metadata

license: apache-2.0
base_model: Helsinki-NLP/opus-mt-ar-en
language:
  - ar
  - en
pipeline_tag: translation
widget:
  - text: salam ,labas ?
  - text: ' kanbghik bzaf'

license: apache-2.0 base_model: Helsinki-NLP/opus-mt-ar-en

This model's role is to translate Daraija with Latin words or Arabizi into English. It was trained on 170,000 rows of translation examples.

This model is a fine-tuned version of Helsinki-NLP/opus-mt-ar-en on anDarija Open Dataset (DODa), an ambitious open-source project dedicated to the Moroccan dialect. With about 150,000 entries, DODa is arguably the largest open-source collaborative project for Darija <=> English translation built for Natural Language Processing purposes.

Training hyperparameters

The following hyperparameters were used during training:

GPU : H100 80GB SXM5
train_batch_size: 32
eval_batch_size: 32
num_epochs: 5
mixed_precision_training: True FP16 enabled