# Kuku Yalanji v12.0 Diagnostic Model

This is the strongest current Bible draft model by heldout chrF, but it is not a
faithful or approved Kuku Yalanji translator.

Version:

```text
v12.0-tagged-direct-plus-reference-bible-gvn-token-4096row-25epoch-batch16
```

Model type:

```text
NLLB seq2seq model with a custom gvn_Latn target token initialized from tpi_Latn
```

Result summary:

```text
train direct/ref chrF: 97.31 / 97.34
heldout direct/ref chrF: 44.37 / 44.36
heldout exact: 0
```

Use:

```text
draft Bible-domain generation and diagnostics only
```

Do not use this model as the authoritative source for approved Bible text or
known dictionary usage examples. Those should remain retrieval-first.

Main artifacts:

```text
merged/
adapter/
model_manifest.json
post_eval_analysis.json
eval_heldout_direct325_predictions_greedy.json
eval_heldout_ref325_predictions_greedy.json
```

The public artifact directory uses symlinks to the archived RunPod result under
the donto resources training archive, so the large weights are stored once.
