Skip to content

Dataset · bigscience

Multilingual prompt corpus for training BLOOMZ and mT0

bigscience/xP3all

Tens of millions of text prompts in 46 languages for training models on instruction-following in zero-shot mode.

What's inside

A large-scale text dataset ranging from 10 to 100 million entries, covering 46 languages and 16 different NLP tasks. This is a cross-lingual pool of prompts and datasets (xP3) created within the BigScience project. Well-known multilingual models BLOOMZ and mT0 were trained on this corpus.

What you can do with it

  • Instruction tuning (SFT) of multilingual language models for working with instructions in different languages
  • Development of zero-shot cross-lingual transfer skills, including low-resource languages
  • Comparison and reproduction of training pipelines for open models at the BLOOMZ level
Tasks
Other
Languages
Akan, Arabic, Assamese, Bambara, Bangla, Catalan, Code, English and 4 more
Size
10M–100M rows
Downloads this month
15.3K
Likes
33

How to load it

from datasets import load_dataset

ds = load_dataset("bigscience/xP3all")