Model Card: NLLB-200 French-Wolof(🇫🇷↔️🇸🇳) Translation Model

Model Details

Model Description

A fine-tuned version of Meta's NLLB-200 (600M distilled) model specialized for French to Wolof translation. This model was trained to improve accessibility of content between French and Wolof languages.

• Developed by: Lahad
• Model type: Sequence-to-Sequence Translation Model
• Language(s): French (fr_Latn) ↔️ Wolof (wol_Latn)
• License: CC-BY-NC-4.0
• Finetuned from model: facebook/nllb-200-distilled-600M

Model Sources

• Repository: Hugging Face - Lahad/nllb200-francais-wolof
• GitHub: Fine-tuning NLLB-200 for French-Wolof

Uses

Direct Use

• Text translation between French and Wolof
• Content localization
• Language learning assistance
• Cross-cultural communication

Out-of-Scope Use

• Commercial use without proper licensing
• Translation of highly technical or specialized content
• Legal or medical document translation where professional human translation is required
• Real-time speech translation

Bias, Risks, and Limitations

1. Language Variety Limitations:

   • Limited coverage of regional Wolof dialects
   • May not handle cultural nuances effectively

2. Technical Limitations:

   • Maximum context window of 128 tokens
   • Reduced performance on technical/specialized content
   • May struggle with informal language and slang

3. Potential Biases:

   • Training data may reflect cultural biases
   • May perform better on standard/formal language

Recommendations

• Use for general communication and content translation
• Verify translations for critical communications
• Consider regional language variations
• Implement human review for sensitive content
• Test translations in intended context before deployment

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Lahad/nllb200-francais-wolof")
model = AutoModelForSeq2SeqLM.from_pretrained("Lahad/nllb200-francais-wolof")

# Translation function
def translate(text, max_length=128):
    inputs = tokenizer(
        text,
        max_length=max_length,
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    )
    
    outputs = model.generate(
        input_ids=inputs["input_ids"],
        attention_mask=inputs["attention_mask"],
        forced_bos_token_id=tokenizer.convert_tokens_to_ids("wol_Latn"),
        max_length=max_length
    )
    
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

Training Details

Training Data

• Dataset: galsenai/centralized_wolof_french_translation_data
• Split: 80% training, 20% testing
• Format: JSON pairs of French and Wolof translations

Training Procedure

Preprocessing

• Dynamic tokenization with padding
• Maximum sequence length: 128 tokens
• Source/target language tags: fr_Latn/wol_Latn

Training Hyperparameters

• Learning rate: 2e-5
• Batch size: 8 per device
• Training epochs: 3
• FP16 training: Enabled
• Evaluation strategy: Per epoch

Evaluation

Testing Data, Factors & Metrics

• Testing Data: 20% of dataset
• Metrics:
  • Cloud Provider:
• Evaluation Factors:
  • Translation accuracy
  • Semantic preservation
  • Grammar correctness

Environmental Impact

• Hardware Type: NVIDIA T4 GPU
• Hours used: 5
• Cloud Provider: [Not Specified]
• Compute Region: [Not Specified]
• Carbon Emitted: [Not Calculated]

Technical Specifications

Model Architecture and Objective

• Architecture: NLLB-200 (Distilled 600M version)
• Objective: Neural Machine Translation
• Parameters: 600M
• Context Window: 128 tokens

Compute Infrastructure

• Training Hardware: NVIDIA T4 GPU
• Training Time: 5 hours
• Software Framework: Hugging Face Transformers

Model Card Contact

For questions about this model, please create an issue on the model's Hugging Face repository.