Overview

This is a fine-tuned version of the model Helsinki-NLP/opus-mt-en-vi on the dataset IWSLT'15 English-Vietnamese. Performance in terms of sacrebleu on the test set is as follows:

• Original opus-mt-en-vi: 29.83
• Fine-tuned opus-mt-en-vi: 37.35

Parameters

• learning_rate=2e-5
• batch_size: 32
• weight_decay=0.01
• num_train_epochs=1

Thoughts

• Model Helsinki-NLP/opus-mt-en-vi is small (around 260MB), and can be easily deployed to a cheap server (e.g., EC2 t2.medium) without a GPU
• Easier and much faster to train compared to t5 or byt5.