LLaMAX: Scaling Linguistic Horizons of LLM by Enhancing Translation Capabilities Beyond 100 Languages

TL;DR

LLaMAX enhances multilingual translation in LLMs by extensive continual pre-training, vocabulary expansion, and data augmentation, achieving high performance across 100+ languages without losing generalization.

Contribution

The paper introduces LLaMAX, a novel multilingual LLM trained with specific strategies to improve low-resource language translation capabilities.

Findings

LLaMAX outperforms existing open-source LLMs by over 10 spBLEU points.

LLaMAX matches the performance of M2M-100-12B on Flores-101.

The model supports over 100 languages with robust translation quality.

Abstract

Large Language Models (LLMs) demonstrate remarkable translation capabilities in high-resource language tasks, yet their performance in low-resource languages is hindered by insufficient multilingual data during pre-training. To address this, we conduct extensive multilingual continual pre-training on the LLaMA series models, enabling translation support across more than 100 languages. Through a comprehensive analysis of training strategies, such as vocabulary expansion and data augmentation, we develop LLaMAX. Remarkably, without sacrificing its generalization ability, LLaMAX achieves significantly higher translation performance compared to existing open-source LLMs (by more than 10 spBLEU points) and performs on-par with specialized translation model (M2M-100-12B) on the Flores-101 benchmark. Extensive experiments indicate that LLaMAX can serve as a robust multilingual foundation model. The code \footnote{\url{https://github.com/CONE-MT/LLaMAX/.}} and the models \footnote{\url{https://huggingface.co/LLaMAX/.}} are publicly available.