The CTX recipe with Stage 1 (domain-adaptive MLM) skipped — contrastive straight off ModernBERT-base. Shows DAPT adds only ~0.1 once Signals A+B are present.

A 149M-parameter ModernBERT-base scientific document embedder trained with citation-context sentences as the primary contrastive signal. Part of the SciEmbed release (paper under double-blind review; author info omitted).

Usage

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("anon-nlp/sciembed-nodapt-ctx")
emb = model.encode(["citation-context supervision for scientific embeddings"],
                   normalize_embeddings=True)

Context length: 512 tokens
Pooling: mean · Output dim: 768 (Matryoshka-truncatable to 512/256/128)
License: MIT

SciRepEval (4-category macro)

Classif.	Regr.	Prox.	Search	Overall
75.3	28.2	80.8	82.6	66.7 ± 0.07

Citation

See the repository README. Paper: SciEmbed: Citation-Context Supervision for Scientific Document Embeddings (under review).