🌍 Language Translation Model

This repository hosts a fine-tuned T5-small-based model optimized for language translation. The model translates text between multiple languages, including English, Spanish, German, French, and Hindi.

📌 Model Details

Model Architecture: T5-small
Task: Language Translation
Dataset: Custom multilingual dataset
Fine-tuning Framework: Hugging Face Transformers
Quantization: Dynamic (int8) for efficiency

🚀 Usage

Installation

pip install transformers torch datasets

Loading the Model

from transformers import T5ForConditionalGeneration, T5Tokenizer
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

model_name = AventIQ-AI/t5-language-translation
model = T5ForConditionalGeneration.from_pretrained(model_name).to(device)
tokenizer = T5Tokenizer.from_pretrained(model_name)

Perform Translation


def translate_text(model, tokenizer, input_text, target_language):
    device = "cuda" if torch.cuda.is_available() else "cpu"
    formatted_text = f"translate English to {target_language}: {input_text}"
    input_ids = tokenizer(formatted_text, return_tensors="pt").input_ids.to(device)
    
    with torch.no_grad():
        output_ids = model.generate(input_ids, max_length=50)
    
    return tokenizer.decode(output_ids[0], skip_special_tokens=True)

# 🔹 **Test Translation**
input_text = "Hello, how are you?"
target_language = "French"  # Options: "Spanish", "German".
translated_text = translate_text(model, tokenizer, input_text, target_language)

print(f"Original: {input_text}")
print(f"Translated: {translated_text}")

📊 Evaluation Results

After fine-tuning, the model was evaluated on a multilingual dataset, achieving the following performance:

Metric	Score	Meaning
BLEU Score	38.5	Measures translation accuracy
Inference Speed	Fast	Optimized for real-time translation

🔧 Fine-Tuning Details

Dataset

The model was trained using a multilingual dataset containing sentence pairs from multiple language sources.

Training Configuration

Number of epochs: 3
Batch size: 8
Optimizer: AdamW
Learning rate: 2e-5
Evaluation strategy: Epoch-based

Quantization

The model was quantized using fp16 quantization, reducing latency and memory usage while maintaining accuracy.

📂 Repository Structure

.
├── model/               
├── tokenizer_config/   
├── quantized_model/     
├── README.md

⚠️ Limitations

The model may struggle with very complex sentences.
Low-resource languages may have slightly lower accuracy.
Contextual understanding is limited to sentence-level translation.