Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains
A research paper presents an end-to-end adaptation of NVIDIA's Nemotron retrieval stack for Modern Greek, including corpus mining, synthetic supervision, retrieval model training, reranker adaptation, reader fine-tuning, and a new benchmark called HERA. A parameter-free BM25 baseline outperforms several off-the-shelf multilingual dense retrieval models on specialist Greek corpora. After fine-tuning on 65,773 Greek retrieval pairs, a Nemotron 1B embedder improves nDCG@10 from 0.362 to 0.835. The learned language competence transfers to general-domain Greek, though the advantage over BM25 remains domain-dependent. A cross-encoder reranker is adapted and demonstrates consistent improvements across specialist domains. A Nemotron 30B-A3B mixture-of-experts reader is LoRA-tuned for grounded generation, increasing judged answer correctness from 29.4% to 66.9%.
Modern Greek is absent from NVIDIA's Nemotron retrieval models and major multilingual retrieval benchmarks. This study adapts the Nemotron retrieval stack for Modern Greek across legal, energy, financial, and medical domains. It includes corpus mining, synthetic supervision, retrieval model training, reranker adaptation, reader fine-tuning, and a new benchmark called HERA. A BM25 baseline outperforms off-the-shelf multilingual dense retrieval models on specialist Greek corpora. Fine-tuning a Nemotron 1B embedder on 65,773 Greek retrieval pairs raises nDCG@10 from 0.362 to 0.835. The competence transfers to general-domain Greek, but the advantage over BM25 is domain-dependent. A cross-encoder reranker is adapted and shows consistent improvements. LoRA-tuning a Nemotron 30B-A3B mixture-of-experts reader for grounded generation increases judged answer correctness from 29.4% to 66.9%.
The adaptation demonstrates that fine-tuning a small embedder (1B parameters) on a modest number of domain-specific retrieval pairs (65,773) can dramatically improve retrieval quality for a low-resource language, surpassing off-the-shelf multilingual models. The use of a parameter-free BM25 baseline highlights that dense retrieval models may not always be superior for specialist domains without adaptation. The combination of a fine-tuned embedder, adapted cross-encoder reranker, and LoRA-tuned reader shows a complete RAG pipeline can be effectively localized.
This work indicates that NVIDIA's Nemotron models can be extended to underserved languages with relatively low data requirements, potentially opening markets in regions where Modern Greek is used in legal, energy, financial, and medical sectors. The approach may be replicable for other low-resource languages, suggesting a pathway for enterprise RAG solutions in non-English domains.
Enables accurate retrieval-augmented generation for Modern Greek in specialist domains, reducing the need for large-scale data collection and training from scratch. This can lower barriers for companies and institutions to deploy AI-powered search and question-answering systems in Greek, improving efficiency in legal, energy, financial, and medical applications.
Next signals to watch include the release of the HERA benchmark and the adapted models, potential expansion to other low-resource languages, and adoption by Greek-speaking enterprises or government services. Further research may explore domain-specific fine-tuning strategies and the trade-offs between dense and sparse retrieval methods in multilingual settings.