Datasets:

OpenLLM-France

/

Claire-Dialogue-French-0.1

like 41

Follow

OpenLLM France 127

Added sample_by: paragraph in the YAML configuration so that the dataset is sampled by paragraph in the Viewer and also in load_dataset, instead of line by line

Thanks a lot @lhoestq !
I could not find out how to configure the viewer.
And your fix goes even further (it change the default behavior of the dataset, to encourage correct use)

Just one thing: can we please let the example code as it is?
even if "sample_by" is now not required with "load_dataset", it cannot be harmful to let it here (so that people can think of other ways of loading the dataset, speech turn by speech turn, for instance).

I could not find out how to configure the viewer.

We added the documentation about it quite recently ;)
https://huggingface.co/docs/hub/datasets-viewer

Just one thing: can we please let the example code as it is?
even if "sample_by" is now not required with "load_dataset", it cannot be harmful to let it here (so that people can think of other ways of loading the dataset, speech turn by speech turn, for instance).

Sure, I just reverted the code snippet change.
You can also document the other sampling strategies if you think they can be useful (possible values are "line", "paragraph" and "document")

Thanks a lot

I forgot to specify the data_files in the configs, sorry. Another PR here to fix this: https://huggingface.co/datasets/OpenLLM-France/Claire-Dialogue-French-0.1/discussions/3

@lhoestq
I tested after this fix (and having cleaned my ~/.cache/huggingface/datasets just in case):
If "sample_by" is not specified, it seems to iterates on lines (i.e. speech turns) not on paragraphs (i.e. full conversations):

In [1]: from datasets import load_dataset

In [2]: dataset = load_dataset("OpenLLM-France/Claire-Dialogue-French-0.1")

In [3]: len(dataset["train"])
Out[3]: 2934084

I expected it to be 36731 after your modification (I understood it would modify the default value of sample_by option):

In [4]: dataset = load_dataset("OpenLLM-France/Claire-Dialogue-French-0.1", sample_by="paragraph")

In [5]: len(dataset["train"])
Out[5]: 36731

Where is it wrong?

... and I have:

In [8]: datasets.__version__
Out[8]: '2.8.0'

Which version of datasets are you using ?
It seems to work fine on my side and in the Viewer

>>> load_dataset("OpenLLM-France/Claire-Dialogue-French-0.1")
DatasetDict({
    train: Dataset({
        features: ['text'],
        num_rows: 36731
    })
    test: Dataset({
        features: ['text'],
        num_rows: 284
    })
})

You should update datasets then ;)