soketlabs/bhasha-wiki-indic

Bhasha Wiki Indic

This dataset has Wikipedia articles pertaining to Indian context.

Dataset Details

Dataset Description

The dataset is built from Wikipedia articles taken from wikimedia/wikipedia. We filtered, cleaned and translated English articles related to India and Indian context out of entire dataset.

Each example has contents of a full cleaned wikipedia article and it's translations in 6 Indian languages.

• Curated by: Soket AI Labs
• Language(s) (NLP): [English, Hindi, Bengali, Gujarati, Tamil, Kannada, Urdu]
• License: [cc-by-sa-3.0]

Uses

The dataset is focussed on Indian factual content for pre-training LLMs where Indian knowledge and contextual understanding is required.

Dataset Structure

Total number of rows: 200820 It has approximately 1.56 billion tokens for all languages. The ratio for number of tokens for each language is roughly same when tokenized with our Indic tokenizer we created which can be found in our model repository Pragna-1b. Here are token counts for each language:

• English: 197.7 millions
• Hindi: 227.5 millions
• Bengali: 289.1 millions
• Gujarati: 206.2 millions
• Tamil: 233.8 millions
• Kannada: 203.5 millions
• Urdu: 207 millions

Each row corresponds to a wikipedia article with the decription of article in source language(english) and translations in 6 indian languages.
The title is in english and descriptions in different languages is represented by column name of format "language_code"_"script". Each description column in different languages is a list of sentences/multiple sentences and can be concatenated to get cleaned article decription.

Each row is of the format:

{'id': '1',
 'url': 'https://simple.wikipedia.org/sample_article',
 'title': 'Sample article',
 'eng_Latn': ['This is a sample...', 'and more information'],
 'hin_Deva': ['यह एक नमूना है'..., 'और अधिक जानकारी'],
 'kan_Knda': ['ಇದು ಒಂದು ಮಾದರಿ...', 'ಮತ್ತು ಹೆಚ್ಚಿನ ಮಾಹಿತಿ'],
 'ben_Beng': ['এটি একটি নমুনা...', 'এবং আরও তথ্য'],
 'guj_Gujr': ['આ એક નમૂનો છે...', 'અને વધુ માહિતી'],
 'tam_Taml': ['இது ஒரு மாதிரி...', 'மேலும் தகவல்'],
 'urd_Arab': ['...یہ ایک نمونہ ہے۔', 'اور مزید معلومات']
}

Dataset Creation

Curation Rationale

We needed to induce knowledge regarding India and Indian context while training our LLM, for which we gathered available Indic content data and also filtered factual data from Wikipedia.

Source Data

Wikpedia english articles from wikimedia/wikipedia

Data Collection and Processing

We filtered out Indian context data from wikimedia/wikipedia dataset's English articles by select keywords. Further we trained a few shot classification model to classify for Indian content vs Not Indian content to narrow down filtered English articles. We cleaned the articles and removed unwanted paragraphs for References etc. We then translated these artices to 6 Indian languages (Hindi, Bengali, Gujarati, Tamil, Kannada, Urdu) using AI4Bharat's IndicTrans2. The dataset has been cleaned and can be used for pre-training multilingual LLMs.

Recommendations

Though we tried to filter as much Indic context articles as possible with high Recall, there might be some non indic articles mixed in them as well.

Citation Information

@ONLINE{bhasha-wiki-indic,
    author = "Soket Labs Technology and Research Private Limited",
    title  = "Bhasha-Wiki-Indic",
    url    = "https://soket.ai"
}