ibahasa Research Data

Raw data behind ibahasa research.

Every figure we publish links back to a file here, pinned to the commit it was cited at. If you want to check a number, this is where you check it.

Benchmark data moved

Benchmark data left this repository on 2026-08-16 and is now maintained at ibahasa/benchmark-nusantara.

The two change at completely different speeds. An article pins its provenance link to a commit SHA so the reader sees the data exactly as it stood when the sentence was written, and a Zenodo release must not move. Benchmark data is the opposite: it changes every time a model is run.

Links published before that date still work. They are pinned to commit SHAs, and a commit permalink serves the tree as it stood at that commit rather than whatever main holds today.

The benchmark snapshots as they stood before the split remain archived in the v0.1.0 Zenodo deposit (10.5281/zenodo.21793396).

Layout

raw/<dataset>/
  MANIFEST.json     file list, byte sizes, sha256 of each file, export date
  <files>           the data itself
INDEX.md            table of every dataset in this repository

Start at INDEX.md.

Filenames are flattened from their original directory structure, so a dataset drawn from several folders cannot produce two files with the same name. The original path of each file is recorded in MANIFEST.json.

Citing a file

Use a permalink pinned to a commit, not to main:

https://github.com/ibahasa/research-data/blob/<commit-sha>/raw/<dataset>/<file>

A link to main shows whatever the file looks like today. A link to a commit shows what it looked like when it was cited, which is usually what you want when you are checking someone's arithmetic.

Each file's sha256 is in MANIFEST.json if you want to verify a download.

Citing this repository

A GitHub permalink is enough to check one file, but it is not a citable identifier on its own: it has no fixed version and nothing indexes it as a dataset. This repository is archived on Zenodo, which mints a permanent, version-pinned DOI for every GitHub Release.

• Cite the exact version you used (recommended, for reproducibility): 10.5281/zenodo.21982949 (v0.2.0)
• Cite "whatever is current" (a concept DOI that always resolves to the latest version): 10.5281/zenodo.21793395

@dataset{huda_2026_ibahasa_research_data,
  author    = {Huda, M. Khoirul},
  title     = {ibahasa Research Data},
  month     = aug,
  year      = 2026,
  publisher = {Zenodo},
  version   = {v0.2.0},
  doi       = {10.5281/zenodo.21982949},
  url       = {https://doi.org/10.5281/zenodo.21982949}
}

CITATION.cff in this repository carries the same identifiers and is read natively by GitHub's "Cite this repository" button on the repo's main page.

Held-out splits

Benchmark item sets are not published here. A portion of those items is held out so it stays outside training corpora and remains usable for measuring contamination later. Releasing them here would end that.

The releasable portion is prepared separately, through tooling that enforces the split rather than relying on anyone remembering it.

Before you cite these numbers

• Most datasets come from a single corpus over a single period. Another corpus would very likely give different figures.
• Verification behind several datasets was done by one person, with no inter-annotator agreement. Where a pass rate appears, it carries an uncertainty we have not measured.
• Benchmark scores are not stable at temperature 0. The same model has scored 75% and then 85% on identical items. Published figures use repeated runs and report variance; single-run numbers in a raw export should not be read as final.
• Some regional-language material covers one dialect variant only. Speakers elsewhere may judge differently, and that has not been tested.

None of this makes the data unusable. It makes it data, with edges. We would rather you knew where the edges are.

License

CC BY 4.0. Attribute to ibahasa.com and link back to the dataset or article you used.

Third-party sources keep their own terms and are attributed in the relevant MANIFEST.json and PROVENANCE files, notably the Leipzig Corpora Collection (CC BY 4.0) and Glottolog.

Contact

Methodology questions, corrections, or a number that looks wrong to you: mkhuda.com.

Corrections are welcome and get published. Several of our own findings are records of us being wrong.