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

2025-05-22

AI Summary

To address cross-lingual toxic transfer—the challenge wherein multilingual large language models (MLLMs) propagate toxicity across languages during global deployment—this paper pioneers and empirically validates the “cross-lingual detoxification” paradigm, enabling generalization of toxicity mitigation across high- and low-resource languages and diverse writing systems. Methodologically, we construct a multilingual toxicity-annotated dataset, design script-agnostic supervised fine-tuning strategies, and establish a cross-distribution evaluation framework covering 504 language–script–resource combinations. Key contributions include: (1) a formal definition and empirical validation of cross-lingual detoxification efficacy; (2) identification of the inherent trade-off between safety enhancement and knowledge retention; and (3) significant reduction in toxicity outputs across multilingual benchmarks while preserving stable performance on non-toxic downstream tasks. All code and datasets are publicly released to ensure reproducibility.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

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 504 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.

Problem

• Mitigating toxicity in multilingual language models
• Transferring detoxification between high and low-resource languages
• Balancing safety and knowledge preservation in detoxification

Innovation

• Supervised fine-tuning for cross-lingual detoxification
• Toxicity mitigation across diverse linguistic contexts
• Evaluating trade-offs between safety and knowledge preservation