Adapting Chat Language Models Using Only Target Unlabeled Language Data

TL;DR

ElChat introduces a base-free approach to adapting chat LLMs to target languages using only unlabeled data. It combines three components—vocabulary expansion on the source chat model, model merging with the source chat model, and copying of special-token weights—to elicit chat and instruction-following abilities while preserving target-language performance. Through extensive multilingual experiments, ElChat outperforms the prior CV method on chat and instruction-following tasks and maintains competitive safety and target-language performance, with comparable inference speedups to VE-based baselines. The work demonstrates a practical path for adapting chat LLMs for low-resource languages when target chat data or aligned base/chat pairs are unavailable, and it frames a roadmap for future improvements in merging strategies and evaluation benchmarks.

Abstract

Vocabulary expansion (VE) is the de-facto approach to language adaptation of large language models (LLMs) by adding new tokens and continuing pre-training on target data. While this is effective for base models trained on unlabeled data, it poses challenges for chat models trained to follow instructions through labeled conversation data. Directly adapting the latter with VE on target unlabeled data may result in forgetting chat abilities. While ideal, target chat data is often unavailable or costly to create for low-resource languages, and machine-translated alternatives are not always effective. To address this issue, previous work proposed using a base and chat model from the same family. This method first adapts the base LLM with VE on target unlabeled data and then converts it to a chat model by adding a chat vector (CV) derived from the weight difference between the source base and chat models. We propose ElChat, a new language adaptation method for chat LLMs that adapts a chat model directly on target unlabeled data, without a base model. It elicits chat abilities by injecting information from the source chat model. ElChat offers more robust and competitive target language and safety performance while achieving superior English, chat, and instruction-following abilities compared to CV.

Figures (10)

• Figure 1: Chat LLM language adaptation with Chat Vector huang-etal-2024-chat and ElChat (ours). Note that a base LM in this paper refers to an LM pre-trained on unlabeled data without any further post-training. A chat LM, on the other hand, is a base model that has been further supervised fine-tuned on labeled conversational data, enabling it to follow instructions.
• Figure 2: ElChat: A three-step adaptation process overview.
• Figure 3: Aggregated mean performance ($\uparrow$) across seven target languages for each model on safety, chat, and instruction-following tasks. Full results are in the Appendix (Tables \ref{['tab:chat_performance_qwen25']} and \ref{['tab:chat_performance_llama31']}).
• Figure 4: Aggregated mean performance across seven target languages for each model (error bars indicate 95% confidence interval). Full results are in Tables \ref{['tab:task_performance_qwen25']} and \ref{['tab:task_performance_llama31']}.
• Figure 5: Aggregated mean inference speedup (tokens/s) across seven target languages. The value next to each bar represents the speedup ratio over Chat. Full results are in the Appendix (Tables \ref{['tab:speed_performance_qwen25']} and \ref{['tab:speed_performance_llama31']}).