Fine-Tuned MarianMT Model for French to English Translation

This model is a fine-tuned version of Helsinki-NLP/opus-mt-fr-en, trained on a mixture of Tatoeba, OpenSubtitle, Europarl, NewsCommentary, NewsDocumentary, and CCMatrix (4.5 million samples).

Live demo

Click here (press restart to run the space)

• Then you input the french text that you want to translate into English.

• Hit submit and the model will output the translation.

Performance and Evaluation

BLEU Scores (higher is better)

• Purpose: Measures the precision of n-grams (word sequences) in the candidate translation compared to reference translations.
• Interpretation: Higher BLEU scores indicate better word-level overlap with the reference.
• Limitations: Does not account for synonyms or sentence structure.

Mix of every source includes Tatoeba, Europarl, NewsCommentary, OpenSubtitles, NewsDocumentary, and CCMatrix

ROUGE-L Scores (higher is better)

• Purpose: Measures the longest common subsequence between the translations and the references.
• Interpretation: Higher ROUGE-L scores indicate better preservation of meaning and content in the translation.
• Limitations: May not fully capture fluency or grammatical correctness.

Usage

from transformers import MarianMTModel, MarianTokenizer
import torch

# Load model and tokenizer
model_path = "nambn0321/NMT_opus_fr_en"
model = MarianMTModel.from_pretrained(model_path)
tokenizer = MarianTokenizer.from_pretrained(model_path)

# Move model to appropriate device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Define a function to translate English to French
def translate(text):
    if not text.strip():
        return "Please enter some text."

    # Tokenize input text
    inputs = tokenizer([text], return_tensors="pt", padding=True, truncation=True).to(device)

    # Generate translation
    with torch.no_grad():
        outputs = model.generate(**inputs)

    # Decode the generated output
    translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return translation

NOM