Parquet is fine and all, but I love the simplicity and simple interoperability of CSV.

You can save a huge amount of overhead just by base64 encoding the vectors, they aren't exactly human readable anyway.

I imagine the resulting file would only be approximately 33% larger than the pickle version.