Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains
A Greek-tuned Nemotron retrieval stack beat its unadapted baseline sharply on specialist RAG tasks.
The paper reports that BM25 outperformed several off-the-shelf multilingual dense retrievers on specialist Greek corpora. After fine-tuning on 65,773 Greek retrieval pairs, the Nemotron 1B embedder raised nDCG@10 from 0.362 to 0.835. A tuned reranker improved results across the specialist domains tested. LoRA-tuning a Nemotron 30B-A3B reader lifted judged answer correctness from 29.4% to 66.9%, with better faithfulness and citations. The authors also release HERA, a large-scale Greek RAG benchmark, along with adapted Sophea Nemo RAG models. HF Daily Papers' note
The paper reports that BM25 outperformed several off-the-shelf multilingual dense retrievers on specialist Greek corpora. After fine-tuning on 65,773 Greek retrieval pairs, the Nemotron 1B embedder raised nDCG@10 from 0.362 to 0.835. A tuned reranker improved results across the specialist domains tested. LoRA-tuning a Nemotron 30B-A3B reader lifted judged answer correctness from 29.4% to 66.9%, with better faithfulness and citations. The authors also release HERA, a large-scale Greek RAG benchmark, along with adapted Sophea Nemo RAG models. HF Daily Papers' note
score 4