CoolFace
Datasetpublic

ajaxdavis/mobtranslate-kuku-yalanji-synthetic-corpus-v2

MobTranslate Kuku Yalanji Synthetic Research Corpus v2 Complete public research release of 20,047 synthetic English-Kuku Yalanji sentence pairs plus the process evidence needed to inspect their production, review, revision, split, and use in the MobTranslate model program. Project-reviewed synthetic research material pending fluent-speaker and elder verification. It is not a speaker-certified dictionary or translation corpus. Identity ISO 639-3: gvn Glottocode:… See the full description on the dataset page: https://huggingface.co/datasets/ajaxdavis/mobtranslate-kuku-yalanji-synthetic-corpus-v2.

sourceHugging Faceotherupdated 2mo agoView on Hugging Face
1likes230downloads
Dataset Card

MobTranslate Kuku Yalanji Synthetic Research Corpus v2

Complete public research release of 20,047 synthetic English-Kuku Yalanji sentence pairs plus the process evidence needed to inspect their production, review, revision, split, and use in the MobTranslate model program.

Project-reviewed synthetic research material pending fluent-speaker and elder verification. It is not a speaker-certified dictionary or translation corpus.

Identity

  • —ISO 639-3: gvn
  • —Glottocode: kuku1273
  • —Project NLLB token: gvn_Latn
  • —Release: 2.0.0-research, 2026-07-11
  • —Sentence pairs: 20,047
  • —Governed v21.1 split: 16,642 train / 1,609 validation / 1,606 test, plus 190 quarantined rows

What is included

  • —data/sentences.jsonl and .csv: every pair with analysis, frame, lexical inventory, evidence, review state, rights state, authorship, and timestamps;
  • —data/dictionary.*: 2,688 source-preserving dictionary records;
  • —data/lexemes.*: 1,870 scholarly working-lexeme records with morphology, usage, evidence, and errata state;
  • —data/synthetic.db: immutable SQLite snapshot of corpus and process tables;
  • —data/reviews.jsonl, revisions.jsonl, lessons.jsonl, and process_events.jsonl;
  • —training/canonical-splits/ and leakage-audited training/v21.1-governed/;
  • —documentation/: operator guide, production examples, grammar cheatsheet, and dictionary errata;
  • —audits/: executable completion audit and machine-readable result;
  • —provenance/: generation, closure, linting, remediation, and export programs.

The original corpus README is preserved as CORPUS-README.md.

Recommended ML input

Use training/v21.1-governed/train.eng-gvn.jsonl for the governed treatment and keep validation, test, and quarantine files separate. Do not train on data/sentences.jsonl and then report results on an included held-out split; that would invalidate the evaluation through leakage.

python
from datasets import load_dataset

dataset = load_dataset(
    "ajaxdavis/mobtranslate-kuku-yalanji-synthetic-corpus-v2",
    "governed_v21_1",
)

Integrity

bash
sha256sum -c SHA256SUMS

The release builder verified SQLite integrity, foreign keys, the pinned dictionary hash, the 20,047-row count, and the canonical corpus-content digest before publication.

Current program decision (2026-07-17)

This corpus remains useful as a controlled retention and regression resource, but it is not evidence for generating another undifferentiated synthetic batch. The 20,047 pairs are correlated products of one documented generation and review process, not 20,047 independent speaker attestations.

The strongest internal v25 baseline misses only 31 of 2,724 directly trained lexical prompts; 26 misses are competing governed headwords rather than unknown forms. Later attested-clause, translated-reader LoRA, and paired OFT screens did not pass promotion gates. The next data program therefore prioritises source-aligned ordinary prose, independent speaker/text clusters, source-backed morphology, sense/variety adjudication, glossary uptake, and fluent review. It does not assign a sentence quota before rights, alignment, deduplication, variety identification, and bilingual review establish the eligible rows.

The dated research status and machine-readable next-corpus contract record the repeatable process intended for future language programs.

Data use and limitations

Read DATA_USE.md. No blanket public-domain, Creative Commons, or commercial license is asserted. The release does not relicense underlying community knowledge, dictionary material, grammatical descriptions, or third-party sources.

The corpus can contain errors in morphology, lexical choice, idiomaticity, pragmatics, register, dialect, cultural appropriateness, and English correspondence. A project database status such as verified means checked against the project's documented source/review process; it does not mean verified by a Kuku Yalanji speaker.

This package deliberately excludes elder-shared evaluation pairs, Bible evaluation controls, audio, personal account data, and model weights. The quarantine file is present for auditability and must not be merged silently into governed splits.

The exact gated v24.3 materialized schedule is ajaxdavis/mobtranslate-kuku-yalanji-v24-3-training, and the model is ajaxdavis/mobtranslate-kuku-yalanji-v24-3.

Citation

See CITATION.cff for the machine-readable citation record.

Project: <https://mobtranslate.com/> Training and hosting guide: <https://mobtranslate.com/docs/kuku-v24-3-model-guide.html>