ibahasa/indommlu-local-languages

IndoMMLU: Local Languages and Cultures (audited subset)

An audited, corrected subset of IndoMMLU (Koto et al., 2023) covering the 9 Local Languages and Cultures subjects: Indonesian primary and secondary school exam questions written in Balinese, Banjarese, Dayak Ngaju, Javanese, Lampung, Madurese, Makassarese, and Sundanese, plus one culture-knowledge subject on Minangkabau customs (answered in standard Indonesian). This is not a dataset we created. It is IndoMMLU's own subset, isolated, checked, and corrected. IndoMMLU's compilation of thousands of real Indonesian school exam questions across nine under-resourced languages is the hard part, and the only reason an audit like this one had anything to work with.

Dataset Description

IndoMMLU compiled 14,981 multiple-choice questions from Indonesian school exams spanning many subjects. 3,700 of those, 24.7%, cover the nine local languages and cultures subjects above. This repository publishes that subset on its own, after two rounds of audit that the original release did not include: a full structural scan of all 3,700 rows (checking option count, key validity, and duplicate option text) and 9 AI reviewers, one per language, reading every row for content problems (corrupted strings, wrong keys, misclassified language) with a human deciding what to fix and what to withhold.

The result is 3,624 rows judged clean or fixed, and 58 rows withheld because they need a native speaker to resolve, not because they were guessed at and hidden.

Use Cases

Evaluating how well language models understand Indonesian regional languages and dialectology, building a smaller and cleaner alternative to downloading the full 14,981-row IndoMMLU just to isolate the local-language subjects, and studying the kind of school-exam knowledge probes that generic multilingual benchmarks rarely cover.

Dataset Structure

full (3,624 rows)

One row per question. Columns:

Column	Meaning
subjek	one of the 9 subjects, human readable (Bahasa Jawa, Budaya Alam Minangkabau, etc.)
id	original IndoMMLU row id
jenjang	school level, derived from the numeric kelas column rather than the source's own level text, which sometimes reads as a misleading merged range like "SD-SMP-SMA"
kelas	numeric grade
soal	the question
opsi_a .. opsi_e	the up to five answer options, empty where a question has fewer
opsi	the same five options as one list column, always five slots, null where empty, for languages/tools that prefer a list over five separate columns
kunci	the correct option letter
kunci_idx	0-based index of kunci into opsi (A=0 .. E=4), for scripts that index by position instead of by letter
kunci_teks	the correct option's text
opsi_duplikat	ya/tidak, whether this row was flagged for having two options with identical text
status_qc	bersih (no issue found) or diperbaiki (a defect was fixed, see catatan_qc)
catatan_qc	empty when bersih, otherwise the specific defect and how it was fixed
sumber	source URL the original exam question was collected from
keterangan_sumber	short free-text note about the source (exam name, semester, etc.)

Breakdown by subject: Sunda 1,152, Jawa 987, Bali 467, Madura 288, Budaya Alam Minangkabau 195, Makassar 156, Lampung 138, Banjar 136, Dayak Ngaju 105. By level: SD 1,958, SMP 926, SMA 740. 32 rows (0.9%) carry opsi_duplikat = ya.

held_for_native_speaker (58 rows)

Same columns as full. These rows were pulled out during the audit because the reviewer found something that needs a native speaker to resolve rather than a guess, for example a term that does not match the vocabulary used consistently elsewhere in the same file for the same concept. catatan_qc explains the specific concern for every row, not a generic placeholder. This config exists so the concern is visible and checkable, not silently dropped.

Dataset Creation

IndoMMLU itself compiled these questions from Indonesian school exam materials circulating online (the sumber column names the page each question was pulled from). This repository did not collect new exam questions. What it added:

1. A full structural scan, not a sample, of all 3,700 rows in the original subset: option count below the minimum, keys outside A to E, keys pointing at an empty option, and options with identical text.
2. Nine AI reviewers, one assigned per language, each reading every row in its language for content problems: corrupted strings (a common pattern was two options merged onto one line, likely an OCR artifact from the original scanned exam sheets), wrong answer keys judged against dictionary or contextual evidence, and questions misclassified into the wrong language subject.
3. A human decision on every flagged row: fix it with a documented reason, or withhold it, never guess silently.
4. A further formatting pass, on top of the three steps above, done in two rounds: soal used a fill-in-the-blank marker of runs of periods (., ...., .........., whatever the original exam happened to use) with no consistent length, and, separately, 907 rows (25%) used the single-glyph ellipsis character (…) instead, sometimes mixed with ASCII periods in the same sentence (both confirmed present in the original source, not introduced by this pass). Both were unified to a fixed five ASCII periods (.....) wherever they marked a blank. The single-glyph ellipsis was deliberately not kept or used as the target format: it means "and so on" in prose and would misrepresent a blank left for an answer. A handful of rows additionally had a blank split in two by a stray character, most likely a leftover from a stripped question number or layout artifact in the original source (for example "1. ....." losing its 1 but keeping the period, or an ellipsis and a period run separated by a space); five rows total were corrected individually, by row id, after manual inspection, not by a general pattern. 2,885 of the 3,624 full rows were touched across both rounds. Every change was diffed against the previous state before publishing, including random samples read in full by hand each round, specifically targeting the rows most likely to break (multi-line verse and the longest, most tangled runs of mixed characters). status_qc and catatan_qc reflect steps 1 to 3 above only and do not record this later formatting pass, since it changed presentation, not content or correctness.

18 rows were dropped entirely as unsalvageable. 71 were fixed with the fix documented in catatan_qc. 58 were withheld into the separate config above.

Duplicate-option detection is a floor, not an exact count. The 32 opsi_duplikat = ya rows are what an exact-text-match scan caught. Manual reading during the audit found at least one case that scan missed: two options merged onto a single line by a scraping or OCR artifact from the source, which reads as one garbled string rather than two identical ones. Anyone doing further quality work on this dataset should not treat 32 as the full extent of structural issues.

Bias, Risks, and Limitations

Public and unprotected against training contamination since October 2023. IndoMMLU's answer keys have been openly downloadable since release. There is no way to tell which models have already trained on this data, unlike a held-out benchmark slice. Any score computed from this subset measures a mix of genuine language competence and memorization, and that mix cannot be separated after the fact. State this plainly wherever a score from this data is reported, not as a footnote.

Not verified by a native speaker of any of the eight languages. The audit was a full structural scan plus a 120-row stratified human sample, plus nine AI reviewers per language, described above. That is real coverage, but it is not the same as a native speaker confirming every answer key is semantically correct. The 58 held_for_native_speaker rows are the cases the audit was confident enough to flag explicitly; there may be others it missed.

Two layers of rights, not one. The CC BY-NC-SA 4.0 license below covers how IndoMMLU's authors compiled and distribute this collection. It does not by itself resolve the copyright status of the original exam questions themselves, which were gathered from third-party exam-bank sites. IndoMMLU's own paper addresses exactly this in its unnumbered Ethical Considerations section, near the end (see the paper): the dataset was collected from publicly available web resources under Indonesian Copyright Law No. 28 of 2014, Article 44, which permits use of a work for research and education when the source is fully cited. This subset keeps that same basis intact, republishing for the same research purpose with the source URL preserved in the sumber column for every row and full citation below. IndoMMLU has also been publicly hosted this way since 2023 without apparent objection, and this subset does not increase that exposure.

Level distribution is uneven across subjects. Sundanese has 718 SD questions alone, 19% of the whole subset, while Banjar, Dayak Ngaju, and Lampung have as few as 10 to 30 questions at the SMP or SMA level. This reflects what exists in the original exam-bank sources, not a sampling decision made here.

Minangkabau is a culture subject, not a language subject. Its questions are written in standard Indonesian and ask about customs, genealogy, and oral history (tambo), unlike the other eight subjects, whose questions are written in the regional language itself. Filter it out by subjek if your use case needs language production or comprehension specifically.

Licensing

CC BY-NC-SA 4.0. IndoMMLU's own GitHub repository and Hugging Face README both state this license in prose, clearly and consistently. The license metadata tag on the original indolem/IndoMMLU repository currently reads mit, most likely a template default left over from setting up the repository rather than a second, different license the authors intended. This card follows the license written out in the authors' own words, the same reading independently reached by SEACrowd/indommlu, an independent republication that also tags it cc-by-nc-sa-4.0.

• BY: attribution required, see Citation below.
• NC: non-commercial use only. This release is non-commercial today. If that changes, this dataset needs to be withdrawn or relicensed with the authors before continued use.
• SA: share-alike. Any further redistribution of this subset's questions must carry the same license.

Citation

@inproceedings{koto-etal-2023-indommlu,
    title = "Large Language Models Only Pass Primary School Exams in {I}ndonesia: {A} Comprehensive Test on {I}ndo{MMLU}",
    author = "Koto, Fajri and Aisyah, Nurul and Li, Haonan and Baldwin, Timothy",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    year = "2023",
    url = "https://arxiv.org/abs/2310.04928"
}

Also cite this repository for the audit and correction work, and link to it at the commit you used (this repository is git based and every revision is addressable).

Dataset Card Contact

Questions about the audit methodology, or a row that looks wrong: see ibahasa.com.