---
language:
- en
dataset_info:
- config_name: 100M_1
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 645762137.0362595
    num_examples: 225498
  - name: validation
    num_bytes: 2863715.5852214186
    num_examples: 1000
  download_size: 389527538
  dataset_size: 648625852.621481
- config_name: 100M_2
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 646074282.0350486
    num_examples: 225607
  - name: validation
    num_bytes: 2863715.5852214186
    num_examples: 1000
  download_size: 389381161
  dataset_size: 648937997.62027
- config_name: 100M_3
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 650730683.5766187
    num_examples: 227233
  - name: validation
    num_bytes: 2863715.5852214186
    num_examples: 1000
  download_size: 390292303
  dataset_size: 653594399.1618401
- config_name: 10M_1
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 64820202.27148681
    num_examples: 22635
  - name: validation
    num_bytes: 2863715.5852214186
    num_examples: 1000
  download_size: 40370445
  dataset_size: 67683917.85670823
- config_name: 10M_2
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 64236004.292101644
    num_examples: 22431
  - name: validation
    num_bytes: 2863715.5852214186
    num_examples: 1000
  download_size: 40412205
  dataset_size: 67099719.87732306
- config_name: 10M_3
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 63250886.13078547
    num_examples: 22087
  - name: validation
    num_bytes: 2863715.5852214186
    num_examples: 1000
  download_size: 40514801
  dataset_size: 66114601.71600689
- config_name: all
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: title
    dtype: string
  - name: text
    dtype: string
  splits:
  - name: train
    num_bytes: 18350156819
    num_examples: 6407814
  download_size: 10723740674
  dataset_size: 18350156819
configs:
- config_name: 100M_1
  data_files:
  - split: train
    path: 100M_1/train-*
  - split: validation
    path: 100M_1/validation-*
- config_name: 100M_2
  data_files:
  - split: train
    path: 100M_2/train-*
  - split: validation
    path: 100M_2/validation-*
- config_name: 100M_3
  data_files:
  - split: train
    path: 100M_3/train-*
  - split: validation
    path: 100M_3/validation-*
- config_name: 10M_1
  data_files:
  - split: train
    path: 10M_1/train-*
  - split: validation
    path: 10M_1/validation-*
- config_name: 10M_2
  data_files:
  - split: train
    path: 10M_2/train-*
  - split: validation
    path: 10M_2/validation-*
- config_name: 10M_3
  data_files:
  - split: train
    path: 10M_3/train-*
  - split: validation
    path: 10M_3/validation-*
- config_name: all
  data_files:
  - split: train
    path: all/train-*
---
This repository contains random subsets of the English wikipedia obtained from 
[`"wikimedia/wikipedia"`](https://huggingface.co/datasets/wikimedia/wikipedia) (`"20231101.en"`). 
It includes two random subsets of the English wikipedia, one containing roughly 10M words total (23k articles), the other containing roughly 100M words total (228K articles). 
These data are intended to be used for the BabyLM challenge. For convenience, the repository also includes the full English wikipedia containing roughly 2.8B words total 
(6.4M articles).

You can load these datasets as follows:
```python
from datasets import load_dataset

ds_10M = load_dataset("eminorhan/wikipedia", "10M")  # 10M word subset

ds_100M = load_dataset("eminorhan/wikipedia", "100M")  # 100M word subset

ds_all = load_dataset("eminorhan/wikipedia", "all")  # the full data (2.8B words)
```
Both subsets come with `train`/`validation` splits, whereas the full data only has a `train` split.
We applied lightweight preprocessing to the article texts using [this script](https://github.com/eminorhan/babylm/blob/master/create_random_wikipedia.py), 
which mainly strips away some sections of the articles like "References", "See also", *etc.*