Closing the Language Gap: Adapting NVIDIA's Nemotron for Modern Greek RAG

What Changed

Modern Greek has historically been a blind spot for high-performance retrieval-augmented generation (RAG) systems. While major multilingual models often handle widely spoken languages with ease, specialized domains—such as legal, energy, financial, and medical sectors—have lacked the necessary linguistic grounding in Greek to support robust RAG pipelines. NVIDIA's Nemotron retrieval models, despite their efficiency, previously lacked support for this language, leaving developers to rely on suboptimal general-purpose multilingual embeddings or sparse retrieval methods like BM25.

A research team has now bridged this gap through an end-to-end adaptation of the Nemotron retrieval stack. This project involved a comprehensive pipeline: corpus mining, synthetic supervision, retrieval model training, reranker adaptation, and reader fine-tuning. The result is a specialized suite of models capable of handling Modern Greek with high precision, alongside the introduction of HERA, the first large-scale benchmark specifically designed to evaluate Greek-language RAG systems.

Technical Details

The adaptation process was methodical, focusing on both the retrieval and generation components of the RAG stack. The researchers began by addressing the lack of training data, utilizing a corpus mining strategy to generate 65,773 Greek retrieval pairs. This synthetic supervision was critical for fine-tuning the Nemotron 1B embedder.

Before this intervention, the team observed that a parameter-free BM25 baseline actually outperformed several off-the-shelf multilingual dense retrieval models on specialist Greek corpora. This highlights a common pitfall in multilingual RAG: dense models often struggle with domain-specific terminology if they haven't been explicitly trained on that language's nuances.

Following the fine-tuning of the 1B embedder, the researchers shifted their attention to the reranking stage. They adapted a cross-encoder reranker, which demonstrated consistent improvements across the targeted specialist domains. This multi-stage approach—combining a dense retriever with a reranker—is standard practice in high-performance RAG, but the adaptation ensures that the semantic matching remains accurate even when dealing with the complex syntax and vocabulary of Modern Greek.

For the generation phase, the team employed LoRA (Low-Rank Adaptation) to fine-tune a Nemotron 30B-A3B mixture-of-experts (MoE) reader. This step was pivotal for grounding the model's output. By focusing on faithfulness and citation quality, the researchers transformed the model from a general-purpose generator into a reliable system capable of providing accurate, source-backed answers in Greek.

Benchmark Analysis

The performance gains achieved through these adaptations are substantial, particularly when comparing the base model to the fine-tuned version. The HERA benchmark provides the first standardized metric for this specific task, revealing clear improvements in retrieval accuracy and answer correctness.

As shown in the table, the nDCG@10 score—a standard metric for evaluating the quality of search results—more than doubled after the adaptation process. Furthermore, the answer correctness for the RAG system saw a significant jump, moving from under 30% to nearly 67%. These figures underscore the necessity of domain-specific and language-specific fine-tuning for high-stakes applications.

Developer Implications

For developers working on RAG systems in languages other than English, this research offers a blueprint for adaptation. The reliance on BM25 as a baseline is a crucial takeaway; it serves as a reality check for developers who might assume that any dense retrieval model is superior to traditional keyword-based search.

1. Data Quality Over Quantity: The creation of 65,773 high-quality retrieval pairs was the catalyst for the performance leap. Developers should prioritize the curation of domain-specific datasets rather than relying solely on massive, general-purpose multilingual datasets.
2. The Reranker Advantage: The consistent gains seen with the adapted cross-encoder reranker suggest that even if a dense retriever is "good enough," a dedicated reranker is essential for production-grade RAG, especially in complex domains like law or medicine.
3. MoE Efficiency: Utilizing a mixture-of-experts architecture for the reader allows for high-performance generation without the prohibitive costs of a dense 30B+ parameter model, making it a viable path for enterprise deployment.

Bottom Line

The adaptation of the Nemotron stack for Modern Greek demonstrates that language-specific fine-tuning is not just an optimization, but a requirement for reliable RAG performance. By releasing both the adapted models and the HERA benchmark, the researchers have provided the community with the tools to replicate this success in other under-represented languages, setting a new standard for multilingual RAG development.

Pneumetron