marian-finetuned-kde4-en-to-fr

This model is a fine-tuned version of Helsinki-NLP/opus-mt-en-fr on the kde4 dataset. It achieves the following results on the evaluation set:

• Loss: 0.0923

Model description

This model builds on the Helsinki-NLP/opus-mt-en-fr Marian MT model. Marian MT is a highly optimized neural machine translation framework built on Transformer architecture. Fine-tuning this model on the KDE4 dataset has enhanced its performance for translating sentences specific to software documentation.

Intended uses & limitations

Uses:

• Translating software documentation or user interfaces from English to French.
• Research purposes for domain-specific translation tasks.

Limitations:

• The model may not perform well on general-purpose text outside the KDE4 domain.
• Potential bias due to training on a limited domain (software documentation).
• Rare or technical terms may not be translated accurately.

Training and evaluation data

The model was fine-tuned on the KDE4 dataset, which consists of software-related documentation and user interface translations.

• Training Data: 80% of the dataset was used for training.
• Validation Data: 10% was used for validation during training.
• Test Data: 10% was held out for evaluation metrics like BLEU.

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
0.1362	1.0	3284	0.1102
0.1145	2.0	6568	0.0964
0.1056	3.0	9852	0.0923

Framework versions

• Transformers 4.46.2
• Pytorch 2.5.1+cu121
• Datasets 3.1.0
• Tokenizers 0.20.3