Datasets:
Upload full validated data
TLDR:
Is there a possibility to upload all the CV data on the dataset? Currently part of the validated files are not included since train, test, dev is only a subset of the whole validated dataset.
Details:
We noticed in the latest versions of the CV dataset the validated audios are not there. Since test train and dev is a subset of the validated, part of the data stays out.
It seems it was a historical decision to leave out the repeated text readings (same text multiple speakers) out of the training dataset, however the multiple recordings (and validation) of a text is currently allowed in the platform. (And honestly we don't think that the reasoning of Kelly Davis applies to the state-of-the-art models)
Because of this we created our own splits and want to upload them as a HF dataset. In order to do it correctly we want to call the data from the original HF dataset so that people will share their information with mozilla. However since there are files missing in the HF dataset, we cannot do this. We really appreciate if not for this release for the upcoming v17 you include all the validated files.
This impacts us most because we are the Catalan language community, which is currently the language with the most recording duration in v17.
Great idea! I can definitely add the validation files in the next release. It'd take a couple of days tho. Would that work?
Thank you very much! No problem at all, in any case we need to recalculate our splits for the v17 in the meantime. So it works perfectly.
Hi Vaibhav, any updates on the v17, will it be published in this organization? we see some unofficial datasets. (Edit: I see that you are working on it here i guess https://huggingface.co/datasets/reach-vb/common_voice_17)