[Important] Dataset splits not clearly shown in HuggingFace

[sgdgp]

Hi,

Its been a long time that this paper has been made public with an available leaderboad. However the Huggingface dataset page still does not provide access to the smaller splits that authors used for various experiments in their paper.

I strongly request the authors to make the smaller splits (smaller than 28k) public. If some specific rule was used to create the splits, please make that available instead.

[ChenZhongPu]

Yes. The dataset card said that there are 598 samples, but the actual size is 28k, where most columns are null. The core split with 598 samples are desired.

[code4DB]

@sgdgp @ChenZhongPu Sorry for the late response and thanks for your interests in our work. We would like to clarify the following two points.

1. Unlike benchmark such as BIRD that provides a small dev dataset. Our smaller dataset in PARROT is kept for the leaderboard, and we release the other two variants of our dataset with a large volume.
2. The column of the dataset in HF is null is due to the display format, where each column corresponds to a database dialect. For a single SQL, not all database dialects are written differently, and we remove the exactly the same one (finally turned to null).

Feel free to contact me if you have any questions :)