Safe at the Margins: A General Approach to Safety Alignment in Low-Resource English Languages -- A Singlish Case Study

TL;DR

This paper presents a scalable safety alignment framework for low-resource languages like Singlish, demonstrating that combining SFT and KTO significantly reduces toxicity while maintaining benchmark performance.

Contribution

It introduces KTO-S and systematically compares safety alignment methods, highlighting superior sample efficiency and toxicity reduction in low-resource language settings.

Findings

SFT+KTO achieves higher safety alignment efficiency.

KTO-S improves stability with KL divergence regularization.

Reduces Singlish toxicity by 99% and generalizes to other datasets.

Abstract

Ensuring the safety of Large Language Models (LLMs) in diverse linguistic settings remains challenging, particularly for low-resource languages. Existing safety alignment methods are English-centric, limiting their effectiveness. We systematically compare Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Kahneman-Tversky Optimization (KTO) for aligning SEA-Lion-v2.1-Instruct, a Llama 3-8B variant, to reduce toxicity in Singlish. Our results show that SFT+KTO achieves superior safety alignment with higher sample efficiency than DPO. Additionally, we introduce KTO-S, which enhances stability via improved KL divergence regularization. Our approach reduces Singlish toxicity by 99\%, generalizes to TOXIGEN, and maintains strong performance on standard LLM benchmarks, providing a scalable framework for safer AI deployment in multilingual contexts.