Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

I turned a food-safety exam blueprint into a 580-item open dataset (and what the schema taught me)

Описание и применение

Создан набор из 580 вопросов по кулинарной безопасности, оформленный в JSONL с обязательными полями темы, задачи, ссылки и происхождения. Это позволило измерять покрытие тем, выявлять пробелы и автоматически подбирать вопросы для повторения.

Задача

Creating a defensible question bank for a 90‑question, scenario‑based food‑safety exam required more than writing 500 questions; each item needed traceability to a rule and source.

Результат

The dataset enabled measurable coverage, targeted drilling, and was reused by others; the trainer tracks mastery per domain, improving learning outcomes.

Рабочий процесс

  • 1. Designed a two‑level taxonomy (topic × task) to measure coverage.
  • 2. Added mandatory citation field (ref) and provenance flag (set).
  • 3. Aggregated items per topic×task, identified coverage gaps (e.g., 37% gap between top and bottom domains).
  • 4. Sorted cells by ascending count and wrote new questions starting from the emptiest cell.
  • 5. Wrote explanations that name the rule and the exam trap to reduce repeat mistakes.
  • 6. Exported data to JSONL and CSV, published on Zenodo under CC BY 4.0.
  • 7. Integrated dataset into the trainer app to track mastery per domain.

Доказательства

I built a free ServSafe Manager exam trainer — foodsafeprep.com — and the hardest part was never the frontend. It was producing a question bank that was defensible: every item traceable to a rule, every rule traceable to a source.

Источник

One JSON object per line.

Источник

I aggregated items per topic × task cell and got 21 cells across 580 items. The distribution was lopsided in a way I would never have guessed by reading through the questions.

Источник

The fix was mechanical: dump the counts to CSV, sort ascending, and write questions starting from the emptiest cell.

Источник

I published the whole set as an open dataset — 580 items, JSONL + CSV, CC BY 4.0 on Zenodo — mostly to force myself to make the schema honest.

Источник

  • Дата записи: 2026-10-09T14:41:49.686Z