AfriGuard: Safety and Robust Refusal For African Languages

Key points

• Large language models (LLMs) designed to interact with users are typically trained to respond appropriately to harmful or toxic content.1
• However, ensuring appropriate responses becomes more challenging in diverse linguistic and cultural contexts, where safety risks may be expressed or interpreted differently.1
• We further find that models often fail to respond appropriately to unsafe prompts not because they explicitly refuse, but because they produce linguistically inadequate responses that result in incidental non-assistance rather than proper refusals.1
• Finally, using AfriGuard, we develop efficient guardrail models that achieve performance competitive with models up to 30x larger, introduce refusal detectors that improve average performance by 43.9% over existing detectors, and demonstrate that supervised fine-tuning can substantially improve overall safety performance.1

Sources (1)

• [1]AfriGuard: Safety and Robust Refusal For African Languages
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 02:21 PM

  In this work, we introduce AfriGuard, a culturally grounded multilingual dataset for studying harmful intent, safety-category understanding, and refusal behavior across 10 African languages and English.

  Large language models (LLMs) designed to interact with users are typically trained to respond appropriately to harmful or toxic content.

Extractive summary: sentences quoted from the sources.