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

This paper adapts NVIDIA’s Nemotron retrieval and generation stack to Modern Greek for specialist domains including law, energy, finance, and medicine. The pipeline covers corpus mining, synthetic supervision, embedding-model training, reranker adaptation, and reader fine-tuning. With 65,773 Greek retrieval pairs, a Nemotron 1B embedder raises nDCG@10 from 0.362 to 0.835, outperforming its untuned version. A LoRA-tuned Nemotron 30B-A3B mixture-of-experts reader increases judged answer correctness from 29.4% to 66.9%, while improving faithfulness and citation quality. The authors also introduce the HERA RAG benchmark and release adapted models and data.