Exploring Two-Phase Continual Instruction Fine-tuning for Multilingual Adaptation in Large Language Models

ACL ARR 2025 February Submission1036 Authors

Abstract:

A key challenge for Large Language Models (LLMs) is improving their Multilingual instruction-following ability over time without deteriorating their ability in languages they already excel at, typically English. This paper studies a two-phase Continual Fine-tuning (CFT) setup toward improving a model's Multilingual adaptability. We study a two-phase CFT process in which an English-only end-to-end instruction fine-tuned LLM from Phase 1 is sequentially fine-tuned on a multilingual instruction dataset. We focus on the open-source MISTRAL-7B and LLAMA-3-8B models and multiple dataset pairs.
Our findings show that our two-phase CFT setup outperforms simultaneous fine-tuning on the mixture of English and Multilingual instruction datasets. Moreover, we observe that the instructions similarity between Phase 1 and Phase 2 datasets plays a crucial role. When instructions are similar, the LLM after Phase 2 fine-tuning retains (or improves) its English performance, while also improving its Multilingual ability. In contrast, for non-similar phase-wise datasets, Phase 2 LLM's English ability deteriorates.
To address this, we explore heuristic-based layer freezing and data replay techniques. We show that these methods enhance multilingual ability while preserving English ability, compared to relevant baselines.

Paper Type: Long

Research Area: Multilingualism and Cross-Lingual NLP

Research Area Keywords: Continual Learning, Multilingual Adaptation, Multilingualism

Languages Studied: arabic, german, spanish, hindi, vietnam, chinese, french, japanese, romanian, russian, thai, turkish, greek, english

Submission Number: 1036