GovReport Training set split

#3
by patrickocal - opened

Hi @Mivg and team,
can you please tell me more about how you (or others) split the data into training, eval and test? I have just created knowledge graphs of the inputs for all three (train, dev and test) and I find that the training set has many more relatively large documents. The test set gave me no memory issues. The validation set was closer to the training set. But the training set had monsters in there (250k, 100k+, plenty over 50k). Some of these gave me memory issues even with an 80gb gpu.
Just wondering why the asymmetry in sampling.
Cheers,
Patrick

Tel Aviv University org

Hi @patrickocal ,
In the SLED work, we did not create or modify any of the datasets, we simply created a different loader for convenience to be able to separate the prefix (e.g. question, query, instruction etc.) from the input. The datasets all remain as they appear in the SCROLLS paper (Shaham et al,.. 2022). Specifically, as mentioned in the SCROLLS paper, the GovReport dataset came from Huang et al. 2021 which provided the split.
As for your memory issues, you may try to clip the inputs to the longest sequence your HW can handle.
Best,
Maor

Perfect: thanks @Mivg . Thanks for the suggestion: I ended up generating the three most problematic KGs slowly on the cpu with 1500gb ram!

patrickocal changed discussion status to closed

Sign up or log in to comment