Tochka-AI's picture
Update README.md
2cb9947 verified
metadata
library_name: transformers
language:
  - ru
pipeline_tag: feature-extraction
datasets:
  - uonlp/CulturaX

ruRoPEBert Sentence Model for Russian language

This is an encoder model from Tochka AI based on the RoPEBert architecture, using the cloning method described in our article on Habr.

CulturaX dataset was used for model training. The hivaze/ru-e5-base (only english and russian embeddings of intfloat/multilingual-e5-base) model was used as the original; this model surpasses it and all other models in quality (at the time of creation), according to the S+W score of encodechka benchmark.

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

The model is trained on contexts up to 2048 tokens in length, but can be used on larger contexts.

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)

model_name = 'Tochka-AI/ruRoPEBert-e5-base-2k'
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

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:

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:

tokenizer.model_max_length = 4096
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-e5-base-512 0.793 0.704 0.457 0.803 0.970 0.788 0.802 0.749 0.328 0.396 0.758 0.679
ruRoPEBert-e5-base-2k 0.787 0.708 0.460 0.804 0.970 0.792 0.803 0.749 0.402 0.423 0.759 0.689
intfloat/multilingual-e5-base 0.834 0.704 0.458 0.795 0.964 0.782 0.803 0.740 0.234 0.373 0.76 0.668

Authors

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