|
--- |
|
language: |
|
- en |
|
- vi |
|
tags: |
|
- translation |
|
license: apache-2.0 |
|
datasets: |
|
- ALT |
|
metrics: |
|
- sacrebleu |
|
--- |
|
|
|
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. |
|
|
|
### Example |
|
``` |
|
%%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] |
|
``` |
|
|
|
### Training results |
|
|
|
MIXED |
|
|
|
| Epoch | Bleu | |
|
|:-----:|:-------:| |
|
| 1.0 | 26.2407 | |
|
| 2.0 | 32.6016 | |
|
| 3.0 | 35.4060 | |
|
| 4.0 | 36.6737 | |
|
| 5.0 | 37.3774 | |
|
|
|
|
|
PURE |
|
|
|
| Epoch | Bleu | |
|
|:-----:|:-------:| |
|
| 1.0 | 37.3169 | |
|
| 2.0 | 37.4407 | |
|
| 3.0 | 37.6696 | |
|
| 4.0 | 37.8765 | |
|
| 5.0 | 38.0105 | |
|
|
|
|
