UCCIX: Irish-eXcellence Large Language Model

TL;DR

The paper addresses the challenge of building large language models for extremely low-resource languages, using Irish as a case study. It introduces UCCIX, an Irish-based LLM derived from Llama 2-13B, and a training pipeline that combines continued pre-training, parallel English-Irish data, and vocabulary expansion to include 10k Irish tokens, followed by supervised instruction fine-tuning. It contributes Irish benchmarking datasets (IrishQA and an Irish version of MT-bench) and demonstrates that UCCIX can outperform much larger models on Irish tasks by up to 12%, with strong instruction-following capabilities and improved tokenizer efficiency. The results suggest a practical blueprint for adapting LLMs to other low-resource or indigenous languages, highlighting the importance of data curation, cross-language transfer, and targeted evaluation.

Abstract

The development of Large Language Models (LLMs) has predominantly focused on high-resource languages, leaving extremely low-resource languages like Irish with limited representation. This work presents UCCIX, a pioneering effort on the development of an open-source Irish-based LLM. We propose a novel framework for continued pre-training of LLMs specifically adapted for extremely low-resource languages, requiring only a fraction of the textual data typically needed for training LLMs according to scaling laws. Our model, based on Llama 2-13B, outperforms much larger models on Irish language tasks with up to 12% performance improvement, showcasing the effectiveness and efficiency of our approach. We also contribute comprehensive Irish benchmarking datasets, including IrishQA, a question-answering dataset, and Irish version of MT-bench. These datasets enable rigorous evaluation and facilitate future research in Irish LLM systems. Our work aims to preserve and promote the Irish language, knowledge, and culture of Ireland in the digital era while providing a framework for adapting LLMs to other indigenous languages.

Figures (1)

• Figure 1: Example of responses generated by our model, UCCIX, and other baselines. We demonstrate the English translation version on the right.