NVIDIA Nemotron adapted for Greek doubles RAG accuracy with 0.835 nDCG

Modern Greek has been a blind spot in enterprise AI: NVIDIA's Nemotron retrieval models don't support it, and major multilingual benchmarks skip it entirely. That's a problem for retrieval-augmented generation (RAG) across legal, energy, financial, and medical applications in Greece. A new arXiv paper from Ayoub Kirouane and Christos Petrocheilos delivers an end-to-end fix, adapting the entire Nemotron stack for Greek and releasing the adapted models plus a new benchmark called HERA.

The team's findings are striking. On specialist Greek corpora, a simple BM25 keyword baseline beat several off-the-shelf multilingual dense retrieval models — a reminder that generic embeddings often stumble on niche languages. But after fine-tuning a Nemotron 1B embedder on 65,773 Greek retrieval pairs, nDCG@10 soared from 0.362 to 0.835, dramatically outperforming the unadapted model. They then adapted a cross-encoder reranker for consistent gains and LoRA-tuned a Nemotron 30B-A3B mixture-of-experts reader for grounded generation, boosting judged answer correctness from 29.4% to 66.9% while improving faithfulness and citation quality. The language competence transfers to general-domain Greek, though the edge over BM25 stays domain-dependent. With HERA and the Sophea Nemo RAG models now public, the work gives Greek-language AI a serious infrastructure upgrade.