README.md

metadata

dataset_info:
  features:
    - name: xml
      dtype: string
    - name: proceedings
      dtype: string
    - name: year
      dtype: string
    - name: url
      dtype: string
    - name: language documentation
      dtype: string
    - name: has non-English?
      dtype: string
    - name: topics
      dtype: string
    - name: language coverage
      dtype: string
    - name: title
      dtype: string
    - name: abstract
      dtype: string
  splits:
    - name: train
      num_bytes: 452838
      num_examples: 310
  download_size: 231933
  dataset_size: 452838
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: cc-by-4.0
task_categories:
  - text-classification

The State of Multilingual LLM Safety Research: From Measuring the Language Gap to Mitigating It

We present a comprehensive analysis of the linguistic diversity of LLM safety research, highlighting the English-centric nature of the field. Through a systematic review of nearly 300 publications from 2020–2024 across major NLP conferences and workshops at *ACL, we identify a significant and growing language gap in LLM safety research, with even high-resource non-English languages receiving minimal attention.

• Paper: https://arxiv.org/abs/2505.24119

Dataset Description

Current version of the dataset consists of annotations for conference and workshop papers collected from *ACL venues between 2020 and 2024, using keywords of "safe" and "safety" in abstracts to identify relevant literature. The data source is https://github.com/acl-org/acl-anthology/tree/master/data, and the paperes are curated by Zheng-Xin Yong, Beyza Ermis, Marzieh Fadaee, and Julia Kreutzer.

Dataset Structure

• xml: xml string from ACL Anthology
• proceedings: proceedings of the conference or workshop the work is published in.
• year: year of publication
• url: paper url on ACL Anthology
• language documentation: whether the paper explicitly reports the languages studied in the work. ("x" indicates failure of reporting)
• has non-English?: whether the work contains non-English language. (0: English-only, 1: has at least one non-English language)
• topics: topic of the safety work ('jailbreaking attacks'; 'toxicity, bias'; 'hallucination, factuality'; 'privacy'; 'policy'; 'general safety, LLM alignment'; 'others')
• language coverage: languages covered in the work (null means English only)
• title: title of the paper
• abstract: abstract of the paper

Citation

@article{yong2025safetysurvey,
  title={The State of Multilingual LLM Safety Research: From Measuring the Language Gap to Mitigating It}, 
  author={Zheng-Xin Yong and Beyza Ermis and Marzieh Fadaee and Stephen H. Bach and Julia Kreutzer},
  year={2025},
  journal={arXiv preprint arXiv:2505.24119},
}

Dataset Card Authors

• Zheng-Xin Yong
• Beyza Ermis
• Marzieh Fadaee
• Stephen H. Bach
• Julia Kreutzer