Fine-tuned NLLB-1.3B Outperforms Few-Shot LLMs in Arabic-Russian Translation

Fine-tuned NLLB-1.3B achieves a BLEU of 16.3 and COMET of 0.738 on a 15.47M-pair corpus, significantly outperforming few-shot LLMs like Aya-Expanse 8B. The study highlights that low lexical overlap and rich morphology make fine-tuned NMT more effective than prompting for this language pair.