Croissant Builder does not have the DatasetBuilder name

[stefanches7]

/!\ PLEASE INCLUDE THE FULL STACKTRACE AND CODE SNIPPET

Short description
Description of the bug.

Environment information

• Operating System: win 11 pro

• Python version: 3.12.9

• tensorflow-datasets/tfds-nightly version: 4.9.8

• tensorflow/tf-nightly version: 2.19.0

• Does the issue still exists with the last tfds-nightly package (pip install --upgrade tfds-nightly) ? yes

Reproduction instructions

import tensorflow_datasets as tfds
import mlcroissant as mlc


builder = tfds.dataset_builders.CroissantBuilder(
    jsonld="export-dara.jsonld"
 )
builder.download_and_prepare()

where export-dara.jsonld is link

If you share a colab, make sure to update the permissions to share it.

Link to logs

Expected behavior
Info of the dataset gets created

Additional context
Facing this issue with multiple dataset croissants on the web, e.g. https://www.scidb.cn/en/detail?dataSetId=067dfc6234b140ef9e3ba7b91c27684d# too

[ccl-core]

Hi @stefanches7 , thank you for raising this issue. IIUC, your croissant does not define any recordset: https://darus.uni-stuttgart.de/api/datasets/export?exporter=croissant&persistentId=doi%3A10.18419/DARUS-1431

The CroissantBuilder uses the recorsets ids to initialize the config names, so in this case as there is no recordset, you will have no configs defined:

datasets/tensorflow_datasets/core/dataset_builders/croissant_builder.py

Line 294 in 2034928

for record_set_id in record_set_ids

I think we could have a better error message there though, I'll send a PR for that soon.

[stefanches7]

@ccl-core and can we not avoid requiring RecordSets altogether?

[ccl-core]

Maybe there is something I miss here, but what would be the point of having a Croissant with no recordSets? RecordSets in Croissant are used to define collections of examples from your distribution. If you do not define any examples, what do you expect TFDS to do with your builder?

I understand that, in theory, you can use Croissant to describe the metadata-part of a dataset only. But in practice, TFDS needs the user to specify how to generate examples (which will then be prepared and stored when calling download_and_prepare). In the CroissantBuilder, we rely on mlcroissant to load examples from the recordsets.

[stefanches7]

@ccl-core naively me having not worked with TFDS much would expect TFDS to handle the metadata and download the data on the HDD for me, before I run my ML pipeline.

Maybe in case of absence of examples, we can take the random examples from the part of the datasets?
It's just way too sad I can not load the Croissant distribution at all in absence of examples, and it seems that data owners also might not consider having examples is a must when publishing their datasets.

[ccl-core]

"before I run my ML pipeline" -> but your ML pipeline runs some examples, i.e., in Croissant naming, records, right? How can you run the ML pipeline on the raw data -- unless the raw data is already structured, in which case you can easily define a recordset for it?

As I said above, I understand we can have a croissant without recordsets e.g. for storing metadata-only info; but I don't think the CroissantBuilder is the right choice in this case. IIUC correctly (please let me know if I don't!), you could easily use mlcroissant for that, without the need of TFDS?

[ccl-core]

Closing this bug for inactivity. Feel free to reopen it if the issue still persists :)