MobTranslate Wajarri v2
Release: v2.0-context-guarded-nr4-s17-context-recovery-r16-s20-20260731
Direction: English (eng_Latn) to Wajarri (wbv_Latn)
Runtime: exact Wajarri v1 project base plus this dynamic PEFT LoRA adapter
Status: evaluation-gated noncommercial research artifact; free-form translation is not validated
Claim boundary
This release passed the frozen closed-set reconstruction, held-out controlled-synthetic noninferiority, retention, and mechanical-degeneration gates declared by the v2 run. It did not pass a natural-language translation gate. There is no independent, speaker-diverse natural final test or fluent-speaker panel. Controlled synthetic rows are grammar- and lexeme-bound test material, not speaker-attested Wajarri.
Known dictionary queries should use deterministic dictionary lookup. The model route must label sentence output as an unverified research preview. A high dictionary reconstruction score does not authorize a general translation-accuracy claim.
Immutable identities
| Artifact | Identity |
|---|---|
| Model repository | ajaxdavis/mobtranslate-wajarri-v2 |
| Release tag | v2.0-context-guarded-nr4-s17-context-recovery-r16-s20-20260731 |
| Selected checkpoint | context_recovery_r16_step_20_nr4 |
| Base repository | ajaxdavis/mobtranslate-wajarri-v1 |
| Exact base revision used | ddae6103913b7cab299d687bee2dc7904fd59459 |
| Existing v1 release tag | v1.0-lcs-s800-20260730 -> 284f5745d052d6ed4882cf8a5be9be447ec3e289 |
| Base weight SHA-256 | 41ea844f30d6af1f2761d71126eb66a6d47c84c3294538bb1851afcd5043fe0e |
| Adapter weight SHA-256 | 05e6c40172d93b99538b3a21a04bbbd14a6c767180f85b83d5f00064a6858fb4 |
| Frozen decoder policy | beam 1, no-repeat 4 |
| Source dataset manifest SHA-256 | 4eb14fa09a31a37574f36e83cf02b4b288eb21f07f497545fda4704980e62a68 |
| Schedule manifest SHA-256 | fa7837ffa414c7e5eea3f7e7ddab3503ad125e017b19027060d59c6d470dd718 |
| Presentation schedule SHA-256 | c8175edf59ab59c7d692a86ad40b2cdff5d92681d34514d60f1223b6abdcdd63 |
The base is not stock NLLB. It contains the frozen wbv_Latn, <lexeme>,
<translate>, and reserved <glossary> tokenizer/model rows. Loading this adapter
against stock facebook/nllb-200-distilled-600M is incompatible.
Data accounting
| Quantity | Count |
|---|---|
| Source-preserving dictionary senses | 1,684 |
| Corpus rows available before experiment selection | 3,569 |
| Unique rows selected for this trajectory | 3,093 |
| Scheduled presentations in this trajectory | 3,712 |
| New controlled synthetic pairs | 157 |
| Total unique controlled synthetic pairs | 211 |
| Held-out controlled synthetic rows | 36 |
| Historical witness rows | 29 |
The 3,712 schedule entries are repeated, token-accounted presentations, not that many independent bilingual observations. The public dataset contains redistributable synthetic, historical, and open lexical material. The exact full training schedule is stored in a separate private Hugging Face dataset because its source terms differ.
Checkpoint census
| Candidate | One-target direct exact | Context exact | Synthetic chrF++ | Retention exact | Mechanical faults |
|---|---|---|---|---|---|
| v1_baseline | 6.45% | 7.01% | 72.53 | 100.00% | 10 |
| context_recovery_r16_step_20_nr4 | 91.26% | 52.14% | 70.90 | 100.00% | 0 |
Baseline synthetic chrF++ was
72.5317; the selected result was
70.9009. The serving-policy
audit changed 2 of
3188 frozen outputs while
holding the weights fixed. Full row-level
predictions and failure classes remain in the mounted experiment record.
Independent release gates
| Gate | Result | Frozen criterion |
|---|---|---|
closed_lexical_reconstruction |
PASS | one-target direct exact rate >= 0.80 with zero mechanical faults across all 3,188 frozen rows |
natural_sentence_generation |
FAIL | independent fluent-speaker review over multiple natural speaker/text clusters |
retention_noninferiority |
PASS | retention exact rate no more than 0.05 below v1 baseline |
synthetic_sentence_noninferiority |
PASS | held-out synthetic chrF2 no more than 2 points below v1 baseline |
natural_sentence_generation remains failed by design. It cannot be averaged away by
the passing reconstruction and retention endpoints.
Assemble the runtime
mkdir -p wajarri-v2
hf download ajaxdavis/mobtranslate-wajarri-v1 --revision ddae6103913b7cab299d687bee2dc7904fd59459 --include 'base/*' 'SHA256SUMS' --local-dir wajarri-v2
hf download ajaxdavis/mobtranslate-wajarri-v2 --revision v2.0-context-guarded-nr4-s17-context-recovery-r16-s20-20260731 --include 'adapter/*' 'ADAPTER-SHA256SUMS' 'release.json' --local-dir wajarri-v2
cd wajarri-v2
grep ' base/' SHA256SUMS | sha256sum -c -
sha256sum -c ADAPTER-SHA256SUMS
Load base/ with Transformers, then attach adapter/ with PEFT. Set
forced_bos_token_id to tokenizer ID 256204 (wbv_Latn). The trained request
prefixes are <lexeme> and <translate>. The frozen policy uses
num_beams=1, do_sample=false,
no_repeat_ngram_size=4,
repetition_penalty=1.0, and
length_penalty=1.0. Decoder selection is a serving
intervention over fixed weights; it is not additional linguistic learning.
Links
- Public dataset: https://huggingface.co/datasets/ajaxdavis/mobtranslate-wajarri-synthetic-corpus-v1
- Exact full schedule: https://huggingface.co/datasets/ajaxdavis/mobtranslate-wajarri-v2-full-training-private (private)
- Shared API Space: https://huggingface.co/spaces/ajaxdavis/alpha-v0-historic
- Reusable language-program playbook: https://mobtranslate.com/docs/operator-guide.html
The NLLB-derived model is subject to CC BY-NC 4.0. Individual data sources retain their own terms. No speaker, community, educational, legal, health, ceremonial, or authoritative approval is claimed.
- Downloads last month
- -
Model tree for ajaxdavis/mobtranslate-wajarri-v2
Base model
facebook/nllb-200-distilled-600M