pretty_name: Multilingual Safety Instructions
language:
- en
- zh
- it
- bn
- ko
- th
- vi
- ar
- jv
- sw
- de
- fr
license: cc-by-nc-4.0
task_categories: - text-generation
tags: - safety
- alignment
- refusal
- multilingual
configs: - config_name: en
data_files: en.csv - config_name: zh
data_files: zh.csv - config_name: it
data_files: it.csv - config_name: bn
data_files: bn.csv - config_name: ko
data_files: ko.csv - config_name: th
data_files: th.csv - config_name: vi
data_files: vi.csv - config_name: ar
data_files: ar.csv - config_name: jv
data_files: jv.csv - config_name: sw
data_files: sw.csv - config_name: de
data_files: de.csv - config_name: fr
data_files: fr.csv
Multilingual Safety Instructions
A multilingual extension of the safety-only instruction–refusal pairs released with the Safety-Tuned LLaMAs project. The original 1,000 harmful-prompt / refusal-response pairs (English) were translated into 11 additional typologically diverse languages with google/gemini-2.0-flash-001. Each language is stored as a separate Hugging Face config.
| Field | Description |
|---|---|
prompt |
Harmful user instruction (translated; en is the original). |
output |
Safe refusal / corrective response (translated; en is the original). |
Languages (configs): en (original), zh, it, bn, ko, th, vi, ar, jv, sw, de, fr.
Intended use
This dataset is intended for alignment research: probing refusal behaviour in language models, building multilingual safety calibration corpora (e.g. detecting refusal-specific neurons), and evaluating cross-lingual safety transfer. The prompt field contains content that may be offensive, harmful, or unsafe — it is published deliberately so that models can be trained or analysed to refuse such queries.
Usage
from datasets import load_dataset
ds_en = load_dataset("iNLP-Lab/multilingual-safety", "en")
ds_zh = load_dataset("iNLP-Lab/multilingual-safety", "zh")
Source and credits
This release re-distributes machine-translated derivatives of the safety_only_data_Instructions.json split from the Safety-Tuned LLaMAs project:
Federico Bianchi, Mirac Suzgun, Giuseppe Attanasio, Paul Röttger, Dan Jurafsky, Tatsunori Hashimoto, James Zou. "Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions." ICLR 2024. arXiv:2309.07875
The original English data is hosted at vinid/safety-tuned-llamas (file data/training/safety_only_data_Instructions.json). All credit for the underlying instruction–refusal pairs belongs to the authors above; this release only contributes the multilingual translations.
@inproceedings{bianchi2024safetytuned,
title = {Safety-Tuned {LL}a{MA}s: Lessons From Improving the Safety of Large Language Models that Follow Instructions},
author = {Bianchi, Federico and Suzgun, Mirac and Attanasio, Giuseppe and R{\"o}ttger, Paul and Jurafsky, Dan and Hashimoto, Tatsunori and Zou, James},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2024}
}
License
Released under CC BY-NC 4.0, matching the non-commercial research intent of the upstream Safety-Tuned LLaMAs release.