Multi-Aya-Redteaming
Content warning: this dataset intentionally contains harmful and adversarial prompts. It is released for safety alignment and robustness work only, and it should be handled under an appropriate content-review process.
Multi-Aya-Redteaming is the multilingual, SFT-ready version of Aya Red-teaming, the human-annotated red-teaming collection released by Cohere Labs with the paper The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm. It is released in English, Italian, French, German and Spanish and contains 3,661 samples in total.
The dataset is a supervised fine-tuning (SFT) resource for safety alignment. The original collection provides harmful prompts annotated for research, with no safe answers; here every prompt is paired with a safe target response, so that the adversarial material can be used directly as instruction-following training data rather than as an evaluation-only resource.
Content warning and intended use
The dataset is designed exclusively for safety alignment, refusal training and multilingual robustness research on a model trained by the project. It must not be used to generate, retrieve or distribute harmful content.
Source dataset
| Dataset | Aya Red-teaming |
| Hugging Face | CohereLabs/aya_redteaming |
| Paper | The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm (Aakanksha et al., 2024) |
| License | Apache 2.0 |
Aya Red-teaming was curated by professional annotators and consists of harmful prompts in 8 languages (Arabic, English, Filipino, French, Hindi, Russian, Serbian and Spanish) across 9 categories of harm — bullying and harassment, discrimination and injustice, graphic material, harms of representation allocation and quality of service, hate speech, non-consensual sexual content, profanity, self-harm, and violence, threats and incitement — with explicit labels for "global" and "local" harm, literal and semantic translations, and an explanation for each prompt. Because it was designed as an evaluation resource, it contains no safe responses.
How the dataset was built
The source prompts were repurposed for SFT by transforming them into instruction-following examples with safe target responses. Every sample was manually reviewed and corrected after generation.
- English, French and Spanish configurations: the corresponding subsets of Aya Red-teaming were used directly. The harmful prompts are kept as they were written by the annotators, and the expected safe responses were generated with
Qwen3-VL-235B-A22B-Instruct. - Italian and German configurations: no source data was available for these languages, so the English prompts were used as seeds. The same model was prompted first to produce a culturally appropriate question in the target language and then to generate the corresponding safe response.
- Safe responses: all answers were drafted by the model as a structured baseline and then checked by human reviewers, who verified alignment with the project safety guidelines, removed unsafe formulations and improved linguistic clarity and consistency across languages.
Dataset structure
Configurations
| Config | Language | Entries | Share |
|---|---|---|---|
it |
Italian | 913 | 24.9% |
de |
German | 835 | 22.8% |
en |
English | 829 | 22.6% |
fr |
French | 754 | 20.6% |
es |
Spanish | 330 | 9.0% |
| Total | 3,661 | 100% |
The configurations are not balanced. For safety-focused resources a balanced distribution across the five target languages is planned wherever feasible, but the final distributions reflect operational constraints and the availability of data judged acceptable.
Fields
| Field | Description |
|---|---|
prompt |
Harmful or adversarial request, in the language of the configuration |
response |
Safe target response to the request, in the same language as prompt |
Usage
from datasets import load_dataset
ds = load_dataset("VillanovaAI/multi-aya_redteaming", "it", split="train")
Limitations
- The dataset intentionally contains harmful and potentially disturbing prompts; it must be handled under a content-review process and is not suitable for unsupervised redistribution or for training models without safety safeguards.
- The Italian and German prompts are model-generated localizations of English seeds rather than prompts written by native red-teamers, so they may be less representative of real misuse in those languages than the English, French and Spanish ones.
- The language configurations are unbalanced, with Spanish being much smaller than the others.
Licensing and attribution
The upstream collection is released under Apache 2.0 and this SFT version is distributed under the same license. Content and red-teaming effort are credited to Cohere Labs and the Aya Red-teaming annotators.
Citation
If you use this dataset, please cite the Aya Red-teaming paper.
@misc{aakanksha2024multilingualalignmentprismaligning,
title={The Multilingual Alignment Prism: Aligning Global and Local Preferences to Reduce Harm},
author={Aakanksha and Arash Ahmadian and Beyza Ermis and Seraphina Goldfarb-Tarrant and Julia Kreutzer and Marzieh Fadaee and Sara Hooker},
year={2024},
eprint={2406.18682},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2406.18682}
}
- Downloads last month
- 98