ajaxdavis/mobtranslate-wajarri-v3-controlled

MobTranslate Wajarri v3 controlled adapter

This is a compact NLLB LoRA adapter for one deliberately bounded English-to-Wajarri task. It accepts a dictionary-resolved Wajarri subject through a nonlinguistic <copy> slot and generates one of six complete, source-supported predicate templates.

It is not a free-form Wajarri translator and has not passed an independent speaker-diverse natural-language gate. Known words must come from deterministic dictionary lookup. Inputs outside the six frozen constructions must be rejected or routed elsewhere.

Immutable identity

Field	Value
Model repository	ajaxdavis/mobtranslate-wajarri-v3-controlled
Version/tag	v3.0.0-controlled-subject-slot-s17-step40-20260802
Base repository	ajaxdavis/mobtranslate-wajarri-v1
Base revision	ddae6103913b7cab299d687bee2dc7904fd59459
Base weights SHA-256	41ea844f30d6af1f2761d71126eb66a6d47c84c3294538bb1851afcd5043fe0e
Adapter weights SHA-256	a4a91d4e6a6ba49509a8ab432c0674686c96b4611f58db8b63089f26707ed3d6
Winning confirmation seed	17
Adapter topology	rank-16 LoRA over q/k/v/out projections and fc1/fc2

Authorized task

The caller first resolves one unique English subject to an approved Wajarri dictionary surface. It then replaces the English subject with <copy>, generates with decoder prefix [decoder_start_token_id, wbv_Latn, <copy>], verifies the raw template exactly, and replaces the single output slot with the dictionary surface.

English construction	Required raw model output
The <copy> is coming towards the speaker.	<copy> yanajimanha.
The <copy> is going away from the speaker.	<copy> yanmanha.
The <copy> is running.	<copy> jamarnimanha.
The <copy> is sitting down.	<copy> nyinarangamanha.
The <copy> is sitting.	<copy> nyinamanha.
The <copy> is standing.	<copy> garrimanha.

Generation is greedy (num_beams=1), deterministic, with max_new_tokens=64, no_repeat_ngram_size=0, repetition_penalty=1.0, and length_penalty=1.0.

Evidence and limits

• Three independently seeded deterministic training runs must pass the 35/35 controlled route, batch-1/batch-16 invariance, retained fixed-utterance noninferiority, synthetic holdout noninferiority, zero mechanical faults, and distinct-weight gates.
• The published winner must reproduce all 35 controlled outputs byte-for-byte under A40 bfloat16 and CPU float32 execution.
• These are controlled template and runtime-parity findings. They do not establish broad sentence translation, naturalness, dialect suitability, or speaker approval.
• Closed dictionary reconstruction remains a research diagnostic; production dictionary answers should be returned from the database.

See docs/COMPLETE-TRAINING-EVALUATION-AND-HOSTING-GUIDE.md, release.json, and the files under evaluation/ and provenance/ for the complete machine-checkable record. The public training dataset is at https://huggingface.co/datasets/ajaxdavis/mobtranslate-wajarri-synthetic-corpus-v1.