Bangla MMLU

87,869 Bangla four-choice exam questions from Bangladeshi admissions, HSC and job tests; lm-eval scores the 14,750-item test split.

Also known as: BN MMLU, titulm-bangla-mmlu, TituLLM Bangla MMLU

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryknowledge
SubcategoryBangla four-choice exam questions from Bangladeshi admissions, HSC and job tests
Page statusactive
Metricaccuracy (acc and length-normalised acc_norm)
Directionhigher_is_better
Unit%
Dataset size87869
PublisherHishab Singapore Pte. Ltd., with University of Central Florida and Qatar Computing Research Institute

What it measures

Bangla MMLU asks a model to pick one of four options for a Bangla multiple-choice question drawn from Bangladeshi educational and professional exams. The TituLLMs paper describes it as inspired by English MMLU, not as a translation of Hendrycks's 57-subject set. Items were curated from open educational sites and textbooks covering university admission, higher secondary, job exams, medical admission and engineering admission. The skill is Bangla-medium school and professional knowledge, including Bangladesh-specific general knowledge.

Task format

Four-option multiple-choice QA in Bangla. lm-eval task `bangla_mmlu` formats the stem plus A–D options and scores by comparing log-likelihoods of the letters. Few-shot examples come from the 175-item `dev` split (`sampler: first_n`). Headline split is `test`.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub