LlamaTurk: Adapting Open-Source Generative Large Language Models for Low-Resource Language
Cagri Toraman
TL;DR
The paper studies adapting English-dominant open-source LLMs to a low-resource language (Turkish) by comparing continual training, instruction fine-tuning, task-specific fine-tuning, and vocabulary extension using Llama-based and MaLA models. It shows continual training improves perplexity, instruction tuning improves perplexity and downstream tasks under certain contexts, while vocabulary extension generally harms performance; task-specific tuning enhances downstream task performance, particularly in sentiment analysis, and larger models help under few-shot settings. Multilingual adaptation provides limited benefits compared with monolingual adaptation. The work offers practical guidance for selecting adaptation strategies under resource constraints and provides open resources for reproducing results.
Abstract
Despite advancements in English-dominant generative large language models, further development is needed for low-resource languages to enhance global accessibility. The primary methods for representing these languages are monolingual and multilingual pretraining. Monolingual pretraining is expensive due to hardware requirements, and multilingual models often have uneven performance across languages. This study explores an alternative solution by adapting large language models, primarily trained on English, to low-resource languages. We assess various strategies, including continual training, instruction fine-tuning, task-specific fine-tuning, and vocabulary extension. The results show that continual training improves language comprehension, as reflected in perplexity scores, and task-specific tuning generally enhances performance of downstream tasks. However, extending the vocabulary shows no substantial benefits. Additionally, while larger models improve task performance with few-shot tuning, multilingual models perform worse than their monolingual counterparts when adapted.