The order of the splits is not preserved

[albertvillanova]

We had a regression and the order of the splits is not preserved. They are alphabetically sorted, instead of preserving original "train", "validation", "test" order.

Check: In branch "main"

In [9]: dataset = load_dataset("adversarial_qa", '"adversarialQA")

In [10]: dataset
Out[10]: 
DatasetDict({
    test: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers', 'metadata'],
        num_rows: 3000
    })
    train: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers', 'metadata'],
        num_rows: 30000
    })
    validation: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers', 'metadata'],
        num_rows: 3000
    })
})

Before (2.15.0) it was:

DatasetDict({
    train: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers', 'metadata'],
        num_rows: 30000
    })
    validation: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers', 'metadata'],
        num_rows: 3000
    })
    test: Dataset({
        features: ['id', 'title', 'context', 'question', 'answers', 'metadata'],
        num_rows: 3000
    })
})

See issues:

• https://huggingface.co/datasets/adversarial_qa/discussions/3
• https://huggingface.co/datasets/beans/discussions/4

This is a regression because it was previously fixed. See:

• Split order is not preserved #6196
• The order of data split names is nondeterministic #5728

[albertvillanova]

After investigation, I think the issue was introduced by the use of the Parquet export:

• Use parquet export if possible #6448

I am proposing a fix.

CC: @lhoestq