Hasan Abed Al Kader Hammoud Mohammad Zbeeb Bernard Ghanem

Abstract

We present Hala, a family of Arabic-centric instruction and translationmodels built with our translate-and-tune pipeline. We first compress a strongARleftrightarrowEN teacher to FP8 (yielding sim2times higherthroughput with no quality loss) and use it to create high-fidelity bilingualsupervision. A lightweight language model LFM2-1.2B is then fine-tuned on thisdata and used to translate high-quality English instruction sets into Arabic,producing a million-scale corpus tailored to instruction following. We trainHala models at 350M, 700M, 1.2B, and 9B parameters, and apply slerp merging tobalance Arabic specialization with base-model strengths. On Arabic-centricbenchmarks, Hala achieves state-of-the-art results within both the "nano"(leq2B) and "small" (7-9B) categories, outperforming their bases. We releasemodels, data, evaluation, and recipes to accelerate research in Arabic NLP.