quechua-chanka-mt-bpe

Small Transformer (≈11.5M params, d_model=256, 4 encoder/decoder layers) trained from scratch for Spanish→Quechua (Chanka) translation, using a standard 16k BPE tokenizer on the target (Quechua) side. Trained as the baseline in a controlled ablation against quechua-chanka-mt-prpe, which uses a morphologically-guided PRPE tokenizer instead.

Part of the extrinsic evaluation in QuechuaTok (Contreras, 2026), which introduces Morphological Boundary Accuracy (MorphAcc) as a tokenizer evaluation metric for agglutinative languages.

Results (test set, n=12,206)

Tokenizer	MorphAcc%	BLEU	chrF
BPE (this model)	6.67	4.18	35.13
PRPE	83.33	7.05	41.10

PRPE outperforms this baseline by +68.7% BLEU / +17.0% chrF, showing that morphologically-guided tokenization yields measurable downstream gains for Quechua NMT. See the paper for full methodology.

Training data

makitty/quechua-chanka-unified-v2, broad_train split (91k parallel sentence pairs).

Citation

@misc{contreras2026quechuatok,
  title={QuechuaTok: Morphological Boundary Accuracy as a Necessary Metric
         for Tokenizer Evaluation in Agglutinative Low-Resource Languages},
  author={Contreras, Maria},
  year={2026},
  eprint={2606.23943},
  archivePrefix={arXiv}
}