mapudungun-nllb-1.3B-arn-es-standard-bpe

Standard BPE: 1.3B, Mapudungun to Spanish

Fine-tuned facebook/nllb-200-distilled-1.3B checkpoint for Mapudungun to Spanish.

NLLB-200's built-in SentencePiece tokenizer, with no external presegmentation.

The corrected paper pipeline reports mean sentence-level chrF++ 45.56 for this condition, size, and direction. This card was generated from the repository's results/paper_stats.json source of truth.

Checkpoint provenance: corrected original-data retrain.

Inference

Download this repository and run:

python inference.py --model byumatrixlab/mapudungun-nllb-1.3B-arn-es-standard-bpe "text to translate"

The helper applies condition-specific Mapudungun preprocessing and removes @@ continuation markers from Mapudungun output. Transformers 4.x is required.

License and data

The base model is NLLB-200 and is distributed under CC-BY-NC 4.0. The fine-tuning corpus is derived from the Duan et al. Mapudungun corpus, distributed under CC BY-NC-SA 3.0. The card uses license: other so neither upstream obligation is hidden; users must comply with both sources' terms.

Citation

@inproceedings{thompson2026mapudungun,
  title     = {Bringing {Mapudungun} into the Modern {MT} Ecosystem: Morphology-Aware Tokenization for {NLLB}-200 Fine-Tuning},
  author    = {Thompson, Isaac M. and Rogers, Brandon M.A. and Ringger, Eric K.},
  booktitle = {Proceedings of the Sixth Workshop on NLP for Indigenous Languages of the Americas (AmericasNLP 2026)},
  year      = {2026}
}