NVIDIA researchers published a full adaptation playbook for running the Nemotron retrieval stack in Modern Greek, covering every layer from corpus mining to a LoRA-tuned mixture-of-experts reader. The paper, submitted August 5, 2026, targets legal, energy, financial, and medical RAG—four regulated sectors where Greek is the operational language and where no production-grade retrieval benchmark existed.

The central finding challenges a common shortcut: BM25 baseline retrieval beats several off-the-shelf multilingual dense retrievers on specialist Greek corpora before any fine-tuning. On HERA—a new specialist-domain benchmark covering legal, energy, financial, and medical Greek text—the unadapted Nemotron 1B embedder scores nDCG@10 of 0.362. After fine-tuning on 65,773 synthetically generated Greek retrieval pairs, it reaches 0.835—a 2.3× gain over BM25. That 0.362 baseline is specific to HERA's specialist-domain queries. A separate evaluation on the CUP Greek book catalog (arxiv 2607.21274) shows the same off-the-shelf Nemotron model reaching nDCG@9 of 0.582 on general academic text—a different dataset, metric, and result set not directly comparable to HERA. The learned representations transfer to general-domain Greek, though BM25 advantage remains domain-dependent outside the training distribution.

Retrieval SystemDatasetMetricScoreNotes
BM25 baselineHERAnDCG@10~0.363Derived: 0.835 ÷ 2.3×
Nemotron 1B (unadapted)HERAnDCG@100.362Off-the-shelf; beaten by BM25
Nemotron 1B (fine-tuned)HERAnDCG@100.835After 65,773 synthetic pairs; 2.3× over BM25
Nemotron 1B (off-the-shelf)CUP Greek book catalognDCG@90.582General academic text; different dataset and metric
FIG. 02 Nemotron 1B retrieval performance before and after fine-tuning on Greek specialist corpora — arxiv.org/abs/2608.05138; arxiv.org/html/2607.21274

The training pipeline is fully scripted. Corpus mining assembles domain-specific Greek text. A synthetic supervision stage generates query–passage pairs without manual annotation. The Nemotron 1B bi-encoder is fine-tuned on 65,773 pairs. A cross-encoder reranker is separately adapted and delivers consistent nDCG improvements across all specialist domains. The full stack—embedder, reranker, reader—maps directly onto NVIDIA's existing NeMo retrieval recipes, which include fine-tuning scripts and distillation paths for the 1B and 8B Nemotron Embed models.

Full Greek RAG adaptation pipeline: from corpus mining to LoRA-tuned MoE reader
FIG. 03 Full Greek RAG adaptation pipeline: from corpus mining to LoRA-tuned MoE reader — arxiv.org/abs/2608.05138

At the reader layer, the team LoRA-tunes a Nemotron 30B-A3B MoE model. Judged answer correctness rises from 29.4% to 66.9%, with simultaneous gains in faithfulness and citation quality. The 30B-A3B architecture carries 31.6 billion total parameters but activates 3.2 billion per forward pass (3.6B including embeddings), per the Nemotron 3 Nano technical report. Inference cost tracks closer to a 3B dense model than a 30B one—the figure that matters for production estimates.

AspectValue
Total parameters31.6 B
Active parameters per forward pass3.2 B
Active parameters incl. embeddings3.6 B
Inference cost comparable to~3B dense model
Answer correctness — baseline (no LoRA)29.4%
Answer correctness — after LoRA tuning66.9%
FIG. 04 Nemotron 30B-A3B MoE reader: architecture specs and LoRA tuning impact on answer correctness — arxiv.org/abs/2608.05138; arxiv.org/pdf/2512.20848 (Nemotron 3 Nano technical report)

HERA, released alongside the paper, is the first large-scale Greek RAG evaluation covering legal, energy, and financial domains. It provides teams a concrete eval target rather than translated English benchmarks, which misrepresent Greek-specific terminology, grammar, and domain vocabulary. Both adapted models and the benchmark are public.

Every component—synthetic pair generation, bi-encoder fine-tuning, cross-encoder reranker adaptation, MoE LoRA tuning—is language-agnostic. Teams building RAG for other underserved languages can replicate this pipeline. Domain-specific corpus collection is the primary bottleneck. Synthetic supervision quality determines results: the 65,773 pairs needed to move nDCG@10 from 0.362 to 0.835 required deliberate query generation, not random sampling.

If your retrieval domain is narrow and your language sits outside the top-10 MTEB languages, treat BM25 as your real baseline—not a multilingual embedding model. Budget for at least 50,000 synthetic retrieval pairs before expecting dense retrieval to deliver.