What does MIHENK measure?
MIHENK measures bilingual reasoning, careful reading, language mastery, and multi-step problem solving across Turkish and English questions in 20 disciplines.
MIHENK
Trunchbull portA bilingual Turkish-English reasoning benchmark spanning 20 disciplines, four difficulty tiers, and both multiple-choice and short-answer tasks.
80 caseschatCC-BY-4.0version 1.0.0 / 80 cases
Run on your modelsTry it now, no account needed
Runs in the standard chat lane with a pinned custom evaluator. No tools or sandbox are required.
What it measures
MIHENK tests whether a model can reason consistently in both Turkish and English rather than simply recall familiar facts. Its public development split moves across science, language, law, programming, economics, and other disciplines while increasing difficulty from L1 to L4.
MIHENK is a bilingual reasoning benchmark with localized Turkish and English questions across 20 disciplines. The public development split contains 40 Turkish and 40 English cases, evenly divided between multiple-choice and short-answer formats and across difficulty levels L1 through L4.
The Trunchbull port preserves the complete 80-case public split and the upstream response contract. Multiple-choice cases use answer-letter extraction; short answers use a seven-word limit, Turkish-aware normalization, accepted aliases, and numeric tolerance.
trunchbull port
Trunchbull packages the complete 80-case public development split with its bilingual metadata and pinned mixed-format reference scorer.
Cases
3 of 80 shown| id | case | runtime |
|---|---|---|
| mihenk-public/bilingual-reasoning | Turkish-English reasoningCompare reasoning performance across 40 Turkish and 40 English public cases. | chat |
| mihenk-public/mixed-formats | Mixed answer formatsEvaluate 40 multiple-choice and 40 constrained short-answer cases with upstream-compatible scoring. | chat |
| mihenk-public/difficulty-tiers | Difficulty tiers L1–L4Inspect balanced performance across four difficulty levels and 20 disciplines. | chat |
Tools
none requiredNo model tools. Cases are answered from the prompt and checked by the grader.
MIHENK, explained
Does this release include MIHENK's private holdout?
No. The Trunchbull release contains all 80 cases in the upstream public development split and does not include or reconstruct private holdout questions.