Training Details
Training Data
Dataset: Custom parallel corpus of English ↔ Wolof
Size: ~84,709 sentence pairs
Preprocessing:
Removed duplicates & missing values
Normalized (lowercased, stripped extra spaces)
Added language tag (>>wol<<) to English inputs
Split: 80% train, 10% validation, 10% test
Training Procedure
Base model: Helsinki-NLP/opus-mt-en-mul
Frameworks: Hugging Face Transformers, Datasets, PyTorch
Batch size: 8
Learning rate: 1e-5
Epochs: 5
Mixed precision: fp16
Speeds, Sizes, Times
Training device: Google Colab (GPU)
Dataset size: 1.3 MB (Excel file)
Model size: ~300MB after fine-tuning
Evaluation
Testing Data
Test set from the custom English–Wolof dataset (~8.4k sentences).
Metrics
BLEU (train): 74.76
BLEU (test): 76.12
Accuracy (exact matches): ~calculated but BLEU preferred.
Results
The model demonstrates strong translation performance for Wolof given the low-resource context, achieving a BLEU score > 76 on test data.
Environmental Impact
H…
