Breaking mBad! Supervised Fine-tuning for Cross-Lingual Detoxification

TL;DR

This paper introduces a cross-lingual detoxification method for large language models, enabling toxicity mitigation across diverse languages and scripts, while analyzing its impact on model performance and safety trade-offs.

Contribution

It proposes a novel cross-lingual detoxification approach and evaluates its effectiveness across numerous settings, addressing toxicity in multilingual LLMs.

Findings

Effective toxicity reduction in multiple languages

Trade-offs between safety and knowledge retention

Robust performance across diverse linguistic settings

Abstract

As large language models (LLMs) become increasingly prevalent in global applications, ensuring that they are toxicity-free across diverse linguistic contexts remains a critical challenge. We explore "Cross-lingual Detoxification", a cross-lingual paradigm that mitigates toxicity, enabling detoxification capabilities to transfer between high and low-resource languages across different script families. We analyze cross-lingual detoxification's effectiveness through 392 extensive settings to evaluate toxicity reduction in cross-distribution settings with limited data and investigate how mitigation impacts model performance on non-toxic tasks, revealing trade-offs between safety and knowledge preservation. Our code and dataset are publicly available at https://github.com/himanshubeniwal/Breaking-mBad.