ajaxdavis/mobtranslate-anindilyakwa-v1

Anindilyakwa experimental scripture-domain baseline v0.2.0

This is a Traditional-Owner-authorized public experimental research preview with a failed automatic quality evaluation. It is not a reliable or fluent Anindilyakwa translator. Its purpose is to make the existing reproducible baseline available for transparent testing while preserving the evidence of its limitations.

Release decision

• Public model release: authorized by direct self-attested Traditional Owner direction
• Hugging Face Space inference: public experimental preview
• MobTranslate homepage inference: public experimental preview
• Public dataset: not created
• Qualified fluent-language review: 0 of 938 required role decisions complete
• Independent blind or escrowed final test: not present

The operator attested that they are a Traditional Owner, directed public publication, confirmed that all permissions were obtained, and accepted responsibility for the program. Identity and underlying permission documents were not independently verified by the agent. This direction authorizes publication; it does not establish linguistic quality or representative free-form competence.

Training identity

Item	Value
Run	anindilyakwa-private-baseline-20260809-a1
Base	facebook/nllb-200-distilled-600M at f8d333a098d19b4fd9a8b18f94170487ad3f821d
Direction	English (eng_Latn) to Anindilyakwa (aoi_Latn)
Target control	aoi_Latn, token ID 256204
Training rows	3,477
Development rows	643
Optimizer updates	330
Presentations	10,584 across all 3,477 training rows
Best development checkpoint	step 330, chrF2 7.7615
Model manifest SHA-256	7f570ac7b30d6005ab9cec77f60a72f9e66845ae27d373e3c05ae45f7fee5337
Merged weight SHA-256	d8d5d1a17976838cf15435e97b4ae882e9e1c128cd0ce515cb42fe863ab2cbc6

The corpus is a benchmark-filtered scripture-domain research edition. It is not representative of unrestricted conversation or general translation.

Mechanical verification

The training mechanism passed its reproducibility checks:

• 330/330 optimizer updates were observed.
• The selected aoi_Latn tied embedding/output row received non-zero gradients and changed.
• Unselected control rows (eng_Latn, <pad>, and </s>) did not change.
• All 72 LoRA-B parameters changed and remained finite.
• Adapter and merged tokenizer reload audits passed.
• The eight merged files independently match the training manifest.

Mechanical success is not translation quality.

Route-separated diagnostic failure

The frozen diagnostic evaluated 959 rows separately as lexical reconstruction, fixed utterances, and natural sentences. The test candidates were excluded from training, but they were not blind or independently escrowed and have no qualified fluent-language adjudication.

Route	Test rows	Exact	chrF2	High-repetition flags
Lexical	131	0	1.8236	131
Fixed utterance	14	0	2.4775	14
Natural sentence	60	0	2.3967	60

Across development and test, all 959 outputs triggered the high-repetition audit and none exactly matched its reference. The quality gate therefore remains failed even though publication is now explicitly authorized as an experimental preview. The full metrics are in evidence/route-evaluation-report.json; predictions remain in the canonical research archive and are not uploaded here.

Intended use

This repository is for public audit, reproducibility work, degeneration analysis, and experimental inference. Do not rely on its output for language learning, production translation, health, legal, ceremonial, educational-assessment, cultural-authority, or fluent-speaker substitution. Every output should be treated as an unverified model prediction that may be repetitive or wrong.

Licences and provenance

The NLLB base model is subject to CC BY-NC 4.0. Source-specific terms and the operator permission attestation also apply. No source corpus or dataset is distributed in this model repository.