# Mi'gmaq NLLB-LoRA 1.0.0-rc1

This is a **noncommercial, research-only** English-to-Mi'gmaq candidate. It is
not speaker-reviewed, community-endorsed, or suitable as an authoritative
translation source.

## Scope and rights

- Training data: sentence examples from the Mi'gmaq/Mi'kmaq Online Talking
  Dictionary, contemporary Listuguj spelling, licensed CC BY-NC 4.0.
- Base model: `facebook/nllb-200-distilled-600M` at revision
  `f8d333a098d19b4fd9a8b18f94170487ad3f821d`, licensed CC BY-NC 4.0.
- Direction: English (`eng_Latn`) to a project-added Mi'gmaq token
  (`mic_Latn`). Mi'gmaq is not a built-in NLLB-200 language token.
- Audio was not used to train this text model.
- Commercial use is not permitted by the source licenses.

The `mic_Latn` embedding was numerically initialized from `tpi_Latn` before
training. That seed is a plumbing choice only; it makes no genealogical,
typological, dialect, or linguistic-equivalence claim.

## Contents

- `merged/`: standalone merged model and tokenizer used by the live research
  preview.
- `adapter/`: LoRA adapter and tokenizer.
- `recovery/checkpoint-1440/`: validation-selected resumable trainer state.
- `model_manifest.json`: exact training arguments, environment, source hashes,
  and model metadata.
- `test-final.json` and `test-final-analysis.json`: the one-shot frozen-test
  predictions and diagnostics.
- `decoding-selection.json`: validation-only decoding selection.
- `frozen-test-access.json`: exclusive frozen-test access ledger.
- `RUN-CONTRACT-v1.0.0-rc1.json`: complete run contract and decision record.
- `SHA256SUMS`: local/public artifact inventory.
- `migmaq-v1-rc1-provenance-20260712T1315Z.tar.gz`: immutable closing process bundle containing
  logs, reports, screenshots, run metadata, dictionary consolidation backups,
  postchecks, and living source documents. It contains 102 files and is
  5,820,145 bytes. SHA-256:
  `9db6b6a52259c4fbf459ed29ddbc34776b3935ca3011dd4c6bdd58c94f04aa59`.

The 2.25 GB merged-model archive was sealed after pod deletion and local model
verification, before the public deployment and dictionary-consolidation steps.
Its bundled run contract therefore records that earlier close point. The
top-level `RUN-CONTRACT-v1.0.0-rc1.json` and timestamped provenance archive are
the final deployment-state records.

## Evaluation ruling

The one-shot 742-row headword-disjoint frozen test produced chrF++ 21.43
(95% bootstrap interval 20.81–22.03), SacreBLEU 1.78, zero exact matches,
zero empty outputs, zero source copies, and zero repeated-trigram outputs.
These checks establish process integrity and absence of several mechanical
failure modes. They do **not** establish grammatical, semantic, pragmatic, or
cultural adequacy. Manual inspection found material lexical and content
failures. Qualified Mi'gmaq review is required before any broader promotion.

Frozen evaluation used CUDA bfloat16. The live host uses CPU float32. Beam
search can choose a different near-tied sequence across those numerical
backends, so live output is not promised to be byte-identical to archived GPU
predictions. The model version, weights, tokenizer, and decoding configuration
remain the same.

See the public model card and evaluation report at:

- https://mobtranslate.com/docs/migmaq-model-card.html
- https://mobtranslate.com/docs/migmaq-evaluation.html
- https://mobtranslate.com/docs/migmaq-model-training.html