# 💬 Sentiment-Analysis-for-Product-Release-Sentiment

A **BERT-based sentiment analysis model** fine-tuned on a product review dataset. It predicts the sentiment of a text as **Positive**, **Neutral**, or **Negative** with a confidence score. This model is ideal for analyzing customer feedback, reviews, or user comments.

---

## ✨ Model Highlights

- 📌 **Architecture**: Based on [`bert-base-uncased`](https://huggingface.co/bert-base-uncased) by Google
- 🔧 **Fine-tuned** on labeled product review data
- 🔍 **3-way sentiment classification**: `Negative (0)`, `Neutral (1)`, `Positive (2)`
- 💾 **Quantized version available** for faster inference

---

## 🧠 Intended Uses

- ✅ Classifying product feedback and user reviews
- ✅ Sentiment analysis for e-commerce platforms
- ✅ Social media monitoring and customer opinion mining

---

## 🚫 Limitations

- ❌ Designed for English texts only
- ❌ May not perform well on sarcastic or ironic inputs
- ❌ May struggle with domains very different from product reviews
- ❌ Input texts longer than 128 tokens are truncated

---

## 🏋️‍♂️ Training Details

- **Base Model**: `bert-base-uncased`
- **Dataset**: Custom-labeled product review dataset
- **Epochs**: 5
- **Batch Size**: 8
- **Max Length**: 128 tokens
- **Optimizer**: AdamW
- **Loss Function**: CrossEntropyLoss (with class balancing)
- **Hardware**: Trained on NVIDIA GPU (CUDA-enabled)

---

## 📊 Evaluation Metrics

| Metric     | Score |
|------------|-------|
| Accuracy   | 0.90  |
| F1         | 0.90  |
| Precision  | 0.90  |
| Recall     | 0.90  |

---

## 🔎 Label Mapping

| Label ID | Sentiment |
|----------|-----------|
| 0        | Negative  |
| 1        | Neutral   |
| 2        | Positive  |

---

## 🚀 Usage Example

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from transformers import DataCollatorWithPadding
import torch
import torch.nn.functional as F

# Load model and tokenizer
model_name = "AventIQ-AI/Sentiment-Analysis-for-Product-Release-Sentiment"
tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForSequenceClassification.from_pretrained(model_name)
model.eval()

# Inference
def predict_sentiment(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    inputs = {k: v.to(quantized_model.device) for k, v in inputs.items()}
    with torch.no_grad():
        logits = quantized_model(**inputs).logits
        probs = F.softmax(logits, dim=1)

    predicted_class_id = torch.argmax(probs, dim=1).item()
    confidence = probs[0][predicted_class_id].item()

    label_map = {0: "Negative", 1: "Positive"}
    label = label_map[predicted_class_id]
    confidence_str = f"confidence : {confidence * 100:.1f}%"

    return label, confidence_str

# Example
print(predict_sentiment(
"The service was excellent and the staff was friendly.")
)
```

---

## 🧪 Quantization

- Applied **post-training dynamic quantization** using PyTorch to reduce model size and speed up inference.
- Quantized model supports CPU-based deployments.

---

## 📁 Repository Structure

```
.
├── model/               # Quantized model files
├── tokenizer/           # Tokenizer config and vocabulary
├── model.safetensors/   # Fine-tuned full-precision model
├── README.md            # Model documentation
```

---

## 📌 Limitations

- May not generalize to completely different domains (e.g., medical, legal)
- Quantized version may show slight drop in accuracy compared to full-precision model

---

## 🤝 Contributing

We welcome contributions! Please feel free to raise an issue or submit a pull request if you find a bug or have a suggestion.