I didn't know it would practically impossible to pull the dataset, lol
Corrupted tar headers and silent crashes every time I try to pull the data. Readme is poorly written as well. Incredibly ambigious. Good luck to ever attempts to work on this. Nothing about this project behaves as advertised.
I just cloned the repo and downloaded all the files with no issues.
Do you have a way to reproduce the "Corrupted tar headers and silent crashes" you're mentioning?
As for "Readme is poorly written as well. Incredibly ambigious." I would be happy to merge improvements to it.
The ambiguity is in the size of the dataset. The reader (me) is under the impression that it's just under a gig (~913 MB), and it's not even close to that. The full unzipped dataset is like 26 gigs. It's a laborious process to pull this data unless you are blessed with a machine that has 8+ cores at its disposal. It's silly to just pull the whole dataset and wish for the best. If I had known this, I would have pulled the first partition (1 out of 41) and then I could diagnose the best way to go about pulling this dataset. Who wants to sit there for an hour while the data is being pulled only for it to be corrupted? But I am ignorant to this dataset and huggingface, and I am operating under the assumption that the data I need to pull is ~913 MB.
So I am a little frustrated because I had some great ideas I wanted to explore, and I am under the impression that it's about 1 GB of data, and it's not. It's not even close to that. So then I am thinking, "Ok, I will try and manually add it to my project directory, it's just like my other machine learning projects," but then I check out the page to pull the data manually, and it's an entire scrolling bar's worth of data headers, and I am sitting here wondering, "Ok, is it all of this? Is it the last year's worth of data? Is it just the first year? etc."
A little hand-holding goes a long way. In the manual portion where all the data can be downloaded, why not add a description or a warning: "You need all of this data to do this project". Instead, it's a very confusing format where half the data was added 2 years ago, and the other half was added a year ago. I think logical people, who like to build assumptions from the ground up, would be confused by this setup. I certainly was. I know other programmers and machine learning practitioners are full-on "just jump off a cliff and work it out," but that's not me. I go over each step methodically.
I digress. Just delete this comment if you disagree.