Model description

byt5-base-zu-en-mt is a machine translation model from isiZulu to English based on a fine-tuned google/byt5-base model. It establishes a strong baseline for automatically translating texts from isiZulu to English.

Limitations and bias

This model is limited by its training dataset. This may not generalize well for all use cases in different domains.

Training data

Specifically, this model is a byt5-base model that was fine-tuned on JW300 isiZulu corpus and LAFAND. The model was trained using isiXhosa (xh_ZA) as the language since the pre-trained model does not initially support isiZulu. Thus, you need to use the xh_ZA for language code when evaluating the model.

Training procedure

This model was trained on NVIDIA V100 GPU

Eval results on Test set (BLEU score)

Fine-tuning byt5-base achieves 29.32 BLEU on LAFAND test set

BibTeX entry and citation info

By David Adelani

Downloads last month
Hosted inference API
This model can be loaded on the Inference API on-demand.