KlondikeDev's picture
Add ArithMark-3 and PIQA decontamination checks (all clean) with scripts and reports
7426416 verified
|
Raw History Blame Contribute Delete
1.02 kB

Decontamination

  • decontam_13gram.py: 13-gram check of every question against ARC-Easy/Challenge, HellaSwag, SciQ, OpenBookQA and WinoGrande evaluation splits. Log: report_13gram_other_benchmarks.txt; 0 rows flagged (the flagged-rows file it writes was empty). PIQA failed to load in this pass (see the log), hence piqa.py.
  • piqa.py: questions + traces vs all PIQA validation and test items (the baber/piqa copy lm-eval uses): 13/10/8-gram overlap. Output: piqa_report.txt (0, 0, 0).
  • check.py: questions + traces vs all 1,000 ArithMark-3 items: exact match, 13/10/8-gram overlap, and a loose numeric check (every number of the item appears in the question and the final answer equals the item's answer). Output: report.txt.
  • strict.py: of the loose numeric matches, keeps pairs with the identical set of numbers and the same answer, ranked by content-word overlap. Output: strict_report.txt.

Paths in the scripts point at the authors' machine; change them to your local copies of the data.