PharmaMCQ-Bench

968 four-option multiple-choice questions on pharmaceutical manufacturing and quality. The questions cover GMP, validation, CAPA, regulatory inspection, and pharmacopoeia topics. Each question comes from a knowledge unit that an LLM extracted from guidance documents and a domain expert validated.

Sources: BPOM CPOB (Indonesian GMP), WHO TRS, ICH Q-series, EMA, FDA Warning Letters, CAPA case studies, FDA software validation guidance, Six Sigma, USP/Ph.Eur., and FAERS.

Format

Field	Description
question	Question text in English. Indonesian regulatory terms are kept, with an English gloss.
choices	4 options
answer	Index (0–3) of the correct choice
answer_letter	A–D
source_type, source_title	Source document
fact_type, facet	Knowledge type and the facet the question tests
vignette	true for scenario questions (case facts stated in the stem)
distractor_origins	mined (a real model error) or synthetic for each distractor
evidence	Source excerpt that supports the answer

The correct answer appears equally often at each position (242 times each).

from datasets import load_dataset
ds = load_dataset("parquet", data_files={"test": "test.parquet"})["test"]

test.csv contains the same items in flat A/B/C/D columns.

Construction

1. Questions: Claude Opus 5.5 wrote 2 questions per unit, each testing a different facet. Case-study items were rewritten as scenario questions.
2. Mined distractors: Claude Haiku 4.5 and Sonnet 5.5 answered every question open-ended, and their wrong answers became candidate distractors (following arXiv:2405.11966).
3. Distractor drafting: Opus wrote realistic distractors from the same domain.
4. Judging: Claude Fable 5.1 judged every item against its evidence. An item was kept only if the gold answer is correct, the question is self-contained, and at least 3 distractors are clearly wrong.

1,342 questions were generated and 968 were kept. The full methodology and findings are in MCQ_METHODOLOGY.md in the source repo.

Limitations

• FAERS: only 15 items, because spontaneous-report associations rarely give clean wrong answers.
• Claude models: Claude models generated and judged the items, so report Claude models' scores separately.
• Expert review: a stratified 105-item sample is pending expert review.