WikiOriginQA / README.md
haneda-future's picture
Update README.md
861bb63 verified
|
Raw History Blame Contribute Delete
1.7 kB
metadata
task_categories:
  - question-answering
language:
  - af
  - ca
  - et
  - ja
  - ka
  - ps
  - sk
  - uk
tags:
  - wikipedia
  - wikidata
  - culture
  - qa
configs:
  - config_name: af
    data_files:
      - split: test
        path: data/af.jsonl
  - config_name: ca
    data_files:
      - split: test
        path: data/ca.jsonl
  - config_name: et
    data_files:
      - split: test
        path: data/et.jsonl
  - config_name: ja
    data_files:
      - split: test
        path: data/ja.jsonl
  - config_name: ka
    data_files:
      - split: test
        path: data/ka.jsonl
  - config_name: ps
    data_files:
      - split: test
        path: data/ps.jsonl
  - config_name: sk
    data_files:
      - split: test
        path: data/sk.jsonl
  - config_name: uk
    data_files:
      - split: test
        path: data/uk.jsonl
license: apache-2.0

WikiOriginQA

WikiOriginQA is an automatically generated QA dataset based on Wikipedia, designed to analyze cultural biases in LLMs. It covers eight languages: Afrikaans, Catalan, Estonian, Japanese, Slovak, Georgian, Pashto, and Ukrainian.

Dataset Structure

The dataset is organized as JSON Lines (.jsonl) files for each language. Each line represents a single QA record.

Languages

The dataset contains a total of 8 languages:

  • Afrikaans (af)
  • Catalan (ca)
  • Estonian (et)
  • Japanese (ja)
  • Slovak (sk)
  • Georgian (ka)
  • Pashto (ps)
  • Ukrainian (uk)

Data Fields

Each record contains the following keys:

  • qid: The Wikidata item ID associated with the concept.
  • id: The Wikipedia Article ID in the respective target language.
  • question: The English-translated the question.
  • original_question: The question written in its original language.
  • answers: A list of valid answers.