README.md

---
dataset_info:
  features:
  - name: idx
    dtype: int64
  - name: original
    dtype: string
  - name: en
    dtype: string
  - name: ru
    dtype: string
  - name: chrf_sim
    dtype: float64
  - name: labse_sim
    dtype: float64
  - name: forward_entailment
    dtype: float64
  - name: backward_entailment
    dtype: float64
  - name: p_good
    dtype: float64
  - name: len_tokens
    dtype: int64
  splits:
  - name: train
    num_bytes: 488921889
    num_examples: 936931
  - name: val
    num_bytes: 4988702
    num_examples: 9568
  - name: test
    num_bytes: 5004482
    num_examples: 9524
  download_size: 323982865
  dataset_size: 498915073
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: val
    path: data/val-*
  - split: test
    path: data/test-*
task_categories:
- text2text-generation
- sentence-similarity
language:
- ru
tags:
- paraphrase-generation
- paraphrase
size_categories:
- 100K<n<1M
---
## Dataset Description

The dataset is obtained by filtering [dataset of russian paraphrases by David Dale](https://huggingface.co/datasets/cointegrated/ru-paraphrase-NMT-Leipzig) with automatic metrics.
The data structure is saved.

Have been deleted:
 - Paraphrases that have cosine LABSE similarity with source sentences < 0.75.
 - Paraphrases that are 2.5 times longer than source sentences. (Most of them are looped errors of back translation)
 - Paraphrases that are similar in spelling to the original texts. Paraphrases that have ChrF++ similarity > 0.6 and consist of 15 or more tokens have been deleted.

Source dataset contained samples with English paraphrases (errors of back translation). This paraphrases have been translated into Russian.

### Supported Tasks and Leaderboards
The dataset can be used to train and evaluate models for paraphrase generation or (if negative sampling is used) for paraphrase detection.