PARA-PHRASER: Fine-Tuned T5 Paraphrasing Model with LoRA

🔍 Overview

A lightweight yet powerful paraphrasing tool powered by a fine-tuned t5-small model with LoRA (Low-Rank Adaptation) for efficient parameter tuning. Achieves high-quality paraphrasing while running efficiently on CPU.

✨ Key Features

• Efficient Fine-Tuning: LoRA configuration with rank=16 (r=16)
• CPU-Friendly: Optimized to run smoothly on local machines
• Web Interface: Clean FastAPI backend with responsive HTML/CSS/JS frontend
• Quality Results: Produces fluent, diverse paraphrases while preserving meaning

🛠️ Technical Details

Model Architecture

• Base Model: t5-small (60M parameters)
• Fine-Tuning Method: LoRA (Low-Rank Adaptation)
  • Rank: 16
  • Target Modules: Attention q/v layers
• Training Data: Custom paraphrasing dataset

Performance

• Inference Time (CPU): ~0.5-1 seconds per sentence
• Memory Usage: <1GB RAM
• Output Quality: Comparable to larger models for most use cases