Sentence-BERT Quantized Model for Text Similarity & Paraphrase Detection

This repository hosts a quantized version of the Sentence-BERT (SBERT) model, fine-tuned on the Quora Question Pairs dataset for text similarity and paraphrase detection. The model computes semantic similarity between two input sentences and has been optimized for efficient deployment using ONNX quantization.

Model Details

Model Architecture: Sentence-BERT (all-MiniLM-L6-v2)
Task: Text Similarity & Paraphrase Detection
Dataset: Quora Question Pairs (QQP)
Quantization: ONNX (Dynamic Quantization)
Fine-tuning Framework: Sentence-Transformers (Hugging Face)

Usage

Installation

pip install sentence-transformers onnxruntime transformers

Loading the Model

Original Fine-tuned Model

from sentence_transformers import SentenceTransformer

# Load the fine-tuned model
model = SentenceTransformer("fine-tuned-model")

# Encode two sentences and compute cosine similarity
sentence1 = "How can I learn Python?"
sentence2 = "What is the best way to study Python?"

emb1 = model.encode(sentence1)
emb2 = model.encode(sentence2)

# Cosine similarity
import numpy as np
score = np.dot(emb1, emb2) / (np.linalg.norm(emb1) * np.linalg.norm(emb2))
print("Similarity Score:", score)

# Threshold to classify as paraphrase
print("Paraphrase" if score > 0.75 else "Not Paraphrase")

Quantized ONNX Model

from onnxruntime import InferenceSession
from transformers import AutoTokenizer
import numpy as np

# Load tokenizer and ONNX session
tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
session = InferenceSession("sbert_onnx/model.onnx")

def encode_onnx(session, tokenizer, sentence):
    inputs = tokenizer(sentence, return_tensors="np", padding=True, truncation=True)
    outputs = session.run(None, dict(inputs))
    return outputs[0][0]

# Encode and compute similarity
emb1 = encode_onnx(session, tokenizer, sentence1)
emb2 = encode_onnx(session, tokenizer, sentence2)
score = np.dot(emb1, emb2) / (np.linalg.norm(emb1) * np.linalg.norm(emb2))
print("Quantized Similarity Score:", score)
print("Paraphrase" if score > 0.75 else "Not Paraphrase")

Performance Metrics

Accuracy: ~0.87
F1 Score: ~0.85
Threshold for classification: 0.75 cosine similarity

Fine-Tuning Details

Dataset

Source: Quora Question Pairs (Kaggle)
Size: 400K+ question pairs labeled as paraphrase or not

Training Configuration

Epochs: 3
Batch Size: 16
Evaluation Steps: 1000
Warmup Steps: 1000
Loss Function: CosineSimilarityLoss

Quantization

Method: ONNX dynamic quantization
Tool: Hugging Face Optimum + ONNX Runtime

Repository Structure

.
├── fine-tuned-model/      # Fine-tuned SBERT model directory
├── sbert_onnx/            # Quantized ONNX model directory
├── test_functions.py      # Code for evaluation and testing
├── README.md              # Project documentation

Limitations

The cosine similarity threshold (0.75) may need tuning for different domains.
ONNX quantization may introduce slight performance degradation compared to full-precision models.
SBERT embeddings do not produce classification logits, only similarity scores.

Contributing

Contributions are welcome! Please open an issue or submit a pull request for bug fixes or improvements.