LaguQA

A benchmark for what a language model knows about 107 Indonesian regional and national songs. Every one was transcribed by hand from a printed songbook into ABC 2.1 notation, and every value traces back to a specific page.

The questions cover bibliographic facts such as composer and region of origin, and reasoning over the notation. Some show a fragment of number notation and ask which song it is. Others ask the model to count bars or name the highest note. Number notation is how music is taught across Indonesia, and no benchmark the author could find tests whether a model can read it.

Part of an undergraduate thesis in Informatics, Universitas Ahmad Dahlan.

Baca dalam bahasa Indonesia

Website	laguqa.github.io
Code	github.com/IRedDragonICY/LaguQA
Demo	IRedDragonICY/LaguQA-Demo
Model	IRedDragonICY/LaguQA-Gemma4-E2B
Mirror	kaggle.com/datasets/ireddragonicy/laguqa

The demo draws and plays each song's notation from the ABC files in this repository, so what a visitor hears is the same material the notation questions are keyed against.

Contents

file	rows	contents
laguqa.csv	107	the main table, one row per song
split.json		the frozen 70/37 split
daftar-isi.csv	107	the book's table of contents, for checking page numbers
laguqa_mc.jsonl	1200	multiple choice A-E, random guessing floor 20%; 25% in the time-signature category, where one option is a signature the book never uses
laguqa_test.jsonl	1002	test, full-regime pairs
laguqa_test_split37.jsonl	766	test, 37 songs never trained on
laguqa_train.jsonl	22046	training, all 107 songs (used for the released model)
laguqa_train_split70.jsonl	14658	training, 70 songs (used for the experiments)
laguqa_manifest.json		SHA-256 of the four open files
laguqa_mc_manifest.json		SHA-256 of the multiple-choice file
abc/	107	ABC 2.1 transcriptions, one file per song
SOURCE.md		where the data comes from, how it was verified, and its limits

There are 21 question categories, from metadata facts to reasoning over notation.

Score floor

Accuracy on this benchmark means nothing measured against zero. The book is dominated by 4/4 time and the key of Do = C, so a guesser that always answers the most common value already collects most of the metadata questions without knowing a single song.

The two test files have separate floors, and the numbers are not interchangeable, because the questions differ.

control	laguqa_test_split37	laguqa_test
always the most common answer	26.9%	33.6%
random within the same category	24.3%	28.6%
answering nothing at all	0.0%	0.0%

The numbers are lenient. All three can be regenerated with scripts/17_controls.py and are scored by the same scorer used on real models. For comparison, Gemma-4-E2B without training scores 19.6% on split37, below a guesser that knows no song at all.

Contamination check

The multiple-choice file carries a canary string on its first line. A model that can reproduce that string was trained on this file, which invalidates its score. A canary stays useful after the data is public.

Limitations

• 50 songs have an inferred time signature rather than a printed one. Pages in the 1 = C style print no time signature. The time_signature_source column marks them, and no time-signature question is generated from those songs.
• 28 ABC files are unverified, so notation-reasoning questions reach only 79 songs.
• Composer is filled for 52 of 107 rows, region of origin for 55 of 107. The book names a composer for national songs and a region for regional ones, almost never both. Those gaps are not missing data. The correct answer for such a row is that the book does not state it, and abstention questions exist to test exactly that.
• Eighteen words in the lyrics are still joined wrongly, because one syllable was transcribed capitalised, as in IndoNesia. The lyric keys for those words carry the wrong spelling, so a model that gets them right is marked wrong. The list is in docs/lirik-perlu-dicek.md.
• Two songs share a title. The book contains two different songs called "Desaku", with different melodies and composers. The title_unique column separates them, and is used only when a question has to say which one it means.

SOURCE.md, included here, has the details.

Licence

The metadata and transcriptions here are released under CC BY-NC 4.0. The source book remains fully copyrighted and its scans are not included. Indonesian Law 28/2014 Article 44 permits quotation for research and education as long as the source is named, and what is published here is designed to stay within that.

Two different layers of rights apply. The book is a compilation: what its publisher holds is the selection and arrangement of the contents, along with the page layout. Each song inside has its own rights status, separate from the book. That is why page scans are never published, while lyrics and melodies are treated according to each song's status.

The rights_class column records the class of each song, and HAK-CIPTA.md explains the basis along with how to file an objection. That file lays out the reasoning behind the classification. It is not legal advice, and it is not a claim that everything in this dataset is free to use.

Citation

Citing this dataset requires citing the source book as well.

@book{pustakabaru2025,
  author    = {{Tim Pustaka Baru}},
  title     = {Koleksi Lengkap Lagu-Lagu Daerah \& Wajib Nasional},
  publisher = {Pustaka Baru Press},
  address   = {Banguntapan, Bantul, Yogyakarta},
  year      = {2025},
  isbn      = {978-602-0874-22-7},
  pages     = {192}
}