Swethamani/Translation-Models-En-G

English-Galician Translation Project

---

Description

This repository focuses on training and evaluating machine learning models for English-to-Galician translation. It uses T5-small as the primary model and Helsinki-NLP/opus-mt-en-gl as the secondary model. The project demonstrates the fine-tuning process, evaluation using automatic and human metrics, and comparative analysis of the models. The models leverage the Hugging Face Transformers library and the Seq2Seq framework, ensuring high-quality and efficient translation capabilities.

---

Features

1. Model Training:
   • Fine-tuning T5-small and Helsinki-NLP/opus-mt-en-gl on a custom dataset of 66,358 samples.
   • Preprocessing includes tokenization, truncation, and mapping source-target pairs.
   • Training configuration includes 3 epochs with learning rate and batch size optimizations.
2. Evaluation:
   • Semantic Similarity: Measured using SentenceTransformer embeddings.
   • Automatic Metrics: BLEU and ROUGE for translation quality.
   • Human Evaluation: Scores for relevance, fluency, coherence, engagement, and creativity.
3. Comparison:
   • Includes variability in outputs and metrics for realistic simulation.
   • Side-by-side evaluation of both models to highlight trade-offs.

---

Model Details

Model Summary

This project involves two fine-tuned models designed for English-to-Galician translation tasks:

• T5-small: A lightweight model, ideal for flexible translations with moderate training resources.
• Helsinki-NLP/opus-mt-en-gl: A high-performance model tailored for domain-specific translation tasks.

Model Description

This model was developed to translate English instructions into Galician outputs using fine-tuned machine learning techniques. It achieves robust translation capabilities through careful training and evaluation.

• Base Model: T5-small
• Languages: English (en), Galician (gl)
• Dataset: Custom dataset with 66,358 examples
• Training Epochs: 3
• Batch Size: 32
• Learning Rate: 5e-5
• Framework: PyTorch with 🤗 Transformers

---

Repository Structure

• training.ipynb: Notebook for fine-tuning the models.
• evaluation.ipynb: Notebook for evaluating outputs using semantic and human-level metrics.
• Galician.json: Dataset containing English-Galician translation pairs.
• translation_models/: Directory to save fine-tuned models.
• scripts/: Helper functions for preprocessing, training, and evaluation.

---

Setup

1. Clone the repository:

   git clone https://github.com/username/Translation-Models-En-Gl.git
   

2. Install dependencies:

   pip install transformers datasets sentencepiece torch sacrebleu rouge-score
   

3. Ensure a GPU is available for training (e.g., Google Colab or local machine with CUDA).

---

Usage

Fine-tuning

• Use the training.ipynb notebook to fine-tune T5-small or Helsinki-NLP models.
• The trained models are saved in the translation_models directory.

Evaluation

• Run the evaluation.ipynb notebook to generate outputs and compute metrics.
• Modify the instruction and input text to test different scenarios.

Example

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Swethamani/Translation-Models-En-G")
model = AutoModelForSeq2SeqLM.from_pretrained("Swethamani/Translation-Models-En-G")

# Translate English to Galician
input_text = "Translate this text into Galician."
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Results

T5-small

• Strengths: Provides flexibility and decent translation quality.
• Limitations: May require more fine-tuning for nuanced sentences.

Helsinki-NLP

• Strengths: Offers superior translation quality for complex or domain-specific texts.
• Limitations: Lacks generalization for other NLP tasks.

---

Evaluation

Metrics

• Semantic Similarity: Using SentenceTransformer embeddings to compare input and output.
• BLEU and ROUGE Scores: For assessing translation quality.
• Human Evaluation: Scored on relevance, fluency, coherence, engagement, and creativity.

---

Uses

Direct Use

• Translating English instructions into Galician.
• Generating Galician outputs based on English input prompts.

Results Comparison

• View the training loss and evaluation scores for both models.
• Compare BLEU, ROUGE, and human evaluation scores to determine performance.

---

Acknowledgements

• Libraries Used: Hugging Face transformers, SentenceTransformer, datasets, sacrebleu, and rouge-score.
• Dataset: Custom-curated English-Galician translations.

---