Multi-AdvBench

Overview

Multi-AdvBench is a multilingual adversarial benchmark dataset derived from the original multilingual_advbench dataset from simonycl.

The original multilingual_advbench dataset consists of prompt–target pairs, where:

• the prompt is a harmful request,
• the target is an unintended (unsafe) response.

The new dataset preserves the original prompts and re-generates the targets using a permissively licensed multimodal model: Qwen3-VL-235B-A22B-Instruct-FP8

As a result, Multi-AdvBench provides a multilingual dataset using the original prompts while ensuring that all generated responses are fully safe. The new dataset therefore consists of prompt–target pairs, where:

• the prompt is a harmful request,
• the target is an expected safe response.

---

Sets processing, target generation and language distribution

The original dataset contains 520 prompt–target pairs. The same 520 pairs are available in each language and in the same order.

Our process was as follows:

• we selected five languages,

• we discarded the targets from each prompt–target pair,

• we created a global list of indices from 0 to 519 and randomly shuffled it,

• we split this list into five disjoint blocks of 104 indices, one block per language,

• for each language, we selected only the prompts at the indices assigned to that language. As a result:

  • each language contains exactly 104 prompt–target pairs
  • no prompt is shared between languages
  • all 520 prompts are used exactly once

• for each prompt we generated a new safe response,

• every prompt-target pair has been manually reviewed.

Language	Percentage	Number of Examples
English	20%	104
Italian	20%	104
French	20%	104
Spanish	20%	104
German	20%	104

---

Intended Use

This dataset provides multilingual adversarial benchmark prompt–target pairs suitable for:

• training LLMs for guardrailing purposes,
• benchmarking model performance across languages,
• developing new models requiring permissively licensed training data.

Because responses were generated using a permissively licensed model, all data can be freely reused for research and commercial model development.

---

Dataset Structure

We maintained the choice made by the authors of the original dataset to name the fields "prompt" and "target". Each example in every language contains the following fields:

Field	Type	Description
lang	string	the language of the prompt–target pair
prompt	string	a harmful prompt
target	string	a safe response to the prompt

---

Licensing

The dataset is distributed under the ODC-BY-1.0 license. Users must comply with all attribution and usage requirements specified by the license.