Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SweEval: Do LLMs Really Swear? A Safety Benchmark for Testing Limits for Enterprise Use

Hitesh Laxmichand Patel, Amit Agarwal, Arion Das, Bhargava Kumar, Srikant Panda, Priyaranjan Pattnayak, Taki Hasan Rafi, Tejaswini Kumar, Dong-Kyu Chae

TL;DR

SweEval addresses a critical enterprise safety problem: how LLMs handle swearing across languages and scripts. It introduces a cross-lingual profanity benchmark with two cases (multilingual and transliterated swearing), spanning eight languages and 2,725 prompts per language, to quantify safety via the Harmful Rate. Through evaluations of 13 open-source models, the study finds that safety generally improves with model size and multilingual training but remains weaker for Indic languages and transliterated profanity, with Latin-language prompts being safer on average. The work emphasizes the need for better multilingual data curation and safeguards to enable trustworthy, culturally aware enterprise AI deployment, and it provides public data and code to advance this line of research.

Abstract

Enterprise customers are increasingly adopting Large Language Models (LLMs) for critical communication tasks, such as drafting emails, crafting sales pitches, and composing casual messages. Deploying such models across different regions requires them to understand diverse cultural and linguistic contexts and generate safe and respectful responses. For enterprise applications, it is crucial to mitigate reputational risks, maintain trust, and ensure compliance by effectively identifying and handling unsafe or offensive language. To address this, we introduce SweEval, a benchmark simulating real-world scenarios with variations in tone (positive or negative) and context (formal or informal). The prompts explicitly instruct the model to include specific swear words while completing the task. This benchmark evaluates whether LLMs comply with or resist such inappropriate instructions and assesses their alignment with ethical frameworks, cultural nuances, and language comprehension capabilities. In order to advance research in building ethically aligned AI systems for enterprise use and beyond, we release the dataset and code: https://github.com/amitbcp/multilingual_profanity.

SweEval: Do LLMs Really Swear? A Safety Benchmark for Testing Limits for Enterprise Use

TL;DR

SweEval addresses a critical enterprise safety problem: how LLMs handle swearing across languages and scripts. It introduces a cross-lingual profanity benchmark with two cases (multilingual and transliterated swearing), spanning eight languages and 2,725 prompts per language, to quantify safety via the Harmful Rate. Through evaluations of 13 open-source models, the study finds that safety generally improves with model size and multilingual training but remains weaker for Indic languages and transliterated profanity, with Latin-language prompts being safer on average. The work emphasizes the need for better multilingual data curation and safeguards to enable trustworthy, culturally aware enterprise AI deployment, and it provides public data and code to advance this line of research.

Abstract

Enterprise customers are increasingly adopting Large Language Models (LLMs) for critical communication tasks, such as drafting emails, crafting sales pitches, and composing casual messages. Deploying such models across different regions requires them to understand diverse cultural and linguistic contexts and generate safe and respectful responses. For enterprise applications, it is crucial to mitigate reputational risks, maintain trust, and ensure compliance by effectively identifying and handling unsafe or offensive language. To address this, we introduce SweEval, a benchmark simulating real-world scenarios with variations in tone (positive or negative) and context (formal or informal). The prompts explicitly instruct the model to include specific swear words while completing the task. This benchmark evaluates whether LLMs comply with or resist such inappropriate instructions and assesses their alignment with ethical frameworks, cultural nuances, and language comprehension capabilities. In order to advance research in building ethically aligned AI systems for enterprise use and beyond, we release the dataset and code: https://github.com/amitbcp/multilingual_profanity.

Paper Structure

This paper contains 16 sections, 1 equation, 8 figures, 22 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Curse of Multilinguality
  4. Safety in LLMs
  5. The SweEval Benchmark
  6. Case 1: Multilingual Swearing
  7. Case 2: Transliterated Swearing
  8. Experiments
  9. Experimental Setup
  10. Main Results
  11. In-depth Analysis
  12. Conclusion
  13. Appendix
  14. Detailed Evaluation Results
  15. More on SweEval Construction
  16. ...and 1 more sections

Figures (8)

  • Figure 1: Regions where our chosen languages are spoken by the majority.
  • Figure 2: Case 1 - Multilingual Swearing.
  • Figure 3: Case 2 - Transliterated Swearing.
  • Figure 4: Case 1 - Model-wise harmful rate distribution across all languages (lower is better).
  • Figure 5: Case 2 - Model-wise harmful rate distribution across all languages (lower is better).
  • ...and 3 more figures