Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains
A Greek RAG study shows generic multilingual embeddings can lose to BM25 in specialist domains, while targeted adaptation sharply improves retrieval and grounded answers. Language-specific evaluation still matters.
The authors trained a Nemotron 1B embedder on **65,773 Greek retrieval pairs**, raising nDCG@10 from **0.362 to 0.835**. A LoRA-tuned Nemotron 30B-A3B reader raised judged answer correctness from **29.4% to 66.9%**.
For specialized or lower-resource languages, benchmark a lexical baseline before choosing a dense retriever, then adapt retrieval, reranking and generation as separate stages. The released **HERA benchmark** provides a Greek RAG test bed.
The authors trained a Nemotron 1B embedder on **65,773 Greek retrieval pairs**, raising nDCG@10 from **0.362 to 0.835**. A LoRA-tuned Nemotron 30B-A3B reader raised judged answer correctness from **29.4% to 66.9%**. For specialized or lower-resource languages, benchmark a lexical baseline before choosing a dense retriever, then adapt retrieval, reranking and generation as separate stages. The released **HERA benchmark** provides a Greek RAG test bed. BM25 beat several off-the-shelf multilingual dense models on the specialist corpora, and the adapted embedder's advantage over BM25 remained domain-dependent. The reported evaluation is specific to Modern Greek and the studied legal, energy, financial and medical material.
This confirms that multilingual dense retrieval is not automatically superior to a lexical baseline in specialist, lower-resource corpora. It turns RAG adaptation into a staged selection problem across retrieval, reranking, and generation, with HERA offering a focused test bed. The domain-dependent margin over BM25 narrows the result to workload-specific evaluation rather than a general prescription for Greek or multilingual RAG.