CulturaX dataset was used for model training. The ai-forever/ruBert-base model was used as the original; this model surpasses it in quality, according to the encodechka benchmark.

The model source code is available in the file modeling_rope_bert.py

The model is trained on contexts up to 512 tokens in length, but can be used on larger contexts. For better quality, use the version of this model with extended context - Tochka-AI/ruRoPEBert-classic-base-2k

Usage

Important: 4.37.2 and higher is the recommended version of transformers. To load the model correctly, you must enable dowloading code from the model's repository: trust_remote_code=True, this will download the modeling_rope_bert.py script and load the weights into the correct architecture. Otherwise, you can download this script manually and use classes from it directly to load the model.

Basic usage (no efficient attention)

python

model_name = 'Tochka-AI/ruRoPEBert-classic-base-512'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name, trust_remote_code=True, attn_implementation='eager')

With SDPA (efficient attention)

model = AutoModel.from_pretrained(model_name, trust_remote_code=True, attn_implementation='sdpa')

Getting embeddings

The correct pooler (mean) is already built into the model architecture, which averages embeddings based on the attention mask. You can also select the pooler type (first_token_transform), which performs a learnable linear transformation on the first token.

To change built-in pooler implementation use pooler_type parameter in AutoModel.from_pretrained function

python

test_batch = tokenizer.batch_encode_plus(["Привет, чем занят?", "Здравствуйте, чем вы занимаетесь?"], return_tensors='pt', padding=True)
with torch.inference_mode():
  pooled_output = model(**test_batch).pooler_output   

In addition, you can calculate cosine similarities between texts in batch using normalization and matrix multiplication:

python

import torch.nn.functional as F
F.normalize(pooled_output, dim=1) @ F.normalize(pooled_output, dim=1).T

Using as classifier

To load the model with trainable classification head on top (change num_labels parameter):

model = AutoModelForSequenceClassification.from_pretrained(model_name, trust_remote_code=True, attn_implementation='sdpa', num_labels=4)

With RoPE scaling

Allowed types for RoPE scaling are: linear and dynamic. To extend the model's context window you need to change tokenizer max length and add rope_scaling parameter.

If you want to scale your model context by 2x:

python

tokenizer.model_max_length = 1024
model = AutoModel.from_pretrained(model_name,
                                  trust_remote_code=True,
                                  attn_implementation='sdpa',
                                  rope_scaling={'type': 'dynamic','factor': 2.0}
                                  ) # 2.0 for x2 scaling, 4.0 for x4, etc..

P.S. Don't forget to specify the dtype and device you need to use resources efficiently.

Metrics

Evaluation of this model on encodechka benchmark:

Model name	STS	PI	NLI	SA	TI	IA	IC	ICX	NE1	NE2	Avg S (no NE)	Avg S+W (with NE)
ruRoPEBert-classic-base-512	0.695	0.605	0.396	0.794	0.975	0.797	0.769	0.386	0.410	0.609	0.677	0.630
ai-forever/ruBert-base	0.670	0.533	0.391	0.773	0.975	0.783	0.765	0.384	-	-	0.659	-

Authors

Sergei Bratchikov (Tochka AI Team, HF, GitHub)
Maxim Afanasiev (Tochka AI Team, HF, GitHub)