OpenTWBench/tw-formosa-bench

Formosa Bench v2

2,665 four-option questions about Taiwan, in Traditional Chinese, answered closed-book: the reader is expected to know Taiwan, not to have read any particular document.

Where the rest of OpenTWBench measures the professional knowledge examined by Taiwan's national examinations, this measures the Taiwan those examinations never ask about: 蔣渭水's《臨床講義》, who designed 衛武營, why 林聰明砂鍋魚頭 switched to turkey, the MRT rules, the reservoirs, the night markets.

The successor to lianghsun/Formosa-bench (349 items, four categories).

Overview

Property	Value
Questions	2,665
Categories	13
Source articles	906
Format	4-choice single-answer (A/B/C/D)
Language	Traditional Chinese (zh-TW)
Framework	Twinkle Eval compatible
Licence	CC BY-SA 4.0, inherited from the corpus

Fields

The row shape is the Twinkle Eval MCQ contract — question, A–D, answer — with provenance alongside, which the harness ignores.

Field	Description
question	Question stem — self-contained, names its subject
A / B / C / D	Option texts
answer	Gold option letter
context	The passage the item was written from — evidence for the key, not part of the question
category	Taiwan.md top-level section
subcategory	Taiwan.md subsection
source_title	Title of the source article
source_path	Path within the Taiwan.md repository

Do not put context in the prompt. This is a closed-book benchmark; the passage travels with each item so the key can be audited against its source.

How it was built

Items were drafted by gpt-oss-120b, one per paragraph-aligned window of Taiwan.md (2,972 windows, median 1,119 characters, boundaries on paragraphs — never mid-sentence). The prompt asks for facts checkable outside the corpus and forbids document-local trivia: registration fees, edition numbers, internal counts, feature-retirement dates.

Filtering is where the quality comes from:

stage	items
generated	2,972
well-formed (format, Traditional Chinese, Taiwan usage)	2,939
keyed answer grounded in the source window (judge: gpt-oss-120b, which caught 100% of planted key defects in calibration)	2,683
after dropping corpus-meta categories, stems that lean on the passage, and near-duplicates	2,665

Automatic guards, each calibrated on real failures:

• Self-contained stems. A stem that opens on a pronoun (她取得美國國籍的 年份是？) names nobody once the source is gone — rejected. Pronouns inside a stem stay legal: 咖波的身體特徵中，牠有多少條腳 resolves itself.
• No passage references. 根據本文，... is scaffolding when the passage is supplied and nonsense when it is not. Bare scaffolding is stripped whole; a qualified opening (根據文中科普派的說法) keeps its qualifier, because 科普派 changes which figure is correct. Compound nouns are exempt — 道歉文中 is "in the apology", not a pointer.
• Option order is permuted from a hash of each item's source window id. The generator put the correct answer in option A 84% of the time; after permutation "always answer A" scores 24.9% and "pick the longest option" 26.6% (chance = 25%).
• Zero Simplified characters and zero mainland vocabulary (網絡/軟件/...), checked with a lexicon that fires zero times across 4.4M characters of Taiwanese prose and 255k examination questions.

Fairness note

gpt-oss-120b wrote and grounded these items, so its own score on this benchmark carries a self-advantage and should not be ranked against other models without that caveat. No other model in the OpenTWBench pool shares that advantage.

Limitations

• Factual accuracy is inherited from Taiwan.md — items are grounded in the corpus, not independently verified against primary sources. Where the corpus takes a scholarly position (e.g. crediting the 1580s Spanish record for "Formosa" rather than the popular Portuguese account), the key follows the corpus.
• Coverage follows the corpus and is uneven: People has 624 items, Politics 36.

Licence and attribution

CC BY-SA 4.0. Taiwan.md is CC BY-SA 4.0 and ShareAlike is contagious, so this dataset carries the same licence — unlike the examination suite, which is Apache 2.0 because examination papers are government works.

Source: Taiwan.md, CC BY-SA 4.0.

Citation

@misc{opentwbench2026,
  title  = {OpenTWBench: An Open Evaluation Suite for Taiwan-Domain Knowledge},
  author = {Huang, Liang-Hsun},
  year   = {2026},
  url    = {https://opentwbench.ai}
}