README.md

---
datasets:
- zefang-liu/phishing-email-dataset
metrics:
- recall
- f1
- precision
base_model:
- distilbert/distilbert-base-uncased
pipeline_tag: text-classification
---

# DistilBERT Spam Classifier

A fine-tuned DistilBERT-based model for phishing email detection, trained on the [Phishing Emails Dataset](https://www.kaggle.com/datasets/subhajournal/phishingemails). This model is optimized for identifying spam and phishing emails with high accuracy.

## Model Overview
- **Base Model**: DistilBERT
- **Fine-Tuning**: Performed on a phishing email dataset to classify emails as spam (1) or non-spam (0).
- **Format**: Available in ONNX format for efficient deployment.

## Architecture
The model extends DistilBERT with a custom classification head:

```python
class DistilBERTSpamClassifier(nn.Module):
def __init__(self, distilbert):
super(DistilBERTSpamClassifier, self).__init__()
self.distilbert = distilbert
self.dropout = nn.Dropout(0.1)
self.relu = nn.ReLU()
self.fc1 = nn.Linear(768, 512)
self.fc2 = nn.Linear(512, 2)
self.softmax = nn.LogSoftmax(dim=1)
```

- **Input**: Tokenized email text (processed via DistilBERT tokenizer).
- **Output**: Log-probabilities for two classes (spam or non-spam).
- **Layers**:
- DistilBERT for contextual embeddings (768 dimensions).
- Dropout (0.1) for regularization.
- Fully connected layers (768 → 512 → 2) with ReLU activation.
- LogSoftmax for classification.

## Performance
Evaluated on a test set of 3,021 samples, the model achieves performance across metrics:

| Class | Precision | Recall | F1-Score | Support |
|-------|-----------|--------|----------|---------|
| Non-Spam (0) | 0.98 | 0.98 | 0.98 | 1,870 |
| Spam (1) | 0.96 | 0.97 | 0.96 | 1,151 |

- **Accuracy**: 97%
- **Macro Avg**: Precision: 0.97, Recall: 0.97, F1-Score: 0.97
- **Weighted Avg**: Precision: 0.97, Recall: 0.97, F1-Score: 0.97

## Usage
1. **Install Dependencies**:
```bash
pip install transformers onnxruntime torch
```

2. **Load the Model**:
Use the ONNX model with a compatible inference engine (e.g., ONNX Runtime). Example:

```python
from transformers import DistilBertTokenizer
import onnxruntime as ort
import numpy as np

# Load tokenizer
tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")

# Load ONNX model
session = ort.InferenceSession("path_to_model.onnx")

# Tokenize input
text = "Your example email text here"
inputs = tokenizer(text, return_tensors="np", padding=True, truncation=True, max_length=512)

# Run inference
outputs = session.run(None, dict(inputs))[0]
prediction = np.argmax(outputs, axis=1)
print("Spam" if prediction == 1 else "Non-Spam")
```

3. **Input Requirements**:
- Text input must be tokenized using the DistilBERT tokenizer.
- Maximum sequence length: 512 tokens.

## Dataset
The model was fine-tuned on the [Phishing Emails Dataset](https://www.kaggle.com/datasets/subhajournal/phishingemails), which contains labeled email samples for spam and phishing detection.

## Limitations
- Only available in ONNX format; no PyTorch or TensorFlow checkpoints.
- Maximum input length is 512 tokens; longer emails are truncated.