We split the original xquad dataset into subsets.
There are 876/161/153 question-answer pairs from 34/7/7 articles for train/validation/test separately.
It is uploaded to Hugging Face as well.
You can easily load it with the datasets library by:
from datasets import load_dataset
dataset = load_dataset("zhufy/xquad_split")