agentlans/multilingual-dataset-index

Multilingual Dataset Index

A curated list of multilingual text datasets available on Huggingface, designed to help users easily find datasets by language—including those for low-resource languages.

Purpose

This index aims to make it easier to find datasets by language, addressing the common issue of inconsistent or unclear language codes across different datasets.

Format

The index is provided as a CSV file with the following columns:

• language: The English name of the language, as listed in the dataset configuration.
• dataset: The Huggingface identifier for the dataset.
• config: The specific configuration name within the dataset that corresponds to the language.
• language_code: A language code inferred from the config name, standardized to the ISO 639-3 format.

Limitations

• Different dialects or varieties of the same language may be listed separately (for example, Arabic, Standard Arabic, Egyptian Arabic, Moroccan Arabic).
• In other cases, one language can include multiple dialects, for example:
  • English includes American English and British English
  • French includes Metropolitan French and Quebec French
• Some languages with multiple scripts are grouped under the same language code, such as:
  • Hindi in Devanagari script vs. Romanized Hindi (Hinglish)
  • Simplified vs. Traditional Chinese
• Users should review the configuration names carefully to choose the right dataset for their needs.
• Automatic language detection may be inaccurate, particularly for low-resource languages. As a result, the actual number of high-quality languages in some datasets may be lower than it appears.
• This index includes only multilingual datasets (those that cover more than one language) and does not include datasets focused solely on a single language.
• Datasets may overlap. Be sure to remove duplicates whenever combining multiple datasets.