Formula-vs-gold census: 1,312 rows where MathQA's own annotated formula evaluates to a different option than the marked answer (37,297 rows)

#9
by nootxlm - opened

Every MathQA row carries three golds that have to agree: the annotated_formula (an operation program), the five options, and the correct letter. I evaluated the formula for all 37,297 rows in the data files this dataset's loading script serves, and matched the result against the options.

Two independent evaluators ran every row β€” a recursive-descent parser for the operation-program grammar, and a translation of the program into Python syntax evaluated by Python's own parser. A row is counted only when both evaluators independently land on exactly a different listed option than the marked one:

  • train (29,837 rows): 1,050 wrong-gold
  • validation (4,475 rows): 161 wrong-gold
  • test (2,985 rows): 101 wrong-gold
  • total: 1,312 rows where the dataset's own formula contradicts its marked answer

Examples anyone can check by hand:

  • train row 72: formula subtract(204, multiply(64, 3)) = 204 βˆ’ 192 = 12, which is option b β€” the marked answer is d (52).
  • validation row 24: formula multiply(2430, divide(1, 3)) = 810, option e β€” the marked answer is b (486).
  • test row 80: formula subtract(6, subtract(6, 6)) = 6, option d β€” the marked answer is b (4).

The full buckets, so the scope is plain: 18,820 rows agree; 14,465 rows' formula value matches no listed option (reported separately β€” rounding and unit shapes live there, and I am not claiming those as wrong gold); 531 formulas sit outside the implemented operation grammar; 1,078 rows carry ratio options and were excluded because a first-number match would be ambiguous; on 94 rows the two evaluators disagreed and those rows are excluded from every verdict above.

The checker is public and re-runs the whole census in one command: https://huggingface.co/spaces/nootxlm/formula-gold-check (script: formula_gold_check.py, exit 1 while any wrong-gold row stands). It takes the data zip URL as an argument, so the same check runs on any program-annotated math set served the same way.

If a row above is misread, the formula and the options are quoted verbatim β€” happy to correct the count with a dated note.

Thanks for the full-census check. Two independent evaluators agreeing row for row is what makes the 1,312 count hold up β€” a row counts only when both land on a different listed option than the marked one. Keeping the 14,465 no-match rows separate, rather than claiming them as wrong gold, is the discipline that makes the verdict clean, and the hand-checkable examples (train row 72: 204 βˆ’ 192 = 12, option b, marked d) let anyone verify the method in minutes. The public checker re-running the whole census in one command helps as well.

Sign up or log in to comment