Adding the train split
@albertvillanova @lhoestq (as recent contributors to Helsinki-NLP datasets, please point me to someone else if needed),
Hi, to make Tatoeba easier to use for everyone, we would like to extend this dataset with the train split, and I am thinking of how to move this forward most elegantly.
The ideal outcome would be to have the dataset segmented into subsets and splits and searchable within Dataset Viewer (see my example dataset here), but this will require uploading the whole ~7TB collection to HF and breaking backward compatibility for the users of this dataset.
Alternatively, we can create a new Dataset and keep them both running for a while.
Thank you for any thoughts!
Hi ! sorry for the late reply
First I noticed the dataset is still based on a dataset loading script so it would be great to remove it and only have Parquet files instead (great for compression and filtering !)
There is a CLI tool to open a Pull Request and convert a script based dataset to Parquet without breaking backward compatibility
datasets-cli convert_to_parquet Helsinki-NLP/tatoeba_mt --trust_remote_code
The using datasets and push_to_hub() you can upload additional splits :)