AventIQ-AI / roberta-paraphrase-detection

Paraphrase Detection with Roberta-base

📌 Overview

This repository hosts the quantized version of the Roberta-base model for Paraphrase Detection. The model is designed to determine whether two sentences convey the same meaning. If they are similar, the model outputs "duplicate" with a confidence score; otherwise, it outputs "not duplicate" with a confidence score. The model has been optimized for efficient deployment while maintaining reasonable accuracy, making it suitable for real-time applications.

🏗 Model Details

• Model Architecture: Roberta-base
• Task: Paraphrase Detection
• Dataset: Hugging Face's quora-question-pairs
• Quantization: Float16 (FP16) for optimized inference
• Fine-tuning Framework: Hugging Face Transformers

🚀 Usage

Installation

pip install transformers torch

Loading the Model

from transformers import RobertaTokenizer, RobertaForSequenceClassification
import torch

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

model_name = "AventIQ-AI/roberta-paraphrase-detection"
tokenizer = RobertaTokenizer.from_pretrained(model_name)
model = RobertaForSequenceClassification.from_pretrained(model_name).to(device)

Paraphrase Detection Inference

def predict_paraphrase(sentence1, sentence2, threshold=0.96):
    inputs = tokenizer(sentence1, sentence2, return_tensors="pt", padding=True, truncation=True).to(device)

    with torch.no_grad():
        outputs = model(**inputs)

    logits = outputs.logits
    probabilities = torch.nn.functional.softmax(logits, dim=-1)
    predicted_class = torch.argmax(probabilities, dim=-1).item()
    confidence = probabilities[0][predicted_class].item()

    label_map = {0: "Not Duplicate", 1: "Duplicate"}

    # Apply a slightly less strict threshold
    if predicted_class == 1 and confidence < threshold:
        return {"sentence1": sentence1, "sentence2": sentence2, "predicted_label": "Not Duplicate", "confidence": confidence}
    else:
        return {"sentence1": sentence1, "sentence2": sentence2, "predicted_label": label_map[predicted_class], "confidence": confidence}

# 🔍 Test Example
test_cases = [
    ("The sun rises in the east.", "The east is where the sun rises."),  # Duplicate
    ("She enjoys playing the piano.", "She loves playing musical instruments."),  # Duplicate
    ("I had a great time at the party.", "The event was really fun."),  # Duplicate

    ("The sky is blue.", "Bananas are yellow."),  # Not Duplicate
    ("The capital of France is Paris.", "Berlin is the capital of Germany."),  # Not Duplicate
    ("I like reading books.", "She is going for a run."),  # Not Duplicate
]
for sent1, sent2 in test_cases:
    result = predict_paraphrase(sent1, sent2)
    print(result)

📊 Quantized Model Evaluation Results

🔥 Evaluation Metrics 🔥

• ✅ Accuracy: 0.7515
• ✅ Precision: 0.6697
• ✅ Recall: 0.5840
• ✅ F1-score: 0.6022

⚡ Quantization Details

Post-training quantization was applied using PyTorch's built-in quantization framework. The model was quantized to Float16 (FP16) to reduce model size and improve inference efficiency while balancing accuracy.

📂 Repository Structure

.
├── model/               # Contains the quantized model files
├── tokenizer_config/    # Tokenizer configuration and vocabulary files
├── model.safetensors/   # Quantized Model
├── README.md            # Model documentation

⚠️ Limitations

• The model may struggle with highly nuanced paraphrases.
• Quantization may lead to slight degradation in accuracy compared to full-precision models.
• Performance may vary across different domains and sentence structures.

🤝 Contributing

Contributions are welcome! Feel free to open an issue or submit a pull request if you have suggestions or improvements.