Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains

Analysis

Why This Paper Matters

Modern Greek is a low-resource language for retrieval-augmented generation (RAG), despite its importance in legal, energy, financial, and medical sectors. NVIDIA's Nemotron retrieval models and major multilingual benchmarks largely ignore Greek, leaving practitioners with suboptimal off-the-shelf solutions. This paper addresses that gap by presenting a complete adaptation pipeline for the Nemotron stack, from corpus mining to reader fine-tuning, and introduces HERA, the first large-scale Greek RAG benchmark. The work is significant because it demonstrates that language-specific adaptation can yield dramatic improvements over generic multilingual models, and it provides a template for adapting RAG systems to other under-resourced languages.

The finding that a simple BM25 baseline beats several off-the-shelf dense retrievers on specialist Greek corpora is a stark reminder that multilingual models often underperform on domain-specific, non-English text. This motivates the need for language- and domain-tuned retrieval models, which the authors deliver. By releasing their adapted models and benchmark, they enable reproducible research and practical deployment for Greek RAG.

Technical Contributions

The paper's key technical contributions include:

• Corpus mining and synthetic supervision: They construct a Greek corpus and generate 65,773 retrieval pairs for fine-tuning, addressing the lack of labeled data.
• Retrieval model adaptation: Fine-tuning a Nemotron 1B embedder on Greek pairs yields a massive nDCG@10 improvement from 0.362 to 0.835, showing the value of language-specific training.
• Reranker adaptation: A cross-encoder reranker is adapted to Greek, providing consistent gains across specialist domains.
• Reader fine-tuning with LoRA: They LoRA-tune a Nemotron 30B-A3B mixture-of-experts reader for grounded generation, improving answer correctness and citation quality.
• HERA benchmark: A new benchmark for Greek RAG, filling a void in evaluation resources.

Results

The adapted embedder achieves an nDCG@10 of 0.835, a 2.3x improvement over the unadapted model's 0.362. The reranker further improves results across domains. For generation, the LoRA-tuned reader increases judged answer correctness from 29.4% to 66.9%, with significant gains in faithfulness and citation quality. However, the advantage over BM25 is domain-dependent, indicating that dense retrieval is not universally superior in low-resource settings.

Significance

This paper demonstrates that adapting retrieval and generation models to a specific language can yield substantial performance gains, even when starting from a strong multilingual base. The release of HERA and the adapted models will accelerate research on Greek-language RAG and provide a blueprint for similar efforts in other low-resource languages. The work also highlights the continued relevance of lexical baselines like BM25, especially in specialized domains where training data is limited. Overall, this contributes to making RAG more accessible and effective for non-English languages, with potential real-world impact in legal, medical, and financial applications.