| # Marian Fine-tuned English-French Translation Model | |
| ## Model Description | |
| This model is a fine-tuned version of `Helsinki-NLP/opus-mt-en-fr`, specifically trained for English to French translation. The base model was further trained on the `KDE4` dataset to improve translation quality for technical and software-related content. | |
| ## Model Training Details | |
| ### Training Dataset | |
| - **Dataset**: KDE4 Dataset (English-French parallel corpus) | |
| - **Split Distribution**: | |
| - Training set: 189,155 examples (90%) | |
| - Test set: 21,018 examples (10%) | |
| ### Training Configuration | |
| - **Base Model**: Helsinki-NLP/opus-mt-en-fr | |
| - **Training Arguments**: | |
| - Learning rate: 2e-5 | |
| - Batch size: 32 (training), 64 (evaluation) | |
| - Number of epochs: 10 | |
| - Weight decay: 0.01 | |
| - FP16 training enabled | |
| - Evaluation strategy: Before and after training | |
| - Checkpoint saving: Every epoch (maximum 3 saved) | |
| - Training device: GPU with mixed precision (fp16) | |
| ## Model Results | |
| ### Evaluation Metrics | |
| The model was evaluated using the BLEU score. The evaluation results before and after training are summarized in the table below: | |
| | **Stage** | **Eval Loss** | **BLEU Score** | | |
| |--------------------|---------------|----------------| | |
| | **Before Training** | 1.700 | 38.97 | | |
| | **After Training** | 0.796 | 54.96 | | |
| ### Training Loss | |
| The training loss decreased over the epochs, indicating that the model was learning effectively. The final training loss was approximately 0.710. | |
| ## Model Usage | |
| ```python | |
| from transformers import pipeline | |
| model_checkpoint = "Prikshit7766/marian-finetuned-kde4-en-to-fr" | |
| translator = pipeline("translation", model=model_checkpoint) | |
| translator("Default to expanded threads") | |
| ``` | |
| ### Example Output | |
| ```plaintext | |
| [{'translation_text': 'Par défaut, développer les fils de discussion'}] | |
| ``` | |