---
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: complex
    dtype: string
  - name: simple
    dtype: string
  - name: split
    dtype: string
  splits:
  - name: train
    num_bytes: 309170607
    num_examples: 795585
  - name: validation
    num_bytes: 38667164
    num_examples: 99448
  - name: test
    num_bytes: 38650132
    num_examples: 99448
  - name: all
    num_bytes: 386487903
    num_examples: 994481
  download_size: 540598777
  dataset_size: 772975806
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
  - split: all
    path: data/all-*
license: cc-by-sa-4.0
task_categories:
- text2text-generation
language:
- en
pretty_name: WikiSplit
size_categories:
- 100K<n<1M
---

Preprocessed version of [WikiSplit](https://arxiv.org/abs/1808.09468).  
Since the [original WikiSplit dataset](https://huggingface.co/datasets/wiki_split) was tokenized and had some noises, we have used the [Moses detokenizer](https://github.com/moses-smt/mosesdecoder/blob/c41ff18111f58907f9259165e95e657605f4c457/scripts/tokenizer/detokenizer.perl) for detokenization and removed text fragments.  
For detailed information on the preprocessing steps, please see [here](https://github.com/nttcslab-nlp/wikisplit-pp/src/datasets/common.py).  
This preprocessed dataset serves as the basis for [WikiSplit++](https://huggingface.co/datasets/cl-nagoya/wikisplit-pp).

## Dataset Description

- **Repository:** https://github.com/nttcslab-nlp/wikisplit-pp
- **Used Paper:** https://arxiv.org/abs/2404.09002