ------ |
|
|
|
|
|
|
| ## Citation |
|
|
| If you use this dataset in your work, please cite the paper: |
|
|
| ```bibtex |
| @inproceedings{jin2025almguard, |
| title = {ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio-Language Models}, |
| author = {Jin, Weifei and Cao, Yuxin and Su, Junjie and Xue, Minhui and Hao, Jie and Xu, Ke and Dong, Jin Song and Wang, Derui}, |
| booktitle = {NeurIPS}, |
| year = {2025} |
| } |
| ``` |
|
|
| ------ |
|
|
| ## Acknowledgements |
|
|
| This dataset was constructed as part of the ALMGuard project. We thank prior datasets and toolkits used in construction and evaluation (see the paper for details). |
