Datasets:
single jsonl file?
#5
by
ppbrown
- opened
The current format seems a bit odd to work with.
120 seperate parquet files?
The various CC12M sets provide a single jsonl.gz file
How about providing one of those?
@ppbrown I don't see any downside in us providing an additional jsonl.gz file, other than having another file format to maintain, but we're heads down on Public Diffusion at the moment (https://source.plus/public-diffusion-private-beta). It shouldn't be too bad to throw one together. I'll try to get one prepared when we update the dataset in the coming weeks. No promises it would be a single file though - we're hoping on releasing more of the metadata so it depends on what the final file size ends up being.
To me, the important stuff is just:
- the url
- the caption
- the width and height
anything else is optional :)