Hugging Face Daily Papers · August 7, 2026 · 4 min read

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

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

If you've tried building RAG in Greek, you've probably hit the same wall we did: most multilingual retrievers haven't seen enough Greek, and there's no benchmark to tell you how bad it is. So we adapted the Nemotron retrieval stack end to end and built HERA, a benchmark covering legal, energy, financial and medical Greek. The finding that surprised us most was plain BM25 beating several off-the-shelf dense models on specialist corpora.. fine-tuning fixed that. Models and benchmark are released, and feedback is welcome, especially from people working on other low-resource languages.</p>\n","updatedAt":"2026-08-07T08:14:26.913Z","author":{"_id":"6338c06c107c4835a05699f9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6338c06c107c4835a05699f9/uQbrMgQySY2UW7z3R9gT5.jpeg","fullname":"Ayoub Kirouane","name":"ayoubkirouane","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":68,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.952118992805481},"editors":["ayoubkirouane"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/6338c06c107c4835a05699f9/uQbrMgQySY2UW7z3R9gT5.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.05138","authors":[{"_id":"6a74940fe1228e04b3237ec8","user":{"_id":"6338c06c107c4835a05699f9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6338c06c107c4835a05699f9/uQbrMgQySY2UW7z3R9gT5.jpeg","isPro":false,"fullname":"Ayoub Kirouane","user":"ayoubkirouane","type":"user","name":"ayoubkirouane"},"name":"Ayoub Kirouane","status":"claimed_verified","statusLastChangedAt":"2026-08-06T16:45:04.625Z","hidden":false},{"_id":"6a74940fe1228e04b3237ec9","user":{"_id":"6799d2770a1d27acbcc635bb","avatarUrl":"/avatars/66af80ba8bba84f9105551fc41314b42.svg","isPro":false,"fullname":"Christos Petrocheilos","user":"cpetos","type":"user","name":"cpetos"},"name":"Christos Petrocheilos","status":"claimed_verified","statusLastChangedAt":"2026-08-07T08:45:04.470Z","hidden":false}],"publishedAt":"2026-08-05T00:00:00.000Z","submittedOnDailyAt":"2026-08-07T00:00:00.000Z","title":"Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains","submittedOnDailyBy":{"_id":"6338c06c107c4835a05699f9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6338c06c107c4835a05699f9/uQbrMgQySY2UW7z3R9gT5.jpeg","isPro":false,"fullname":"Ayoub Kirouane","user":"ayoubkirouane","type":"user","name":"ayoubkirouane"},"summary":"Modern Greek is absent from NVIDIA's Nemotron retrieval models and from major multilingual retrieval benchmarks, despite being important for retrieval-augmented generation (RAG) in legal, energy, financial, and medical applications. We present an end-to-end adaptation of the 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. Our study shows that 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 and substantially outperforms its unadapted counterpart. The learned language competence transfers to general-domain Greek, although the advantage over BM25 remains domain-dependent. We further adapt a cross-encoder reranker and demonstrate consistent improvements across specialist domains. Finally, we LoRA-tune a Nemotron 30B-A3B mixture-of-experts reader for grounded generation, increasing judged answer correctness from 29.4% to 66.9% while significantly improving faithfulness and citation quality. We also introduce HERA, the first large-scale Greek benchmark for retrieval-augmented generation, and release our adapted models and benchmark to support future research on Greek-language RAG systems.","upvotes":16,"discussionId":"6a749410e1228e04b3237eca","organization":{"_id":"68f14d43f5266b706cf99e17","name":"KIEFERSA","fullname":"KIEFER","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68f14c3c6ede47be458ff051/7iTmA80GKtq3CNQDhvUoc.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6338c06c107c4835a05699f9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6338c06c107c4835a05699f9/uQbrMgQySY2UW7z3R9gT5.jpeg","isPro":false,"fullname":"Ayoub Kirouane","user":"ayoubkirouane","type":"user"},{"_id":"698d85c8a82ef9f4893bc3b0","avatarUrl":"/avatars/6664ce50983cb96346b94e750e8440ef.svg","isPro":false,"fullname":"Christos Porikis","user":"chrispor96","type":"user"},{"_id":"67ccb4e91bacab3ae8c5ebc1","avatarUrl":"/avatars/3632fbe99ea8f7b36d90e10c4029b24d.svg","isPro":false,"fullname":"Giaples Georgios","user":"gigigiapl","type":"user"},{"_id":"6799d2770a1d27acbcc635bb","avatarUrl":"/avatars/66af80ba8bba84f9105551fc41314b42.svg","isPro":false,"fullname":"Christos Petrocheilos","user":"cpetos","type":"user"},{"_id":"698d86a50eca9dd7d1179d76","avatarUrl":"/avatars/ee6f2050f992e2852492e041ff617aed.svg","isPro":false,"fullname":"Hasan Degismez","user":"hasankiefer","type":"user"},{"_id":"63cc5c5180ba2ca41528499a","avatarUrl":"/avatars/3b0a2c4813b1f8e7c6e14d3b7829ed9c.svg","isPro":false,"fullname":"Hs Dz","user":"pupdz","type":"user"},{"_id":"68a6dd8ecccf20993295dafe","avatarUrl":"/avatars/64651415a91e34ef688aee4c49d1afb9.svg","isPro":false,"fullname":"John Karvounas","user":"jkarvounas","type":"user"},{"_id":"698d8f2d7cd86ed5aac6b612","avatarUrl":"/avatars/dc7cca7a251a2f28b0c149aa12f3f759.svg","isPro":false,"fullname":"Themistoklis Nikolis","user":"kiefer-dev-1","type":"user"},{"_id":"6a75a625e7c0b16367422744","avatarUrl":"/avatars/595c601db74ddcb885203d5f32eb1416.svg","isPro":false,"fullname":"Lampis Papakostas","user":"lampispap","type":"user"},{"_id":"6a39091e977556eb25bc44fa","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/eJAo8xa1nlfD-EYp7WJZ_.png","isPro":false,"fullname":"Petr Levtonov","user":"TheLevti","type":"user"},{"_id":"6a34f40fc88b1fc59a2e5d99","avatarUrl":"/avatars/87d5267bcab507c417b172b167d1ab57.svg","isPro":false,"fullname":"Kimis Perros","user":"kperros-kiefer","type":"user"},{"_id":"6a75aaef22e97c5cc2df05ae","avatarUrl":"/avatars/2da12b8af7b5a23106e08f25f44c2748.svg","isPro":false,"fullname":"Apsotolos Stavridis","user":"Apostolos82","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"68f14d43f5266b706cf99e17","name":"KIEFERSA","fullname":"KIEFER","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/68f14c3c6ede47be458ff051/7iTmA80GKtq3CNQDhvUoc.png"},"query":{}}">

Papers

arxiv:2608.05138

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

Published on Aug 5

· Submitted by

Ayoub Kirouane on Aug 7

KIEFER

Upvote

Authors:

Ayoub Kirouane ,

Christos Petrocheilos

Abstract

Modern Greek is absent from NVIDIA's Nemotron retrieval models and from major multilingual retrieval benchmarks, despite being important for retrieval-augmented generation (RAG) in legal, energy, financial, and medical applications. We present an end-to-end adaptation of the 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. Our study shows that 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 and substantially outperforms its unadapted counterpart. The learned language competence transfers to general-domain Greek, although the advantage over BM25 remains domain-dependent. We further adapt a cross-encoder reranker and demonstrate consistent improvements across specialist domains. Finally, we LoRA-tune a Nemotron 30B-A3B mixture-of-experts reader for grounded generation, increasing judged answer correctness from 29.4% to 66.9% while significantly improving faithfulness and citation quality. We also introduce HERA, the first large-scale Greek benchmark for retrieval-augmented generation, and release our adapted models and benchmark to support future research on Greek-language RAG systems.

View arXiv page View PDF Add to collection