Bug: all of the data from lhq are from DeepMind-Mathematics, while no code/proof related samples are found. And there are duplicated samples
#4
by
HarryPot
- opened
Hi, Thanks for your great release.
I have been looking into the lhq subset, and found out that all of the samples are simple math from DeepMind-Mathematics.
I try to sample from the entire dataset dozens of times while not a single one is code/proof related sample.
Could you please check this?
Update: I can further confirm that the lhq subset includes only the DeepMind-Mathematics set.
I calculated the length of the content and order by descending and found out they are from the DeepMind-Mathematics set.
But according to the report, there should be some long context proof text.
Further, when counting with distinct rows, it shows 6.8M, not the 7.03M, indicating duplicate samples.
HarryPot
changed discussion title from
It seem all of the data from lhq subset are from the DeepMind-Mathematics
to Bug: all fo the data of lhq are from DeepMind-Mathematics, while no code/proof related samples are found
HarryPot
changed discussion title from
Bug: all fo the data of lhq are from DeepMind-Mathematics, while no code/proof related samples are found
to Bug: all of the data from lhq are from DeepMind-Mathematics, while no code/proof related samples are found. And there are duplicated samples