[feat req] Alphabetical ordering for splits in dataset viewer

[sanchit-gandhi]

Link

https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0

Description

Currently, the datasets splits for the viewer are displayed in a seemingly random order, see example for Common Voice 11:

It would be easier to traverse the list of possible splits if they were arranged alphabetically!

[severo]

WDYT @huggingface/datasets? We respect the order given by https://github.com/huggingface/datasets/blob/main/src/datasets/inspect.py#L382: is the order of the splits important or not? We generally expect train to be the first split, which would not occur with ["train", "test", "eval"] for example if they were ordered alphabetically.

[severo]

Note: the same question exists for the configs

[lhoestq]

The order doesn't matter - you can sort it if you want

[severo]

I wonder if we want to manage this on the datasets-server side or on the Hub. And I'm not sure what the strategy should be. For long lists, it seems a good idea to sort alphabetically. For ["train", "test", "eval"], I'm not sure if ["eval", "test", "train"] is better, as an ML process starts with training, etc. and it's surely why the splits in the dataset script were written in this order.

[polinaeterna]

maybe sort only configs? it can also be done directly in the script.

[severo]

maybe @julien-c or @gary149 want to chime in

[severo]

maybe sort only configs?

yes, it's a good idea @polinaeterna. The splits are generally only 5-6 max, while there are a lot of datasets with a high number of configurations. Alphabetically sorting seems more useful for lists with a large number of items.

[severo]

Done with #623

I manually refreshed the cache for https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/viewer/ab/train, but other datasets will remain with the old order until the cache is invalidated. I don't think it's worth refreshing all the cache just for this.