nemotron-3-embed-8b-medical

An embedding model for biomedical and clinical text. It is built for searching the medical literature, retrieving evidence for clinical and consumer health questions, checking scientific claims against published work, powering RAG over clinical corpora, and organising large collections of research by subject.

Mean(Task) 68.80 on MTEB(Medical, v1).

It is built on nvidia/Nemotron-3-Embed-8B-BF16, already a strong biomedical retriever. The transformer weights are unchanged from it. What this model adds is a clustering head, a low-rank projection fitted on external PubMed literature and applied when clustering, worth +10.3 on MedrxivClusteringS2S.v2 and +8.2 on P2P.

Results

MTEB(Medical, v1), mteb 2.18.6.

Task	Score
PublicHealthQA	90.88
MedicalQARetrieval	87.16
TRECCOVID	86.78
TRECCOVID-PL	85.17
SciFact	83.33
SciFact-PL	80.71
CMedQAv2-reranking	79.91
CUREv1	66.92
MedrxivClusteringP2P.v2	42.77  (base 34.62)
NFCorpus	42.32
MedrxivClusteringS2S.v2	41.96  (base 31.65)
CmedqaRetrieval	37.64
Mean(Task)	68.80

Base figures are the same checkpoint without the head, under the same prompt.

The clustering head

Embedding spaces are organised for general similarity, so k-means over raw medical embeddings often groups papers by topic rather than by clinical subject. The head is fitted to correct that. We grouped PubMed abstracts by clinical subject, 46k for the title head and 25k for the title and abstract head, reduced the embedding space to 512 dimensions with PCA, then took the discriminant directions in that reduced space. The PCA step is necessary: LDA at 4096 dimensions fails on small-sample covariance estimation. The resulting basis is orthonormalised and applied as a rank-51 projection returning 4096-dimensional vectors, so output shape is unchanged.

Usage

Needs transformers>=5.2.0 and sentence-transformers>=5.4.1, same as the base model.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("minetta/nemotron-3-embed-8b-medical")

q = model.encode(["what causes myocarditis?"], prompt_name="query")
d = model.encode(["Myocarditis is an inflammation..."], prompt_name="document")

For clustering, apply the head:

import numpy as np
from huggingface_hub import hf_hub_download

h = np.load(hf_hub_download("minetta/nemotron-3-embed-8b-medical",
                            "clustering_head/head_s2s.npz"))
mu, P = h["mu"].astype(np.float32), h["P"].astype(np.float32)

titles = ["Tocilizumab in severe COVID-19 pneumonia",
          "Hypertension prevalence among adults in rural Bangladesh"]
x = model.encode(titles, prompt_name="clustering").astype(np.float32)
z = ((x - mu) @ P.T) @ P
z /= np.maximum(np.linalg.norm(z, axis=1, keepdims=True), 1e-12)

head_s2s.npz for titles, head_p2p.npz for title+abstract. The projection returns 4096-dimensional vectors, so output dimensionality is unchanged.

Zero-shot

100% zero-shot on MTEB(Medical, v1). No MTEB task dataset, or anything derived from one, was used at any stage. The head is fitted on external PubMed literature; every document was checked for n-gram overlap against all twelve evaluation corpora before use. The gate, the counts and what it removed are in CONTAMINATION.md.

Limitations

• The head is fitted for clinical subject categories and is not expected to transfer to unrelated taxonomies.
• default_prompt_name is clustering; pass prompt_name="document" when encoding passages.
• Chinese and Polish performance is inherited from the base model.
• Benchmark runs truncate at 4096 tokens, matching the base model's own MTEB configuration. The checkpoint accepts 32k, but long documents at large batch sizes will exhaust an 80GB card.
• Scores were produced with mteb 2.18.6.

Acknowledgements

Thanks to NVIDIA for Nemotron-3-Embed-8B-BF16. It is a very capable base model and a strong starting point for medical work.

---

Built by minetta.ai