B

BERTA

Developed by sergeyzh
BERTA is obtained by distilling the embedding vectors of the FRIDA model into LaBSE-ru-turbo, which is used to calculate the embedding vectors of Russian and English sentences and supports multiple prefix tasks.
Downloads 7,089
Release Time : 3/10/2025

Model Overview

The BERTA model is a pre-trained model for calculating the embedding vectors of Russian and English sentences. It is obtained by distilling the embedding vectors of the FRIDA model into LaBSE-ru-turbo, retaining the Russian-English sentence embedding and prefix functions.

Model Features

Multi-prefix support
Supports multiple prefix tasks, such as semantic similarity, paraphrase recognition, natural language inference, etc., and optimizes task performance through different prefixes.
Distillation optimization
By distilling the embedding vectors of the FRIDA model into LaBSE-ru-turbo, it retains high performance while reducing the model complexity.
Multilingual support
Supports the calculation of sentence embeddings for Russian and English, suitable for cross-lingual tasks.

Model Capabilities

Calculate sentence embedding vectors
Semantic text similarity calculation
Paraphrase recognition
Natural language inference
Sentiment analysis
Toxicity recognition

Use Cases

Text classification
News title classification
Classify news titles with an accuracy of up to 0.891.
Accuracy 0.891
Movie review classification
Classify movie reviews by sentiment with an accuracy of 0.678.
Accuracy 0.678
Information retrieval
News retrieval
Used for news retrieval tasks, with an NDCG@10 of 0.816.
NDCG@10 0.816
Question-answering retrieval
Used for question-answering retrieval tasks, with an NDCG@10 of 0.710.
NDCG@10 0.710
Semantic similarity
Russian STS benchmark
Calculate the semantic similarity of Russian sentences with a Pearson correlation coefficient of 0.822.
Pearson correlation coefficient 0.822
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
© 2025AIbase