En Vi

🚀 CLAck/en-vi: English-Vietnamese Translation Model

This project presents a fine-tuned MarianMT model for English-Vietnamese translation, trained in two phases on specific datasets.

🚀 Quick Start

This is a finetuning of a MarianMT pretrained on English-Chinese. The target language pair is English-Vietnamese. The first phase of training (mixed) is performed on a dataset containing both English-Chinese and English-Vietnamese sentences. The second phase of training (pure) is performed on a dataset containing only English-Vietnamese sentences.

💻 Usage Examples

Basic Usage

%%capture
!pip install transformers transformers[sentencepiece]

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
# Download the pretrained model for English-Vietnamese available on the hub
model = AutoModelForSeq2SeqLM.from_pretrained("CLAck/en-vi")

tokenizer = AutoTokenizer.from_pretrained("CLAck/en-vi")
# Download a tokenizer that can tokenize English since the model Tokenizer doesn't know anymore how to do it
# We used the one coming from the initial model
# This tokenizer is used to tokenize the input sentence
tokenizer_en = AutoTokenizer.from_pretrained('Helsinki-NLP/opus-mt-en-zh')
# These special tokens are needed to reproduce the original tokenizer
tokenizer_en.add_tokens(["<2zh>", "<2vi>"], special_tokens=True)

sentence = "The cat is on the table"
# This token is needed to identify the target language
input_sentence = "<2vi> " + sentence 
translated = model.generate(**tokenizer_en(input_sentence, return_tensors="pt", padding=True))
output_sentence = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]

📚 Documentation

Training results

MIXED

PURE

📄 License

This project is licensed under the Apache-2.0 license.

📦 Additional Information