Datasets:
fix-missing-configs
This PR aims to fix missing configurations introduced by my previous conversion to parquet files. Indeed, I only converted the clean configuration that was the default one.
To avoid such misunderstanding in the future, I updated the configuration description on the dataset card. Indeed, it was not reported that the validation and test splits in the given training configurations were all the same and duplicates of the validation and test configurations (same number of rows each time, I verified by checking manually for config clean_sa that confirmed it). Moreover, microset configuration was broken and now fixed, even when starting from the last commit before the parquet files conversion, due to wrong manifest file (clean_000000.json that did not correspond to clean_000000.tar which is used to build microset).