# DistilBERT Quantized Model for Sentiment Analysis on Yelp Polarity Dataset

This repository hosts a quantized version of the DistilBERT model, fine-tuned for sentiment analysis tasks on the Yelp Polarity dataset. The model has been optimized using post-training quantization to make it suitable for resource-constrained environments while maintaining high accuracy.

## Model Details

- **Model Architecture:** DistilBERT Base Uncased  
- **Task:** Sentiment Analysis  
- **Dataset:** Yelp Polarity  
- **Quantization:** Dynamic Quantization (INT8 on Linear layers)  
- **Fine-tuning Framework:** Hugging Face Transformers  

---

### Installation

```sh
pip install transformers datasets evaluate scikit-learn torch

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load trained model and tokenizer
model = AutoModelForSequenceClassification.from_pretrained("./results")
tokenizer = AutoTokenizer.from_pretrained("./results")

# Set model to eval mode
model.eval()

# 10 Sample review texts
sample_texts = [
    "The food was absolutely wonderful!",
    "Terrible experience. I will never come back.",
    "Average service, but the food was decent.",
    "I loved the ambiance and the staff was super friendly!",
    "Worst food I've had in a long time.",
    "Highly recommend this place for a date night.",
    "The waiter was rude and the food was cold.",
    "Amazing pizza, will order again!",
    "They took too long to serve and it was overpriced.",
    "Best customer service and delicious desserts!"
]

# Predict and print results
for text in sample_texts:
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
    with torch.no_grad():
        outputs = model(**inputs)
        prediction = torch.argmax(outputs.logits, dim=-1).item()
        sentiment = "Positive" if prediction == 1 else "Negative"
    print(f"Text: {text}\\nPredicted Sentiment: {sentiment}\\n")

Quantization

import os
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

# Load the fine-tuned model
model = AutoModelForSequenceClassification.from_pretrained("./results")

# Apply dynamic quantization
quantized_model = torch.quantization.quantize_dynamic(
    model,
    {torch.nn.Linear},
    dtype=torch.qint8
)

# Define path
quantized_model_path = "./results/quantized_model"

# Create directory if it doesn't exist
os.makedirs(quantized_model_path, exist_ok=True)

# Save the quantized model weights
torch.save(quantized_model.state_dict(), f"{quantized_model_path}/pytorch_model.bin")

# Save config and tokenizer
model.config.save_pretrained(quantized_model_path)
tokenizer = AutoTokenizer.from_pretrained("./results")
tokenizer.save_pretrained(quantized_model_path)

print("✅ Quantized model saved at:", quantized_model_path)

Performance Metrics
Accuracy: Approx. 95% on Yelp Polarity Test Subset

Precision, Recall, F1-score: Computed during evaluation using scikit-learn

Fine-Tuning Details
Dataset
Source: Yelp Polarity (via Hugging Face Datasets)

Train samples used: 50,000

Test samples used: 10,000

Training
Number of epochs: 3

Batch size: 16

Evaluation strategy: Per epoch

Learning rate: 2e-5

Weight decay: 0.01

Repository-Structure

.
├── results/                     # Contains fine-tuned and quantized model files
│   ├── pytorch_model.bin        # Quantized model weights
│   ├── config.json              # Model config
│   └── tokenizer/               # Tokenizer files
├── logs/                        # Training logs
├── README.md                    # Model documentation
 

Limitations

The model is trained only on Yelp reviews and may not generalize to other domains.

Post-training quantization may cause minor accuracy degradation compared to full-precision models.