Modern Greek has been largely absent from NVIDIA's Nemotron retrieval models and major multilingual benchmarks, creating a gap for retrieval-augmented generation (RAG) in legal, energy, financial, and medical applications. A new study from Hugging Face presents an end-to-end adaptation of the Nemotron retrieval stack specifically for Modern Greek, addressing this gap with corpus mining, synthetic supervision, and model fine-tuning.
The research introduces HERA, the first large-scale Greek benchmark for RAG, covering specialist domains. Surprisingly, a parameter-free BM25 baseline outperformed several off-the-shelf multilingual dense retrieval models on specialist Greek corpora. However, after fine-tuning on 65,773 Greek retrieval pairs, a Nemotron 1B embedder improved nDCG@10 from 0.362 to 0.835, substantially outperforming its unadapted counterpart.
The team also adapted a cross-encoder reranker, achieving consistent improvements across domains. For generation, they LoRA-tuned a Nemotron 30B-A3B mixture-of-experts reader, increasing judged answer correctness from 29.4% to 66.9% while improving faithfulness and citation quality. The learned language competence transferred to general-domain Greek, though the advantage over BM25 remained domain-dependent.
All adapted models and the HERA benchmark are released to support future research on Greek-language RAG systems. The authors welcome feedback, especially from researchers working on other low-resource languages.