RuModernBERT-small

The Russian version of the modernized bidirectional encoder-only Transformer model, ModernBERT. RuModernBERT was pre-trained on approximately 2 trillion tokens of Russian, English, and code data with a context length of up to 8,192 tokens, using data from the internet, books, scientific sources, and social media.

	Model Size	Hidden Dim	Num Layers	Vocab Size	Context Length	Task
deepvk/RuModernBERT-small [this]	35M	384	12	50368	8192	Masked LM
deepvk/RuModernBERT-base	150M	768	22	50368	8192	Masked LM

Notice ⚠️

The patched tokenizer is provided under the patched-tokenizer revision.

Details

We observed that several Russian lowercase letters were split into multiple subword tokens. This can be problematic for tasks like Named Entity Recognition (NER), where it is important that the first token of a word is a semantically meaningful unit.

To address this, we release a patched revision of the tokenizer with minimal but targeted change. Six common Russian lowercase letters (а, е, и, н, о, т) are now encoded as single tokens. These tokens were assigned to [unusedX] slots in the vocabulary. Corresponding BPE merges were added to ensure proper single-token encoding during inference. To preserve compatibility with the pretrained model each new token was initialized with the embedding of its uppercase counterpart both in tok_embedding and lm_head. To prevent duplicate vectors and maintain robustness, a small amount of Gaussian noise was added during initialization with gamma 1e-3.

We evaluated the patched model on 20 tasks from the RuMTEB benchmark and did not observe any statistically significant differences in performance compared to the original version. If your task is sensitive to tokenization granularity, such as in NER, we recommend using this updated version.

Usage example:

from transformers import AutoTokenizer, AutoModelForMaskedLM

model_id = "deepvk/RuModernBERT-small"

# You can specify revision
revision = "patched-tokenizer"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = AutoModelForMaskedLM.from_pretrained(model_id, revision=revision, attn_implementation="flash_attention_2")

Usage

Don't forget to update transformers and install flash-attn if your GPU supports it.

from transformers import AutoTokenizer, AutoModelForMaskedLM

# Prepare model
model_id = "deepvk/RuModernBERT-small"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForMaskedLM.from_pretrained(model_id, attn_implementation="flash_attention_2")
model = model.eval()

# Prepare input
text = "Мама мыла [MASK]."
inputs = tokenizer(text, return_tensors="pt")
masked_index = inputs["input_ids"][0].tolist().index(tokenizer.mask_token_id)

# Make prediction
outputs = model(**inputs)

# Show prediction
predicted_token_id = outputs.logits[0, masked_index].argmax(axis=-1)
predicted_token = tokenizer.decode(predicted_token_id)
print("Predicted token:", predicted_token)
# Predicted token:  посуду

Training Details

This is the small version with 35 million parameters.

Tokenizer

We trained a new tokenizer following the original configuration. We maintained the size of the vocabulary and added the same special tokens. The tokenizer was trained on a mixture of Russian and English from FineWeb.

Dataset

Pre-training includes three main stages: massive pre-training, context extension, and cooldown. Unlike the original model, we did not use the same data for all stages. For the second and third stages, we used cleaner data sources.

Data Source	Stage 1	Stage 2	Stage 3
FineWeb (En+Ru)	✅	❌	❌
CulturaX-Ru-Edu (Ru)	❌	✅	❌
Wiki (En+Ru)	✅	✅	✅
ArXiv (En)	✅	✅	✅
Book (En+Ru)	✅	✅	✅
Code	✅	✅	✅
StackExchange (En+Ru)	✅	✅	✅
Social (Ru)	✅	✅	✅
Total Tokens	1.3T	250B	50B

Context length

In the first stage, the model was trained with a context length of 1,024. In the second and third stages, it was extended to 8,192.

Evaluation

To evaluate the model, we measure quality on the encodechka and Russian Super Glue (RSG) benchmarks. For RSG, we perform a grid search for optimal hyperparameters and report metrics from the dev split.

For a fair comparison, we compare the RuModernBERT model only with raw encoders that were not trained on retrieval or sentence embedding tasks.

Russian Super Glue

Model	RCB	PARus	MuSeRC	TERRa	RUSSE	RWSD	DaNetQA	Score
deepvk/deberta-v1-distill	0.433	0.56	0.625	0.590	0.943	0.569	0.726	0.635
deepvk/deberta-v1-base	0.450	0.61	0.722	0.704	0.948	0.578	0.760	0.682
ai-forever/ruBert-base	0.491	0.61	0.663	0.769	0.962	0.574	0.678	0.678
deepvk/RuModernBERT-small [this]	0.555	0.64	0.746	0.593	0.930	0.574	0.743	0.683
deepvk/RuModernBERT-base	0.556	0.61	0.857	0.818	0.977	0.583	0.758	0.737

Encodechka

	Model Size	STS-B	Paraphraser	XNLI	Sentiment	Toxicity	Inappropriateness	Intents	IntentsX	FactRu	RuDReC	Avg. S	Avg. S+W
cointegrated/rubert-tiny	11.9M	0.66	0.53	0.40	0.71	0.89	0.68	0.70	0.58	0.24	0.34	0.645	0.575
deepvk/deberta-v1-distill	81.5M	0.70	0.57	0.38	0.77	0.98	0.79	0.77	0.36	0.36	0.44	0.665	0.612
deepvk/deberta-v1-base	124M	0.68	0.54	0.38	0.76	0.98	0.80	0.78	0.29	0.29	0.40	0.653	0.591
answerdotai/ModernBERT-base	150M	0.50	0.29	0.36	0.64	0.79	0.62	0.59	0.10	0.22	0.20	0.486	0.431
ai-forever/ruBert-base	178M	0.67	0.53	0.39	0.77	0.98	0.78	0.77	0.38	🥴	🥴	0.659	🥴
DeepPavlov/rubert-base-cased	180M	0.63	0.50	0.38	0.73	0.94	0.74	0.74	0.31	🥴	🥴	0.621	🥴
deepvk/RuModernBERT-small [this]	35M	0.64	0.50	0.36	0.72	0.95	0.73	0.72	0.47	0.28	0.26	0.636	0.563
deepvk/RuModernBERT-base	150M	0.67	0.54	0.35	0.75	0.97	0.76	0.76	0.58	0.37	0.36	0.673	0.611

Citation

@misc{deepvk2025rumodernbert,
    title={RuModernBERT: Modernized BERT for Russian},
    author={Spirin, Egor and Malashenko, Boris and Sokolov Andrey},
    url={https://huggingface.co/deepvk/rumodernbert-base},
    publisher={Hugging Face}
    year={2025},
}