Skip to content

Dataset · CohereLabs

Global-MMLU-Lite Multilingual Benchmark

CohereLabs/Global-MMLU-Lite

Compact version of the Global-MMLU benchmark by Cohere Labs across 23 languages for fast LLM evaluation.

What's inside

A text dataset of 10 to 100 thousand records, representing a lightweight version of the global MMLU benchmark. In version 3.0 (May 2026), it covers 23 languages including English, Arabic, Spanish, Czech, Hungarian, Slovak, and Tajik. The dataset is designed for fast and less resource-intensive evaluation of language model knowledge across different linguistic environments.

What you can do with it

  • Fast multilingual validation of LLM checkpoints during training
  • Comparing model knowledge in low-resource languages without running the full Global-MMLU
Languages
Arabic, Bangla, Czech, Welsh, English, Spanish, French, Hindi and 4 more
Size
10K–100K rows
Downloads this month
24.6K
Likes
45

How to load it

from datasets import load_dataset

ds = load_dataset("CohereLabs/Global-MMLU-Lite")