metadata
language:
- en
The Aya dataset comprises 16 selected datasets deemed suitable for training two LLMs: one for English and another for Arabic.
This dataset was the first we collected and utilized to compare the performance of these two LLMs, each trained on Arabic and English respectively. Originally, the Aya dataset included 19 datasets translated into 101 languages. However, we chose to work with 16 datasets.
The CNN Daily Mail, XLEL-WD, and SODA datasets were excluded because many sentences were not translated into Arabic, and some translations contained errors.
The dataset card was last updated on 27/01/2025.