walledai/CSRT

dataset_info:
features:

• name: prompt
  dtype: string
• name: original
  dtype: string
• name: harms
  sequence: string
  splits:
• name: train
  num_bytes: 65394
  num_examples: 315
  download_size: 45231
  dataset_size: 65394
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
    task_categories:
• text-generation
  tags:
• codeswitch
• multilingual
  size_categories:
• n<1K

Dataset derived from CSRT: Evaluation and Analysis of LLMs using Code-Switching Red-Teaming Dataset submitted to the NeurIPS 2024 Datasets and Benchmarks track

We introduce code-switching red-teaming, a simple yet effective red-teaming technique that simultaneously tests the multilingual capabilities and safety of LLMs

Keywords

LLM, Evaluation, Safety, Multilingual, Red-teaming, Code-switching

Abstract

Recent studies in large language models (LLMs) shed light on their multilingual ability and safety, beyond conventional tasks in language modeling. Still, current benchmarks reveal their inability to comprehensively evaluate them and are excessively dependent on manual annotations. In this paper, we introduce code-switching red-teaming (CSRT), a simple yet effective red-teaming technique that simultaneously tests multilingual understanding and safety of LLMs. We release the CSRT dataset, which comprises 315 code-switching queries combining up to 10 languages and eliciting a wide range of undesirable behaviors. Through extensive experiments with ten state-of-the-art LLMs, we demonstrate that CSRT significantly outperforms existing multilingual red-teaming techniques, achieving 46.7% more attacks than existing methods in English. We analyze the harmful responses toward the CSRT dataset concerning various aspects under ablation studies with 16K samples, including but not limited to scaling laws, unsafe behavior categories, and input conditions for optimal data generation. Additionally, we validate the extensibility of CSRT, by generating code-switching attack prompts with monolingual data.

Citation

If you find the dataset useful, please cite the following work:

@article{yoo2024csrt,
  title={CSRT: Evaluation and Analysis of LLMs using Code-Switching Red-Teaming Dataset},
  author={Yoo, Haneul and Yang, Yongjin and Lee, Hwaran},
  journal={arXiv preprint arXiv:2406.15481},
  year={2024}
}