A compact yet powerful Russian cross-encoder reranker fine-tuned from deepvk/RuModernBERT-small (34M parameters) on the ARGA100/ru-reranker-mega dataset (3.2M query-document pairs).

Achieves NDCG@10=0.74444 on MTEB RuBQReranking — competitive with models 8–16× larger, making it the most efficient Russian reranker available.

Benchmark Results (MTEB RuBQReranking)

Model	Params	NDCG@10	Params/Eff.
`bge-reranker-v2-m3-en-ru`	375M	0.84387	1.0
`BAAI/bge-reranker-v2-m3`	568M	0.84384	1.0
`DiTy/cross-encoder-russian-msmarco`	178M	0.78572	2.3×
`mxbai-rerank-large-v1`	435M	0.77927	2.1×
`bce-reranker-base_v1`	278M	0.74599	2.7×
`ARGA100/ru-reranker-modernbert-small` (ours)	34M	0.74444	11.0×
`mxbai-rerank-base-v2`	494M	0.71268	1.9×
`mxbai-rerank-base-v1`	184M	0.71159	2.5×
`BAAI/bge-reranker-base`	278M	0.70550	2.4×
`BAAI/bge-reranker-large`	560M	0.69430	1.9×
`gte-reranker-modernbert-base`	150M	0.65751	3.0×
`mxbai-rerank-xsmall-v1`	71M	0.62594	4.7×

Params/Eff. = relative NDCG@10 per million parameters vs. the top model.

Our 34M model delivers 98% of BCE-reranker-base_v1's performance (0.74444 vs 0.74599) while being 8× smaller and 3× faster at inference.

Model Details

Base model: deepvk/RuModernBERT-small (34M params)
Architecture: CrossEncoder with ModernBertForSequenceClassification
Max sequence length: 2048 tokens
Training data: ARGA100/ru-reranker-mega — 3.2M Russian query-document pairs
Loss: BinaryCrossEntropyLoss with label smoothing
Training: 3 epochs, greedy soup of top checkpoints
Language: Russian

Usage

python

from sentence_transformers import CrossEncoder

model = CrossEncoder("ARGA100/ru-reranker-modernbert-small", max_length=2048)

# Score a query-document pair
pairs = [
    ["сколько калорий в яйце", "В одном курином яйце содержится около 70-80 ккал"],
    ["сколько калорий в яйце", "Яичный белок практически не содержит жиров"],
]
scores = model.predict(pairs)
print(scores)  # Higher = more relevant

# Rerank documents
query = "лучшие рестораны москвы"
documents = [
    "Топ-10 ресторанов Москвы с авторской кухней",
    "Как приготовить борщ дома",
    "Ресторан White Rabbit вошел в рейтинг лучших",
]
ranks = model.rank(query, documents, top_k=3)
for r in ranks:
    print(f"Score: {r['score']:.4f} | Doc: {documents[r['corpus_id']]}")

Training Details

Hardware: NVIDIA Tesla V100 16GB
Framework: sentence-transformers 5.6.0, transformers 5.12.1
Batch size: 16
Learning rate: 4.7e-5
Optimizer: AdamW
Warmup: 500 steps
Epochs: 3
Checkpoint selection: Greedy soup of all 5 checkpoints (best single ckpt: 0.72704, best soup: 0.74444)

Dataset

The model was trained on ARGA100/ru-reranker-mega — a deduplicated collection of 3.2M Russian query-document relevance pairs assembled from multiple public sources.

Citation

If you use this model, please cite:

bibtex

@misc{ru-reranker-modernbert-small,
  author = {ARGA100},
  title = {RuModernBERT-small Russian Cross-Encoder Reranker},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ARGA100/ru-reranker-modernbert-small}}
}