Fine-tuned NLLB-200 Distilled for French-to-Wolof Translation

This model is a fine-tuned version of the NLLB-200 Distilled 600M model designed specifically for Wolof-to-French translation.

The fine-tuning process leveraged the methodologies and algorithmic insights presented in the paper "Advancing Wolof-French Sentence Translation: Comparative Analysis of Transformer-Based Models and Methodological Insights" by Kane et al., which provides a comprehensive analysis of transformer architecture adaptations for Wolof-French sentence translation.

Model Description

By adapting and fine-tuning NLLB-200 Distilled, a state-of-the-art multilingual sequence-to-sequence model, this model improves translation quality for the French-Wolof language pair, focusing on enhanced representation and transfer learning techniques suited to low-resource languages.

Intended Use

Translate text from Wolof to French.
Support NLP applications and research involving Wolof language processing, especially machine translation.
Enable further fine-tuning or transfer learning for similar African language pairs.

How to Use

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_name = "Oumar199/nllb_wolof_french"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

inputs = tokenizer("Jàmm nga am ? Ndax ganaar gi ñor na ?", return_tensors="pt")
outputs = model.generate(**inputs)
translation = tokenizer.decode(outputs, skip_special_tokens=True)

print(translation) # Expected output: Bonjour. la poule est-elle mûre ?

Limitations and Bias

Wolof is a low-resource language, so performance might vary with complex or uncommon phrases.
Translation quality depends on the similarity of input text to training distribution.
The choice of generation hyperparameters (such as temperature, number of beams, top-k, max length, etc.) can significantly influence output quality. For optimal performance, these parameters should be carefully fine-tuned for your specific use case or evaluation setting. Default values may not yield best results and experimentation is recommended.

Training Data

Fine-tuned on a curated parallel corpus of French-Wolof sentences assembled to capture diverse linguistic phenomena as outlined in Kane et al.

Training Procedure

Fine-tuning used the NLLB-200 Distilled architecture with hyperparameters and techniques inspired by the comparative analysis in the Kane et al. paper, optimizing for BLEU score and translation fluency.

Evaluation

Evaluated using BLEU and ROUGE scores against a held-out Wolof-French test set, demonstrating improved translation accuracy and quality compared to baseline NLLB-200 models without fine-tuning.

Model card created to ensure transparency, reproducibility, and assist the community interested in African language NLP.

Oumar199
/

nllb_wolof_french