StaRSE

StaRSE stands for Static Russian Sentence Embeddings. It is a compact Russian sentence embedding model implemented as a Sentence-Transformers StaticEmbedding endpoint.

The model is intended for CPU-friendly semantic similarity, clustering, classification features, and retrieval-style first-stage representations when a full Transformer encoder is too expensive to run at high throughput.

StaRSE has 61.51M logical embedding parameters (120,138 × 512), stored as binary sign bits plus one FP32 L2 norm per vocabulary entry, making model.safetensors only 7.79 MiB. The Safetensors badge shows incorrect value.

Performance

Evaluation is reported on MTEB(rus, v1.1) across 23 tasks. The main score is mean_task_main_score = 51.16.

Task type	Tasks	Mean score
Classification	9	56.81
Clustering	3	51.80
MultilabelClassification	2	35.01
PairClassification	1	52.50
Reranking	2	41.88
Retrieval	3	39.09
STS	3	62.18

Usage

Install Sentence Transformers:

pip install -U sentence-transformers

Load the model with trust_remote_code=True.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("BorisTM/starse", trust_remote_code=True)

sentences = [
    "Партитуры Чайковского часто звучат в консерватории.",
    "Балетная сцена хранит музыку Щелкунчика.",
    "Футбольная команда выиграла матч.",
]

embeddings = model.encode(sentences, normalize_embeddings=True)
similarities = model.similarity(embeddings, embeddings)
print(embeddings.shape)           # (3, 512)
print(tuple(similarities.shape))  # (3, 3)
print(similarities)
# tensor([[1.0000, 0.3521, 0.0626],
#         [0.3521, 1.0000, 0.0420],
#         [0.0626, 0.0420, 1.0000]])

Citation

@misc{starse2026,
  title = {StaRSE: Compact Russian Sentence Embeddings with a Sign-Coded Static Encoder},
  year = {2026},
  url = {https://huggingface.co/BorisTM/starse}
}