README.md

metadata

license: mit
task_categories:
  - question-answering
language:
  - en
  - de
  - fr
  - zh
  - ur
pretty_name: CommonsenseQA Multilingual
configs:
  - config_name: en
    data_files:
      - split: train
        path: data/en/train.jsonl
      - split: validation
        path: data/en/validation.jsonl
  - config_name: de
    data_files:
      - split: train
        path: data/de/train.jsonl
      - split: validation
        path: data/de/validation.jsonl
  - config_name: fr
    data_files:
      - split: train
        path: data/fr/train.jsonl
      - split: validation
        path: data/fr/validation.jsonl
  - config_name: zh
    data_files:
      - split: train
        path: data/zh/train.jsonl
      - split: validation
        path: data/zh/validation.jsonl
  - config_name: ur
    data_files:
      - split: train
        path: data/ur/train.jsonl
      - split: validation
        path: data/ur/validation.jsonl
  - config_name: multi-lang
    data_files:
      - split: train
        path: data/multi-lang/train.jsonl
  - config_name: multi-lang-with-urdu
    data_files:
      - split: train
        path: data/multi-lang-with-urdu/train.jsonl

CommonsenseQA Multilingual

Multilingual version of CommonsenseQA with GPT-4o-mini/GPT-5-mini generated Chain-of-Thought explanations.

Languages

• English (en) - original
• German (de) - translated
• French (fr) - translated
• Chinese (zh) - translated
• Urdu (ur) - translated

Splits

Single Language Splits

• train: 8,196 examples per language
• validation: 1,221 examples per language

Multi-Language Splits

• multi-lang: 8,196 examples (4 languages: en, fr, de, zh - ~2,049 per language, non-overlapping)
• multi-lang-with-urdu: 8,195 examples (5 languages: en, fr, de, zh, ur - 1,639 per language, non-overlapping)

Important: In multi-language splits, each question appears in ONLY ONE language. There is no overlap.

Usage

from datasets import load_dataset

# Load a specific language
ds = load_dataset("alibashir/CommonsenseQA", "ur")

# Load multi-lang (4 languages)
ds = load_dataset("alibashir/CommonsenseQA", "multi-lang")

# Load multi-lang-with-urdu (5 languages)
ds = load_dataset("alibashir/CommonsenseQA", "multi-lang-with-urdu")

Fields

• question: The question with answer choices
• cot: Chain-of-thought explanation
• answer: Correct answer (A-E)
• lang: Language code (only in multi-lang splits)

Changelog

• v3: Added Urdu (ur) translations and multi-lang-with-urdu split
• v2: Removed extra conclusion sentences from translations
• v1: Initial release with en, de, fr, zh