Hugging Face's logo

Datasets:
sabilmakbar
/
sea_wiki

Tasks:
Text Generation
Fill-Mask
Sub-tasks:
language-modeling
masked-language-modeling
Languages:
Achinese
Balinese
Central Bikol
Tags:
Wikipedia
Southeast Asia (SEA)
Dialect
Banyumasan Dialect of Javanese (Ngapak)
SEA-related Languages
SEA Local Languages
License:
Dataset card Files Community
6
sea_wiki / README.md
sabilmakbar's picture
sabilmakbar
Fix tag in source_datasets
b687797
preview code
|
raw
history blame
18.4 kB

SEA Wikipedia Data Repository

license: cc-by-sa-3.0

Welcome to SEA Wikipedia Data Repository. The datasets are extracted from Wikipedia HF and processed using the scripts available in this repository for reproducibility purpose.

Getting Started

To read the datasets directly

Use one of the following code chunks to load it from HuggingFace Hub: You can refer to the 2nd args of config name using the following script

dataset = load_dataset(
  "sabilmakbar/sea_wiki",
  "seawiki_dedup_all" # a config name, can be "seawiki_dedup_all" or "seawiki_with_countries_all", or "seawiki_with_countries_dedup_all" , defaults to "seawiki_dedup_all"
)

Or you can provide both lang and date_stamp (or just lang only by assuming the date_stamp will take the newest one)

dataset = load_dataset(
  "sabilmakbar/sea_wiki",
  lang = "id", # see README for complete lang choices
  date_stamp="20230901"
)

Or you can provide a country params with similar fashion to lang args (providing both country and lang will prioritize the lang kwarg)

dataset = load_dataset(
  "sabilmakbar/sea_wiki",
  lang = "id", # see the splits for complete lang choices
  date_stamp="20230901"
)

FAQS

What are the available languages provided in dataset and from which country?

You may check the following tables to understand the current coverage of this dataset (languages, countries, data size & volume).

1. Table of Countries and its Country Code

Country Code Country Name Wiki Info
brn Brunei Wiki Link
idn Indonesia Wiki Link
khm Cambodia Wiki Link
lao Laos Wiki Link
mmr Myanmar Wiki Link
mys Malaysia Wiki Link
sgp Singapore Wiki Link
tha Thailand Wiki Link
tls East Timor Wiki Link
vnm Vietnam Wiki Link

2. Table of Languages and Countries of its speakers

Lang Code Lang Name Country Codes Spoken Wiki Info Total Data Total Size (MiB rounded)
ace Acehnese idn Wiki Link 12904 4.64
ban Balinese idn Wiki Link 19837 16.56
bjn Banjarese idn Wiki Link 10437 6.35
bug Buginese idn Wiki Link 9793 1.98
gor Gorontalo idn Wiki Link 14514 5.71
km Khmer khm Wiki Link 11994 98.37
id Indonesian idn Wiki Link 654287 1049.93
jv Javanese idn Wiki Link 72667 66.54
lo Lao lao Wiki Link 5014 14.53
mad Madurese idn Wiki Link 1192 1.54
map_bms Banyumasan
(Dialect of Javanese)		 idn Wiki Link 11832 4.83
mnw Mon mmr Wiki Link 3296 45.13
min Minangkabau idn Wiki Link 225858 110.99
ms Malay mys, sgp, brn, idn Wiki Link 346186 391.43
my Burmese mmr Wiki Link 109310 298.85
nia Nias idn Wiki Link 1650 1.85
shn Shan mmr Wiki Link 13945 32.19
su Sundanese idn Wiki Link 61494 45.21
tet Tetum tls, idn Wiki Link 1465 1.39
th Thai tha Wiki Link 159719 966.00
vi Vietnamese vnm Wiki Link 1288680 1528.79

3. Table of Token Statistics for Covered Languages

The token statistics is generated using tiktoken using encoder for GPT-4.

Lang Code Total Token Avg Token per Article Min Token Max Token Token Deciles List
ace 1,370,829 105.61899992295247 3 9,659 [38.0, 52.0, 54.0, 69.0, 76.0, 84.0, 90.0, 123.0, 126.0]
ban 5,924,610 287.44893503469024 5 24,364 [97.0, 144.0, 165.0, 187.0, 209.0, 245.0, 276.0, 315.0, 421.0]
bjn 1,935,505 184.28115776444827 2 30,170 [36.0, 38.0, 39.0, 40.0, 42.0, 51.0, 82.0, 151.0, 367.0]
bug 553,693 55.54147858360919 1 13,951 [31.0, 42.0, 43.0, 46.0, 48.0, 50.0, 52.0, 55.0, 57.0]
gor 1,575,766 103.05860039241334 2 5,525 [55.0, 58.0, 60.0, 62.0, 64.0, 66.0, 69.0, 75.0, 96.0]
id 325,411,713 491.22975561670967 1 198,597 [54.0, 93.0, 123.0, 145.0, 180.0, 226.0, 332.0, 543.0, 1068.0]
jv 23,528,314 321.95284619594963 2 342,156 [48.0, 60.0, 75.0, 88.0, 117.0, 175.0, 270.0, 420.0, 772.0]
km 54,559,721 4,758.391854177568 1 1,110,771 [160.0, 293.0, 452.0, 693.0, 1032.0, 1609.0, 2644.0, 4745.0, 9607.0]
lo 9,395,636 1,918.6514192362672 3 107,154 [134.0, 184.2, 285.0, 494.0, 658.0, 894.6, 1258.0, 1971.2, 4153.8]
mad 611,736 513.2013422818792 14 17,093 [80.1, 110.2, 135.0, 161.0, 194.0, 242.0, 302.7, 531.4, 1167.1]
map_bms 1,307,244 110.41844750401216 1 20,629 [20.0, 21.0, 22.0, 24.0, 30.0, 35.0, 36.0, 38.0, 111.0]
min 33,114,184 146.54109358681606 3 58,387 [81.0, 91.0, 96.0, 108.0, 119.0, 135.0, 156.0, 168.0, 170.0]
mnw 31,595,647 9,659.3234484867 6 1,450,765 [425.0, 601.0, 629.0, 682.0, 763.0, 2103.0, 4255.0, 7724.0, 14517.0]
ms 121,343,673 348.64363228892813 1 68,545 [32.0, 40.0, 49.0, 63.0, 105.0, 138.0, 216.0, 362.0, 788.0]
my 189,439,447 1,740.8673761015998 10 1,376,658 [164.0, 269.0, 350.0, 508.0, 559.0, 578.0, 605.0, 892.4, 3369.0]
nia 795,527 464.134772462077 8 18,650 [59.0, 61.0, 63.0, 65.0, 67.0, 86.0, 239.1, 623.4, 1249.7]
shn 23,125,637 1,692.6977748499487 2 204,094 [460.0, 480.0, 585.0, 679.0, 715.0, 740.0, 756.0, 780.0, 1580.9]
su 14,710,124 239.07627297697022 1 99,456 [41.0, 43.0, 45.0, 49.0, 70.0, 146.0, 216.0, 219.0, 419.0]
tet 487,016 332.6612021857924 4 24,287 [30.3, 47.0, 66.9, 101.0, 164.0, 177.0, 187.0, 248.6, 604.4]
th 330,964,733 2,072.8566695476807 1 289,150 [231.0, 390.0, 546.0, 727.0, 969.0, 1276.0, 1741.0, 2533.0, 4361.0]
vi 546,481,258 424.3163404275143 3 246,463 [46.0, 64.0, 71.0, 80.0, 86.0, 92.0, 120.0, 240.0, 824.0]

Some other languages in SEA that are already exists its Wiki Index at Wikimedia might be missing from this list. Any lang update PR is greatly appreciated!

How does the data being preprocessed? What makes it different from loading it directly from Wikipedia HF?

The data available in here are processed with following flows:

  1. Raw data is being deduplicated on title and text (text-content from a given article), to remove articles containing boilerplate text (template text that are used usually for unavailable informations or asking for contributions of content in that article), which usually deemed noisy for NLP data.
  2. Furthermore, the title and text data are being checked for string-matching duplication (duplication of text that are being pre-processed, i.e symbols removed, HTML tags striped, or ASCII-chars/UTF-8 chars validated). You may check this dedup_raw_wiki_data.py script to understand its implementation.

How do I extract new Wikipedia Dataset of SEA languages?

You may check to the script extract_raw_wiki_data.py to understand its implementations, or you can adjust the bash provided in extract_raw_wiki_data_sea.sh to extract it on your own.

How do I extract new Wikipedia Dataset of SEA languages?

You may visit this Wikipedia Dump Index to check any latest available data and this link Wikipedia Language Coverage to map into any languages that you're wanting to extract. Please note that this dataset is extensible to any languages of your choice.

To replicate the whole dataset generation process

  1. Set-up a new Python/Conda Environment (recommended Python version: 3.9.6 to 3.9.18 or 3.10.0 to 3.10.13) and install the requirements on requirements.txt use this codebase via pip install -r requirements.txt.

  2. Activate the chosen Python/Conda environment which the requirements are being installed.

  3. Force install multiprocess==0.70.15 by using pip install multiprocess==0.70.15 to avoid this issue (there's no other workaround for now)

  4. Run this sh script for extractions from Wikiedia HF using sh extract_raw_wiki_data_sea.sh
    This script will run extract_raw_wiki_data.py to construct the Wiki Dataset.

  5. Run this sh script for deduplications from extracted data in Step 4 using sh dedup_raw_wiki_data_sea.sh
    This script will run dedup_raw_wiki_data.py to do Wiki Dataset Clenasing. Please note that the cleansing process can be language/dialect specific.

Citation Info: 

@ONLINE{wikidump,
    author = "Wikimedia Foundation",
    title  = "Wikimedia Downloads",
    url    = "https://dumps.wikimedia.org"}
@ONLINE{wikipedia-hf,
    title  = "Huggingface Wikipedia Dataset",
    url    = "https://huggingface.co/datasets/wikipedia"}