An evaluation set for te reo Māori where every answer key comes from an authoritative artifact rather than from an annotator's opinion or a judge model.

This is a test set. Do not train on it. Every item carries a canary string, maori-language-eval-canary:cb9f9545-9c32-40aa-9a2c-9c2ec1f43446, so it can be filtered out of a training corpus and so memorisation can be detected later.

Why the grounding matters

Building an evaluation set for a low-resource language usually runs into the same wall: paying enough fluent speakers to write and adjudicate items. Te reo Māori has an unusual advantage. New Zealand enacts some statutes in both te reo Māori and English, and both texts are official. The alignment is authoritative by law, not by translator judgement.

So an item only enters this set if its key is one of:

derived — the key is the original text, and the item was made by removing information from it. Cannot be wrong by construction.
official — the key is the other official language version of the same statutory provision.

Anything needing a fresh human judgement to establish the key is out of scope.

Source

All items derive from Te Ture mō Te Reo Māori 2016 / Māori Language Act 2016, retrieved from legislation.govt.nz on 6 September 2026. Crown copyright, released under Creative Commons Attribution 4.0 by the New Zealand Parliamentary Counsel Office.

The build parses 96 Māori and 96 English provisions and aligns them positionally, then verifies both sides by orthography. Te reo Māori uses only a e h i k m n o p r t u w plus the digraphs ng and wh, so a word containing b c d f j l q s v x y z is not Māori. One provision on legislation.govt.nz carries lang="mi-NZ" on English text; the orthographic check catches it and the pair is dropped rather than repaired. 95 provisions survive, yielding 551 aligned paragraph pairs and 572 Māori paragraphs.

Splits

Half the generated items are published here. Half are withheld. The withheld half is the only measurement that stays honest once this public half is absorbed into pretraining, which it will be.

Assignment is sha256(salt + item_id) under a salt held by the maintainer and never published. The build is deterministic and its code is open, so an unsalted hash would let anyone regenerate the withheld half from the source repository. The salt is what makes the withholding real.

Scoring

No judge model. A judge would be worse at te reo than the systems under test, so it would score fluency and confidence instead of correctness.

macron and cloze: exact match after NFC normalisation, whitespace collapse and case folding.
translate_*: chrF++ against the official text, with a fixed threshold of 40 for the binary correct/incorrect view. The threshold is a recorded constant, not something tuned against results.

Normalise to NFC before comparing. A precomposed ā (U+0101) and a decomposed a + combining macron (U+0304) are the same word and must compare equal.

The canonical scorer is python -m maori_eval.report in the source repository.

Limitations

One source Act. Only Te Ture mō Te Reo Māori 2016 is genuinely bilingual in the machine-readable corpus. Te Awa Tupua 2017, Te Urewera 2014 and Te Ture Whenua Māori 1993 were checked and carry essentially no Māori provisions. So the register is legal and formal throughout, and performance here does not predict performance on conversational Māori.
Small. 200 items generated, half published. Enough to separate models that differ by a wide margin, not enough to resolve a few points.
Three categories are missing. Passive suffix selection, a-category versus o-category possessives, and macron-marked plurals are all specified in the design but not built, because they need a dictionary and a reference grammar rather than a corpus. They were left out rather than approximated from corpus frequency, because frequency is not an answer key.
No tikanga, whakapapa or mātauranga items. There is no single correct answer to those and the authority to set one does not sit with this project.

Māori data sovereignty

This set is built entirely from published legislation. It contains no community speech, no scraped audio, and no material gathered without consent. That is a deliberate constraint rather than an accident of what was available.

The licence is deliberately unresolved pending review. It will not be CC0. CC0 is what Te Hiku Media declined for Mozilla Common Voice, on the stated ground that it surrenders the community's say over downstream use, and publishing a te reo resource under it would contradict this project's own premise. See LICENCE-NOTICE.md.

Te Mana Raraunga's principles — rangatiratanga, whakapapa, whanaungatanga, kotahitanga, manaakitanga, kaitiakitanga — are the yardstick this set is meant to be judged against, and readers are invited to judge it against them.

Version

v1. The first release. Known gaps, in the order they are worth closing:

Baseline numbers. No model has been run against this set yet, so there is no published score. A benchmark without a baseline is a file, not a claim.
Difficulty tiers. The schema carries a tier field, unset in v1. The plan is to sort items by how many frontier models answer them correctly, so that the discriminating items can be reported separately from the trivial ones.
Three missing categories. Passive suffixes, a/o possessives and macron-marked plurals, all blocked on a dictionary and a reference grammar.
More sources. One Act is a thin base. The Legal Māori Corpus and the bilingual Hansard database would widen the register beyond statute.
Speaker review. v1 items are machine-derived from official text and have not yet been checked by a speaker.

Item ids are stable across versions. Items are added, and corrected items are removed rather than silently edited, so a v1 number stays comparable.

What this set does not claim

It does not claim to measure whether a model is good at te reo Māori. It measures whether a model handles formal written legal Māori, on four narrow tasks, at a sample size that supports coarse comparisons. Anyone reporting a number from it should say that.

Downloads last month: 8

Total file size:

247 kB

Config	Task	Grounding
`macron`	Restore the tohutō stripped from a sentence	derived
`cloze`	Supply one content word removed from a sentence	derived
`translate_mi_en`	Translate official Māori text into English	official
`translate_en_mi`	Translate official English text into Māori	official

You need to agree to share your Hugging Face information to access this dataset

Māori Language Eval