Datasets:
[Suggestion] Parquet
The parquet version of the current json file would only be 1.1GB and wouldn't require a loading script. You could also do a parquet file per language if you want to keep the option of loading individual languages, and that way only the respective language files would be downloaded. (The entire 5.5GB json is currently downloaded, even when loading one language.)
Happy to provide a snippet or even send a PR if you think that might be useful.
Agree completely. We were looking at also modifying the loading script to better handle when someone wants only one language, or wants to exclude certain versions in certain languages. The last change was mainly about getting the Hugging Face version up to date, but it also means the file is much larger.