Mapping to URLs

#3
by eclique - opened

Hi, thank you for uploading!
Is there a way to match this data to the original LAION URLs?

@eclique
I think it's enough to find it with the ID.

Is there any issue?

For now, it seems like there is no other method than the ID.

I was comparing it with this dataset and the IDs don't seem to match, unless I'm missing something?
https://huggingface.co/datasets/laion/laion400m

@eclique
I checked the laion/laion400m you shared, but there's no ID?

This data was created using parquet files that only include metadata as described in the URL and caption metadata dataset.

The ID for this data is the SAMPLE_ID mentioned in the laion blog.

If you want to match IDs and URLs, rather than looking in the laion/laion400m on the hub, try using the parquet files available on laion-400M on Kaggle.

When this dataset was created, parquet files hosted on the-eye.eu were used, but now the server is down so you won't be able to download those parquet files.

And here's the source code used for creating it at that time: HF_builders/src/Image/Laion400m.py

When recreating the laion400m-1 dataset, only these parquet files were used:

"part-00000-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00001-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00002-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00003-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00004-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00005-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00006-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"
"part-00007-5b54c5d5-bbcf-484d-a2ce-0d6f73df1a36-c000"

or oxen-ai parquet

jp1924 changed discussion status to closed

Sign up or log in to comment