five

sabilmakbar/indo_wiki

收藏
Hugging Face2023-11-03 更新2024-03-04 收录
下载链接:
https://hf-mirror.com/datasets/sabilmakbar/indo_wiki
下载链接
链接失效反馈
官方服务:
资源简介:
--- annotations_creators: - no-annotation language_creators: - crowdsourced language: - ace - ban - bjn - bug - gor - id - jv - mis - min - ms - nia - su - tet license: - cc-by-sa-3.0 - gfdl multilinguality: - multilingual source_datasets: - Wikipedia-HF task_categories: - text-generation - fill-mask task_ids: - language-modeling - masked-language-modeling pretty_name: Wikipedia Archive for Indonesian Languages & Local Languages tags: - Wikipedia - Indonesian - Sundanese - Javanese - Malay - Dialect - Javanese Dialect (Banyumase/Ngapak) - Indonesian Language - Malay Language - Indonesia-related Languages - Indonesian Local Languages dataset_info: - config_name: indowiki_all features: - name: url dtype: string - name: title dtype: string - name: text dtype: string splits: - name: ace num_bytes: 4875688 num_examples: 12932 - name: ban num_bytes: 17561379 num_examples: 20243 - name: bjn num_bytes: 6669628 num_examples: 10460 - name: bug num_bytes: 3297641 num_examples: 15877 - name: gor num_bytes: 6007726 num_examples: 14572 - name: id num_bytes: 1103106307 num_examples: 657990 - name: jv num_bytes: 70335030 num_examples: 73150 - name: map_bms num_bytes: 5215803 num_examples: 13574 - name: min num_bytes: 116481049 num_examples: 227024 - name: ms num_bytes: 416001194 num_examples: 367463 - name: nia num_bytes: 1938378 num_examples: 1651 - name: su num_bytes: 47489084 num_examples: 61557 - name: tet num_bytes: 1452716 num_examples: 1465 download_size: 1803193334 dataset_size: 1800431623 - config_name: indowiki_dedup_all features: - name: url dtype: string - name: title dtype: string - name: text dtype: string splits: - name: ace num_bytes: 4867838 num_examples: 12904 - name: ban num_bytes: 17366080 num_examples: 19837 - name: bjn num_bytes: 6655378 num_examples: 10437 - name: bug num_bytes: 2072609 num_examples: 9793 - name: gor num_bytes: 5989252 num_examples: 14514 - name: id num_bytes: 1100932403 num_examples: 654287 - name: jv num_bytes: 69774853 num_examples: 72667 - name: map_bms num_bytes: 5060989 num_examples: 11832 - name: min num_bytes: 116376870 num_examples: 225858 - name: ms num_bytes: 410443550 num_examples: 346186 - name: nia num_bytes: 1938121 num_examples: 1650 - name: su num_bytes: 47410439 num_examples: 61494 - name: tet num_bytes: 1447926 num_examples: 1460 download_size: 1793103024 dataset_size: 1790336308 - config_name: indowiki_dedup_id_only features: - name: url dtype: string - name: title dtype: string - name: text dtype: string splits: - name: train num_bytes: 1100932403 num_examples: 654287 download_size: 1103131493 dataset_size: 1100932403 --- # **Indonesian Wikipedia Data Repository** --- license: cc-by-sa-3.0 --- Welcome to Indonesian Wikipedia Data Repository. The datasets are extracted from [Wikipedia HF](https://huggingface.co/datasets/wikipedia) and processed using the scripts available in this repository for reproducibility purpose. # **FAQS** ### What are the available languages provided in dataset? Please check the following table. | Lang Code | Lang Desc | Wiki Info | Total Data | Total Size (bytes) | | :---: | :----: | :--- | ---: | ---: | | ace | Acehnese | [Wiki Link](https://en.wikipedia.org/wiki/Acehnese_language) | 12904 | 4867838 | | ban | Balinese | [Wiki Link](https://en.wikipedia.org/wiki/Balinese_language) | 19837 | 17366080 | | bjn | Acehnese | [Wiki Link](https://en.wikipedia.org/wiki/Banjarese_language) | 10437 | 6655378 | | bug | Buginese | [Wiki Link](https://en.wikipedia.org/wiki/Buginese_language) | 9793 | 2072609 | | gor | Gorontalo | [Wiki Link](https://en.wikipedia.org/wiki/Gorontalo_language) | 14514 | 5989252 | | id | Indonesian | [Wiki Link](https://en.wikipedia.org/wiki/Indonesian_language) | 654287 | 1100932403 | | jv | Javanese | [Wiki Link](https://en.wikipedia.org/wiki/Javanese_language) | 72667 | 69774853 | | map_bms | Banyumasan <br />(Dialect of Javanese) | [Wiki Link](https://en.wikipedia.org/wiki/Banyumasan_dialect) | 11832 | 5060989 | | min | Minangkabau | [Wiki Link](https://en.wikipedia.org/wiki/Minangkabau_language) | 225858 | 116376870 | | ms | Malay | [Wiki Link](https://en.wikipedia.org/wiki/Malay_language) | 346186 | 410443550 | | nia | Nias | [Wiki Link](https://en.wikipedia.org/wiki/Nias_language) | 1650 | 1938121 | | su | Sundanese | [Wiki Link](https://en.wikipedia.org/wiki/Sundanese_language) | 61494 | 47410439 | | tet | Tetum | [Wiki Link](https://en.wikipedia.org/wiki/Tetum_language) | 1465 | 1452716 | ### How do I extract new Wikipedia Dataset of Indonesian languages? You may check to the script [_```extract_raw_wiki_data.py```_](https://huggingface.co/datasets/sabilmakbar/indo_wiki/blob/main/extract_raw_wiki_data.py) to understand its implementations, or you can adjust the bash provided in [_```extract_raw_wiki_data_indo.sh```_](https://huggingface.co/datasets/sabilmakbar/indo_wiki/blob/main/extract_raw_wiki_data_indo.sh) to extract it on your own. Please note that this dataset is extensible to any languages of your choice. ### How do I extract new Wikipedia Dataset of Indonesian languages? You may visit this [Wikipedia Dump Index](https://dumps.wikimedia.org/backup-index.html) to check any latest available data and this link [Wikipedia Language Coverage](https://meta.wikimedia.org/wiki/List_of_Wikipedias#All_Wikipedias_ordered_by_number_of_articles) to map into any languages that you're wanting to extract. ### 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 no-available 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 validated). You may check this [ ```cleanse_wiki_data.py```](https://huggingface.co/datasets/sabilmakbar/indo_wiki/blob/main/cleanse_wiki_data.py) script to understand its implementation. # 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/indo_wiki", "indo_wiki_dedup_data" # a config name, can be "indo_wiki_raw_data" or "indowiki_dedup_id_only", defaults to "indo_wiki_dedup_data" ) ``` Or you can provide both ```lang``` and ```date_stamp``` (providing only one will thrown an error) ``` dataset = load_dataset( "sabilmakbar/indo_wiki", lang = "id", # see the splits for complete lang choices date_stamp="20230901" ) ``` ### 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. Run this ```sh``` script for extractions from Wikimedia Dump: ```sh extract_raw_wiki_data_indo.sh```. 4. Run this ```sh``` script of deduplication: ```sh dedup_raw_wiki_data_indo.sh```. ## 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"} ```
提供机构:
sabilmakbar
原始信息汇总

数据集概述

基本信息

  • 数据集名称: Wikipedia Archive for Indonesian Languages & Local Languages
  • 标签: Wikipedia, Indonesian, Sundanese, Javanese, Malay, Dialect, Javanese Dialect (Banyumase/Ngapak), Indonesian Language, Malay Language, Indonesia-related Languages, Indonesian Local Languages
  • 许可证: cc-by-sa-3.0, gfdl
  • 多语言性: 多语言
  • 源数据集: Wikipedia-HF
  • 任务类别: 文本生成, 填充掩码
  • 任务ID: 语言建模, 掩码语言建模

语言信息

  • 支持语言: ace, ban, bjn, bug, gor, id, jv, mis, min, ms, nia, su, tet

数据集配置

配置: indowiki_all

  • 特征:
    • url: string
    • title: string
    • text: string
  • 分割:
    • ace: 12932个样本, 4875688字节
    • ban: 20243个样本, 17561379字节
    • bjn: 10460个样本, 6669628字节
    • bug: 15877个样本, 3297641字节
    • gor: 14572个样本, 6007726字节
    • id: 657990个样本, 1103106307字节
    • jv: 73150个样本, 70335030字节
    • map_bms: 13574个样本, 5215803字节
    • min: 227024个样本, 116481049字节
    • ms: 367463个样本, 416001194字节
    • nia: 1651个样本, 1938378字节
    • su: 61557个样本, 47489084字节
    • tet: 1465个样本, 1452716字节
  • 下载大小: 1803193334字节
  • 数据集大小: 1800431623字节

配置: indowiki_dedup_all

  • 特征:
    • url: string
    • title: string
    • text: string
  • 分割:
    • ace: 12904个样本, 4867838字节
    • ban: 19837个样本, 17366080字节
    • bjn: 10437个样本, 6655378字节
    • bug: 9793个样本, 2072609字节
    • gor: 14514个样本, 5989252字节
    • id: 654287个样本, 1100932403字节
    • jv: 72667个样本, 69774853字节
    • map_bms: 11832个样本, 5060989字节
    • min: 225858个样本, 116376870字节
    • ms: 346186个样本, 410443550字节
    • nia: 1650个样本, 1938121字节
    • su: 61494个样本, 47410439字节
    • tet: 1460个样本, 1447926字节
  • 下载大小: 1793103024字节
  • 数据集大小: 1790336308字节

配置: indowiki_dedup_id_only

  • 特征:
    • url: string
    • title: string
    • text: string
  • 分割:
    • train: 654287个样本, 1100932403字节
  • 下载大小: 1103131493字节
  • 数据集大小: 1100932403字节
搜集汇总
数据集介绍
main_image_url
构建方式
该数据集源自维基百科多语言语料库,经由HuggingFace上的Wikipedia数据集提取并进一步加工而成。构建流程首先对原始数据进行基于标题与文本内容的去重处理,以移除包含模板文本(如无信息或请求贡献的样板内容)的噪声条目。随后,对清洗后的标题与文本执行字符串匹配去重,剔除符号移除、HTML标签剥离及ASCII字符验证后仍重复的文本。最终形成涵盖13种印尼及周边地区语言的语料集合,包括印尼语、马来语、爪哇语、巽他语等,并提供去重前后的多个配置版本,以便研究者按需选用。
特点
该数据集最显著的特点在于其多语言覆盖与精细去重策略。它囊括了从使用人数众多的印尼语(超过65万条)到小众语言如尼亚斯语(仅1651条)的广泛语种,为低资源语言的自然语言处理研究提供了宝贵资源。数据集的去重机制不仅过滤了冗余样板文本,还通过字符串匹配进一步消除语义重复,显著提升了语料质量。此外,数据以统一格式存储,每条记录包含URL、标题与文本字段,便于直接用于语言建模或掩码语言建模等任务。
使用方法
使用该数据集时,可通过HuggingFace的datasets库直接加载。用户可指定配置名称(如'indo_wiki_dedup_all')选择去重后的完整语料,或通过'indo_wiki_dedup_id_only'仅获取印尼语子集。若需按语言拆分,可提供语言代码(如'id')与时间戳参数,灵活抽取特定语种数据。对于希望复现构建流程的研究者,数据集仓库提供了完整的提取与去重脚本,只需配置Python环境并运行Shell命令即可从维基百科转储文件生成自定义版本,确保研究的可重复性。
背景与挑战
背景概述
在全球自然语言处理研究蓬勃发展的浪潮中,语言资源的丰富性与多样性成为推动模型性能提升的关键基石。印度尼西亚作为拥有超过700种语言的多元文化国度,其本土语言在数字化进程中长期面临语料匮乏的困境,阻碍了相关语言技术的进步。在此背景下,由研究者Sabil Makbar主导、依托HuggingFace平台发布的sabilmakbar/indo_wiki数据集应运而生,其创建时间可追溯至2023年。该数据集的核心研究问题在于系统性地整合印度尼西亚及其周边地区(如马来西亚、东帝汶)的13种语言与方言的维基百科语料,涵盖印尼语、爪哇语、巽他语等主流语言,亦囊括尼亚斯语、特图姆语等低资源语言。通过去重与清洗流程,数据集提供了约180万篇高质量文档,为语言建模、掩码填充等任务提供了标准化基准。其影响力不仅体现在填补了东南亚多语言语料库的空白,更为跨语言迁移学习与低资源语言模型研究提供了关键数据支撑。
当前挑战
该数据集所面临的挑战首先体现在领域问题层面:印度尼西亚语系中多数语言属于低资源语言,现有预训练模型在缺乏充足语料的情况下难以捕捉其独特的语法结构与词汇特征,导致文本生成与理解任务性能受限。此外,语言间的方言变体(如班尤马山语作为爪哇语的方言)加剧了模型对语言边界的辨识难度。在构建过程中,研究人员遭遇了多重技术挑战:其一,维基百科原始语料中存在大量模板文本(如无信息占位符与贡献邀请模板),需通过基于标题与文本内容的双重去重机制消除噪声;其二,跨语言数据的不平衡性显著,印尼语语料占据总规模的60%以上,而尼亚斯语与特图姆语样本不足两千条,如何在数据稀疏条件下保持模型泛化能力成为亟待解决的难题;其三,字符编码、HTML标签残留等预处理问题要求定制化的清洗流水线,以确保语料的纯净度与一致性。
常用场景
经典使用场景
在自然语言处理领域,sabilmakbar/indo_wiki数据集最为经典的使用场景在于为印度尼西亚语及当地多种地方语言提供大规模、高质量的无监督预训练语料。该数据集汇集了来自维基百科的亚齐语、巴厘语、爪哇语、巽他语、马来语等十余种语言文本,经过精心的去重与清洗处理,剔除了模板噪声与冗余内容,为语言模型的预训练奠定了坚实的数据基础。研究者可基于此数据集训练自回归或掩码语言模型,从而捕获印尼语系丰富的词汇、句法与语义特征,显著提升模型在该语系上的泛化能力。
解决学术问题
该数据集有效解决了低资源印尼地方语言在自然语言处理研究中长期面临的数据匮乏与质量参差不齐的学术难题。通过系统性地整合并去重来自维基百科的多语言文本,indo_wiki为语言建模、掩码语言建模等任务提供了规模可观且结构统一的基准语料,使得研究者能够针对亚齐语、尼亚斯语、特图姆语等资源稀缺的语言开展可靠的统计学习与模型评估。这一数据资源的开放,极大地推动了印尼语系多语言自然语言理解与生成研究的进展,为后续跨语言迁移学习与低资源语言技术突破奠定了方法论与实证基础。
衍生相关工作
该数据集衍生了多项具有影响力的经典研究工作,包括针对印尼语系的多语言预训练模型如IndoBERT及其变体,以及基于跨语言迁移学习的低资源语言建模方法。研究者常以indo_wiki作为基准语料,对比不同去重策略与数据增强技术对语言模型性能的影响,并在此基础上提出针对印尼语形态丰富性优化的分词与子词建模方案。此外,该数据集还催生了若干关于维基百科数据质量评估与清洗流水线的系统研究,为自然语言处理社区构建可复现、可扩展的多语言数据基础设施提供了重要参考范例。
以上内容由遇见数据集搜集并总结生成
二维码
社区交流群
二维码
科研交流群
商业服务