USER-base

Universal Sentence Encoder for Russian (USER) is a sentence-transformer model for extracting embeddings exclusively for Russian language. It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.

This model is initialized from deepvk/deberta-v1-base and trained to work exclusively with the Russian language. Its quality on other languages was not evaluated.

Usage

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

from sentence_transformers import SentenceTransformer

queries = [
  "Когда был спущен на воду первый миноносец «Спокойный»?",
  "Есть ли нефть в Удмуртии?"
]
passages = [
  "Спокойный (эсминец)\nЗачислен в списки ВМФ СССР 19 августа 1952 года.",
  "Нефтепоисковые работы в Удмуртии были начаты сразу после Второй мировой войны в 1945 году и продолжаются по сей день. Добыча нефти началась в 1967 году."
]

model = SentenceTransformer("deepvk/USER-base")
# Prompt should be specified according to the task (either 'query' or 'passage').
passage_embeddings = model.encode(passages, normalize_embeddings=True, prompt_name='passage')
# For tasks other than retrieval, you can simply use the `query` prompt, which is set by default.
query_embeddings = model.encode(queries, normalize_embeddings=True)

However, you can use model directly with transformers

import torch.nn.functional as F
from torch import Tensor, inference_mode
from transformers import AutoTokenizer, AutoModel

def average_pool(
  last_hidden_states: Tensor,
  attention_mask: Tensor
) -> Tensor:
    last_hidden = last_hidden_states.masked_fill(
      ~attention_mask[..., None].bool(), 0.0
    )
    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]

# You should manually add prompts when using the model directly. Each input text should start with "query: " or "passage: ".
# For tasks other than retrieval, you can simply use the "query: " prefix.
input_texts = [
  "query: Когда был спущен на воду первый миноносец «Спокойный»?",
  "query: Есть ли нефть в Удмуртии?",
  "passage: Спокойный (эсминец)\nЗачислен в списки ВМФ СССР 19 августа 1952 года.",
  "passage: Нефтепоисковые работы в Удмуртии были начаты сразу после Второй мировой войны в 1945 году и продолжаются по сей день. Добыча нефти началась в 1967 году."
]

tokenizer = AutoTokenizer.from_pretrained("deepvk/USER-base")
model = AutoModel.from_pretrained("deepvk/USER-base")

batch_dict = tokenizer(
  input_texts, padding=True, truncation=True, return_tensors="pt"
)
with inference_mode():
  outputs = model(**batch_dict)
  embeddings = average_pool(
    outputs.last_hidden_state, batch_dict["attention_mask"]
  )
  embeddings = F.normalize(embeddings, p=2, dim=1)

# Scores for query-passage
scores = (embeddings[:2] @ embeddings[2:].T) * 100
# [[55.86, 30.95],
#  [22.82, 59.46]]
print(scores.round(decimals=2))

⚠️ Attention ⚠️

Each input text should start with "query: " or "passage: ". For tasks other than retrieval, you can simply use the "query: " prefix.

Training Details

We aimed to follow the bge-base-en model training algorithm, but we made several improvements along the way.

Initialization: deepvk/deberta-v1-base

First-stage: Contrastive pre-training with weak supervision on the Russian part of mMarco corpus.

Second-stage: Supervised fine-tuning two different models based on data symmetry and then merging via LM-Cocktail:

  1. We modified the instruction design by simplifying the multilingual approach to facilitate easier inference. For symmetric data (S1, S2), we used the instructions: "query: S1" and "query: S2", and for asymmetric data, we used "query: S1" with "passage: S2".

  2. Since we split the data, we could additionally apply the AnglE loss to the symmetric model, which enhances performance on symmetric tasks.

  3. Finally, we combined the two models, tuning the weights for the merger using LM-Cocktail to produce the final model, USER.

Dataset

During model development, we additional collect 2 datasets: deepvk/ru-HNP and deepvk/ru-WANLI.

Symmetric Dataset Size Asymmetric Dataset Size
AllNLI 282 644 MIRACL 10 000
MedNLI 3 699 MLDR 1 864
RCB 392 Lenta 185 972
Terra 1 359 Mlsum 51 112
Tapaco 91 240 Mr-TyDi 536 600
Opus100 1 000 000 Panorama 11 024
BiblePar 62 195 PravoIsrael 26 364
RudetoxifierDataDetox 31 407 Xlsum 124 486
RuParadetox 11 090 Fialka-v1 130 000
deepvk/ru-WANLI 35 455 RussianKeywords 16 461
deepvk/ru-HNP 500 000 Gazeta 121 928
Gsm8k-ru 7 470
DSumRu 27 191
SummDialogNews 75 700

Total positive pairs: 3,352,653
Total negative pairs: 792,644 (negative pairs from AIINLI, MIRACL, deepvk/ru-WANLI, deepvk/ru-HNP)

For all labeled datasets, we only use its training set for fine-tuning. For datasets Gazeta, Mlsum, Xlsum: pairs (title/text) and (title/summary) are combined and used as asymmetric data.

AllNLI is an translated to Russian combination of SNLI, MNLI, and ANLI.

Experiments

As a baseline, we chose the current top models from the encodechka leaderboard table. In addition, we evaluate model on the russian subset of MTEB, which include 10 tasks. Unfortunately, we could not validate the bge-m3 on some MTEB tasks, specifically clustering, due to excessive computational resources. Besides these two benchmarks, we also evaluated the models on the MIRACL. All experiments were conducted using NVIDIA TESLA A100 40 GB GPU. We use validation scripts from the official repositories for each of the tasks.

Model Size (w/o Embeddings) Encodechka (Mean S) MTEB (Mean Ru) Miracl (Recall@100)
bge-m3 303 0.786 0.694 0.959
multilingual-e5-large 303 0.78 0.665 0.927
USER (this model) 85 0.772 0.666 0.763
paraphrase-multilingual-mpnet-base-v2 85 0.76 0.625 0.149
multilingual-e5-base 85 0.756 0.645 0.915
LaBSE-en-ru 85 0.74 0.599 0.327
sn-xlm-roberta-base-snli-mnli-anli-xnli 85 0.74 0.593 0.08

Model sizes are shown, with larger models visually distinct from the others. Absolute leaders in the metrics are highlighted in bold, and the leaders among models of our size is underlined.

In this way, our solution outperforms all other models of the same size on both Encodechka and MTEB. Given that the model is slightly underperforming in retrieval tasks relative to existing solutions, we aim to address this in our future research.

FAQ

Do I need to add the prefix "query: " and "passage: " to input texts?

Yes, this is how the model is trained, otherwise you will see a performance degradation. Here are some rules of thumb:

  • Use "query: " and "passage: " correspondingly for asymmetric tasks such as passage retrieval in open QA, ad-hoc information retrieval.
  • Use "query: " prefix for symmetric tasks such as semantic similarity, bitext mining, paraphrase retrieval.
  • Use "query: " prefix if you want to use embeddings as features, such as linear probing classification, clustering.

Citations

@misc{deepvk2024user,
    title={USER: Universal Sentence Encoder for Russian},
    author={Malashenko, Boris and  Zemerov, Anton and Spirin, Egor},
    url={https://huggingface.co/datasets/deepvk/USER-base},
    publisher={Hugging Face}
    year={2024},
}
Downloads last month
21,044
Safetensors
Model size
124M params
Tensor type
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Datasets used to train deepvk/USER-base

Collection including deepvk/USER-base