Full-census check: no duplicate questions, no conflicting gold (17,584 rows, both configs)

#29
by nootxlm - opened

A full-census grouping check on this dataset's parquet export, run today: every row of both configs (main and socratic), both splits — 17,584 rows in total.

Grouping each split by its question (checked two ways: byte-exact, and folded — whitespace-collapsed and lowercased) and comparing the final answer after the #### marker inside each group:

  • main train: 7,473 rows, 7,473 distinct questions, 0 conflicting answers
  • main test: 1,319 rows, 1,319 distinct questions, 0 conflicting answers
  • socratic train: 7,473 rows, 7,473 distinct questions, 0 conflicting answers
  • socratic test: 1,319 rows, 1,319 distinct questions, 0 conflicting answers

No question appears twice in any split, under either grouping, so there is no duplicate-question gold conflict to find — the answer key is internally consistent at full census, not just on a sample. Sharing this because duplicate-with-different-gold is a fault that does turn up in other widely used datasets, and a clean full-census result here is worth having on record for anyone choosing training or evaluation data. The grouping script is a small standalone file; happy to share it or adapt it to another split shape.

Sign up or log in to comment