Data Studio Error: Convert JSONL incorrectly

[zxccade]

Describe the bug

Hi there,

I uploaded a dataset here https://huggingface.co/datasets/V-STaR-Bench/V-STaR, but I found that Data Studio incorrectly convert the "bboxes" value for the whole dataset. Therefore, anyone who downloaded the dataset via the API would get the wrong "bboxes" value in the data file.

Could you help me address the issue?

Many thanks,

Steps to reproduce the bug

The JSONL file of V_STaR_test_release.jsonl has the correct values of every "bboxes" for each sample.

But in the Data Studio, we can see that the values of "bboxes" have changed, and load the dataset via API will also get the wrong values.

Expected behavior

Fix the bug to correctly download my dataset.

Environment info

• datasets version: 2.16.1
• Platform: Linux-5.14.0-427.22.1.el9_4.x86_64-x86_64-with-glibc2.34
• Python version: 3.10.16
• huggingface_hub version: 0.29.3
• PyArrow version: 19.0.0
• Pandas version: 2.2.3
• fsspec version: 2023.10.0

[lhoestq]

Hi ! Your JSONL file is incompatible with Arrow / Parquet. Indeed in Arrow / Parquet every dict should have the same keys, while in your dataset the bboxes have varying keys.

This causes the Data Studio to treat the bboxes as if each row was missing the keys from other rows.

Feel free to take a look at the docs on object segmentation to see how to format a dataset with bboxes: https://huggingface.co/docs/datasets/object_detection