Multi-Safety-Dataset
Content warning: this dataset intentionally contains prompts that aim to elicit harmful, illegal or otherwise unsafe behaviour. It is released for safety alignment and robustness work only.
Multi-Safety-Dataset is a multilingual collection of safety-related prompt–answer pairs released in English, Italian, French, German and Spanish. It contains 1,250 samples in total, evenly distributed across the five languages (250 samples per language).
The dataset is a supervised fine-tuning (SFT) resource for safety alignment: each sample pairs a potentially harmful request, phrased the way a real user would formulate it, with a curated response that declines the request or redirects it towards acceptable behaviour. It was built to address the underrepresentation of safety-related instructions in the project's multilingual instruction-tuning mixture, where the available safety resources were either English-only or translated rather than natively written.
How the dataset was built
The dataset was produced through a mixed human–model pipeline that combines scalability with human oversight.
- Prompt collection: an internal survey was conducted among project collaborators, who were asked to provide lists of potentially harmful prompts in their respective native languages. This gave a natively phrased prompt pool for each of the five languages and ensured that the prompts are:
- Natively formulated in each target language, rather than translated from English,
- Contextually relevant, i.e. grounded in the cultural and linguistic context of the target language,
- Reflective of realistic misuse scenarios, as reported by speakers of the language.
- Draft generation: the initial answers were generated with
Qwen3-VL-235B-A22B-Instruct, prompted to produce structured draft responses aligned with safety best practices. Model-generated drafts were used as a structured baseline instead of authoring all answers from scratch, which would have significantly increased development time. - Manual review and correction: a comprehensive human review phase was applied to every draft answer in order to:
- verify factual correctness,
- ensure alignment with the project safety guidelines,
- remove potentially unsafe formulations,
- improve linguistic clarity and consistency across languages.
The released version is the result of the completed review phase: every sample was checked by a native or expert speaker, and the final statistics match the planned 250 samples per configuration.
Dataset structure
Configurations
| Config | Language | Entries | Share |
|---|---|---|---|
en |
English | 250 | 20% |
it |
Italian | 250 | 20% |
fr |
French | 250 | 20% |
de |
German | 250 | 20% |
es |
Spanish | 250 | 20% |
| Total | 1,250 | 100% |
Fields
| Field | Description |
|---|---|
question |
Potentially harmful request, written natively in the language of the configuration |
answer |
Curated response to the request, in the same language as question |
Usage
from datasets import load_dataset
ds = load_dataset("VillanovaAI/Multi-Safety-Dataset", "en", 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.
- Coverage is limited to the five target languages.
Licensing and attribution
The dataset is distributed under ODC-BY. The prompts were authored by the project collaborators and no third-party data is redistributed; the answers were generated with Qwen3-VL-235B-A22B-Instruct, whose license was verified as compatible with downstream commercial use, and then manually reviewed and corrected.
- Downloads last month
- 2