GemmAr: Enhancing LLMs Through Arabic Instruction-Tuning

Summary

The paper introduces InstAr-500k, a new Arabic instruction dataset of 500,000 examples designed to improve LLM performance in Arabic. Researchers fine-tuned the open-source Gemma-7B model using InstAr-500k and evaluated it on downstream tasks, achieving strong results on Arabic NLP benchmarks. They then released GemmAr-7B-V1, a model specifically tuned for Arabic NLP tasks. Why it matters: This work addresses the lack of high-quality Arabic instruction data, potentially boosting the capabilities of Arabic language models.