FinanceRateCalc's picture
mirror refresh from data/hf-mirrors/denominator-reasoning at 16468eb8ca
998aa78 verified
|
Raw History Blame Contribute Delete
1.79 kB
metadata
license: cc-by-4.0
task_categories:
  - question-answering
language:
  - en
tags:
  - evaluation
  - reasoning
  - procedural
  - statistics
  - finance
pretty_name: Denominator Reasoning (procedurally generated)
size_categories:
  - n<1K

Denominator Reasoning: does a system understand what a rate is?

240 tasks, generated by a public script from a seed. Each gives the five outcome counts of a block of mortgage applications and asks for a denial rate under a stated denominator (60 tasks), under a denominator defined on the spot (60), converted from one denominator to another (60), or the inverse: which denominator produced a quoted rate (60; exactly one subset of outcomes fits).

Nothing to memorise. The counts are drawn around the shape of the real 2025 HMDA cells and perturbed by the seed; another seed gives another set of the same difficulty. This file is the public split (seed 2026). Scoring is arithmetic, no model: scripts/build_denominator_tasks.py::score.

Why. The same public file yields "FHA denial rates" from 10.6% to 37.4% depending on loan purpose and denominator, and answer engines quote them interchangeably. Recalling a published number is not the skill; applying a definition to counts is. These tasks isolate that skill from recall.

Limits. One narrow skill (ratios over stated subsets), English only, one domain's vocabulary. A perfect score shows the arithmetic and the reading of a definition, not understanding of mortgage lending. The counts are synthetic; no figure here is a statistic about any lender, place or year.

Generator and reference scorer: https://github.com/financeratecalc/financeratecalc.github.io/blob/main/scripts/build_denominator_tasks.py. Real cells: https://financeratecalc.com/denominator-map.html. Licence: CC BY 4.0.