ayoubkirouane/Sophea-Nemo-Embedding

Sophea-Nemo-EmbeddingGreek + English dense retriever — fine-tuned NVIDIA Nemotron-3-Embed-1B

Sophea-Nemo-Embedding is a Greek + English dense retrieval embedder, fine-tuned from NVIDIA Nemotron-3-Embed-1B for Greek and English retrieval. It is the first stage of a two-stage Greek RAG stack; the matching second stage is Sophea-Nemo-Reranker.

• Creator: Kiefer SA
• Base model: nvidia/Nemotron-3-Embed-1B-BF16 (Ministral-3 backbone, bidirectional, mean pooling with include_prompt=true, 2048-d Matryoshka-sliceable, L2-normalized)
• Languages: Greek + English
• Prompts: query: on queries, passage: on documents (both sides prompted)
• Training: full fine-tune, InfoNCE (in-batch + 7 hard negatives), 1 epoch, max-len 4096, bf16

Greek is out-of-distribution for the base. Nemotron-3-Embed lists 34 languages and Greek is not one of them — which is exactly why the base retrieves poorly on Greek and why this fine-tune exists.

Why it matters — base vs fine-tuned

On our held-out 5-domain Greek+English retrieval eval (cosine IR, query: /passage: prompts), the fine-tune roughly 2.5×'s the base embedder's retrieval quality — on every domain.

Read. Mean nDCG@10 0.331 → 0.818 (+0.486), with the largest gains on the hardest domains (finance +0.53, medical +0.54). Recall@10 climbs from a mean 0.425 to 0.931 — i.e. the right passage is now almost always in the top-10.

Size comparison — vs base Qwen3-Embedding (off-the-shelf)

Same 5-domain retrieval eval, same corpus/prompts. Your fine-tuned 1B embedder beats the base Qwen3-Embedding-8B by a wide margin (nDCG@10):

Read. Mean nDCG@10 0.818 (ours, 1B) vs 0.664 (8B), 0.665 (4B), 0.585 (0.6B). Note the Qwen 8B ≈ 4B — scaling the off-the-shelf model barely helps on Greek, whereas a 1B fine-tuned on-domain wins by +0.153 over the 8×-larger model.

End-to-end two-stage pipeline

The two models are designed to run together: the embedder retrieves the top-50 candidates, the reranker re-orders them. Measured end-to-end on the same 5-domain Greek+English eval (150 queries/set, correct query:/passage: prompts):

Read. The full stack reaches mean nDCG@10 0.818 (Recall@10 0.927) — the strongest configuration measured. The reranker adds the most where the embedder is weakest (finance +0.061, legal +0.037), i.e. it sharpens precision on the hard cases. On energy it slightly regresses (-0.022) and on val the gain is marginal, because the embedder alone is already very strong there and leaves little headroom — for the easiest, already-high-recall domains you can skip the rerank stage if latency matters.

Usage (sentence-transformers)

from sentence_transformers import SentenceTransformer
m = SentenceTransformer("KIEFERSA/Sophea-Nemo-Embedding")
q = m.encode(["Ποια είναι η πρωτεύουσα της Ελλάδας;"], prompt="query: ", normalize_embeddings=True)
d = m.encode(["Η Αθήνα είναι η πρωτεύουσα της Ελλάδας."], prompt="passage: ", normalize_embeddings=True)
print((q @ d.T))   # cosine similarity

Both sides must be prompted (query: / passage: ) — the base uses include_prompt=true, so dropping the prefix is a distribution mismatch. Embeddings are L2-normalized (use cosine / dot).

Serving with vLLM

vllm serve KIEFERSA/Sophea-Nemo-Embedding --runner pooling

Exposes an OpenAI-compatible /v1/embeddings endpoint. vLLM does not auto-apply the model's ST prompts, so prepend them client-side — query: on queries, passage: on documents:

from openai import OpenAI
c = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
name = "KIEFERSA/Sophea-Nemo-Embedding"
q = c.embeddings.create(model=name, input=["query: Ποια είναι η πρωτεύουσα της Ελλάδας;"]).data[0].embedding
d = c.embeddings.create(model=name, input=["passage: Η Αθήνα είναι η πρωτεύουσα της Ελλάδας."]).data[0].embedding

License

Fine-tune of nvidia/Nemotron-3-Embed-1B-BF16 — governed by OpenMDW-1.1 (base Ministral-3 is Apache-2.0); commercial use permitted under those terms. Review the base model's license before use.