Datasets:
Cleaning scripts
Hey, thanks for the dataset. I found it very useful to have access to input and output token length :)
I was wondering if you could share the cleaning scripts? Basically, what you did to go from ccdv/govreport-summarization
dataset to this one!
Thanks 🤗
Glad you find it useful! I don't have the original script on hand, as it was on a machine that I had to wipe sadly.
IIRC, I used clean-text to clean it, pretty much just default settings with the exception of lower=False
. I then just dropped duplicates (if any) for summary/report, and then used the long-t5 tokenizer to get token lengths for every input report + summary, and filtered the dataset to only contain report
with 8192 tokens or less and 1024 tokens or less in summary
.
While it's not exactly the same as what was used for this, you can see a script with the general idea here. hope that helps!