Skip to content

Dataset · MBZUAI

3.4M instructions and responses across 52 languages

MBZUAI/Bactrian-X

3.4 million instruction-response pairs across 52 languages, translated from English alpaca-52k and dolly-15k.

What's inside

The dataset contains 3.4 million instruction-response pairs across 52 languages. 67k English instructions from alpaca-52k and dolly-15k were translated into 51 languages via the Google Translate API. Languages include Afrikaans, Arabic, Azerbaijani, Bengali, Czech, German, and others; the format is not specified in the card.

What you can do with it

  • SFT of a multilingual instruction model
  • LoRA fine-tuning for low-resource languages
  • Evaluating cross-lingual instruction-following
Tasks
Text generation
Languages
Afrikaans, Arabic, Azerbaijani, Bangla, Czech, German, English, Spanish and 4 more
Size
1M–10M rows
Downloads this month
2K
Likes
126

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("MBZUAI/Bactrian-X")