Spam Detection with BERT

This repository contains an implementation of a Spam Detection model using BERT (Bidirectional Encoder Representations from Transformers) for binary classification (Spam / Ham). The model is trained on the prithivMLmods/Spam-Text-Detect-Analysis dataset and leverages Weights & Biases (wandb) for comprehensive experiment tracking.


πŸ—‚οΈ Summary of Uploaded Files

File Name Size Description Upload Status
.gitattributes 1.52 kB Tracks files stored with Git LFS. Uploaded
README.md 8.78 kB Comprehensive documentation for the repository. Updated
config.json 727 Bytes Configuration file related to the model settings. Uploaded
model.safetensors 438 MB Model weights stored in safetensors format. Uploaded (LFS)
special_tokens_map.json 125 Bytes Mapping of special tokens for tokenizer handling. Uploaded
tokenizer_config.json 1.24 kB Tokenizer settings for initialization. Uploaded
vocab.txt 232 kB Vocabulary file for tokenizer use. Uploaded

πŸ› οΈ Overview

Core Details:

  • Model: BERT for sequence classification
    Pre-trained Model: bert-base-uncased
  • Task: Spam detection - Binary classification task (Spam vs Ham).
  • Metrics Tracked:
    • Accuracy
    • Precision
    • Recall
    • F1 Score
    • Evaluation loss

πŸ“Š Key Results

Results were obtained using BERT and the provided training dataset:

  • Validation Accuracy: 0.9937
  • Precision: 0.9931
  • Recall: 0.9597
  • F1 Score: 0.9761

πŸ“ˆ Model Training Details

Model Architecture:

The model uses bert-base-uncased as the pre-trained backbone and is fine-tuned for the sequence classification task.

Training Parameters:

  • Learning Rate: 2e-5
  • Batch Size: 16
  • Epochs: 3
  • Loss: Cross-Entropy

πŸš€ How to Use the Model

1. Clone the Repository

git clone <repository-url>
cd <project-directory>

2. Install Dependencies

Install all necessary dependencies.

pip install -r requirements.txt

or manually:

pip install transformers datasets wandb scikit-learn

3. Train the Model

Assuming you have a script like train.py, run:

# Import necessary libraries
from datasets import load_dataset, ClassLabel
from transformers import BertTokenizer, BertForSequenceClassification, Trainer, TrainingArguments
import torch
from sklearn.metrics import accuracy_score, precision_recall_fscore_support

# Load dataset
dataset = load_dataset("prithivMLmods/Spam-Text-Detect-Analysis", split="train")

# Encode labels as integers
label_mapping = {"ham": 0, "spam": 1}
dataset = dataset.map(lambda x: {"label": label_mapping[x["Category"]]})
dataset = dataset.rename_column("Message", "text").remove_columns(["Category"])

# Convert label column to ClassLabel for stratification
class_label = ClassLabel(names=["ham", "spam"])
dataset = dataset.cast_column("label", class_label)

# Split into train and test
dataset = dataset.train_test_split(test_size=0.2, stratify_by_column="label")
train_dataset = dataset["train"]
test_dataset = dataset["test"]

# Load BERT tokenizer
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")

# Tokenize the data
def tokenize_function(examples):
    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=128)

train_dataset = train_dataset.map(tokenize_function, batched=True)
test_dataset = test_dataset.map(tokenize_function, batched=True)

# Set format for PyTorch
train_dataset.set_format(type="torch", columns=["input_ids", "attention_mask", "label"])
test_dataset.set_format(type="torch", columns=["input_ids", "attention_mask", "label"])

# Load pre-trained BERT model
model = BertForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=2)

# Move model to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Define evaluation metric
def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    predictions = torch.argmax(torch.tensor(predictions), dim=-1)
    precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average="binary")
    acc = accuracy_score(labels, predictions)
    return {"accuracy": acc, "precision": precision, "recall": recall, "f1": f1}

# Training arguments
training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",  # Evaluate after every epoch
    save_strategy="epoch",        # Save checkpoint after every epoch
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    num_train_epochs=3,
    weight_decay=0.01,
    logging_dir="./logs",
    logging_steps=10,
    load_best_model_at_end=True,
    metric_for_best_model="accuracy",
    greater_is_better=True
)

# Trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=test_dataset,
    compute_metrics=compute_metrics
)

# Train the model
trainer.train()

# Evaluate the model
results = trainer.evaluate()
print("Evaluation Results:", results)

# Save the trained model
model.save_pretrained("./saved_model")
tokenizer.save_pretrained("./saved_model")

# Load the model for inference
loaded_model = BertForSequenceClassification.from_pretrained("./saved_model").to(device)
loaded_tokenizer = BertTokenizer.from_pretrained("./saved_model")

# Test the model on a custom input
def predict(text):
    inputs = loaded_tokenizer(text, return_tensors="pt", padding="max_length", truncation=True, max_length=128)
    inputs = {k: v.to(device) for k, v in inputs.items()}  # Move inputs to the same device as model
    outputs = loaded_model(**inputs)
    prediction = torch.argmax(outputs.logits, dim=-1).item()
    return "Spam" if prediction == 1 else "Ham"

# Example test
example_text = "Congratulations! You've won a $1000 Walmart gift card. Click here to claim now."
print("Prediction:", predict(example_text))

πŸ”— Dataset Information

The training dataset comes from Spam-Text-Detect-Analysis available on Hugging Face:

Dataset size: - 5.57k entries

✨ Weights & Biases Integration

Why Use wandb?

  • Monitor experiments in real time via visualization.
  • Log metrics such as loss, accuracy, precision, recall, and F1 score.
  • Provides a history of past runs and their comparisons.

Initialize Weights & Biases

Include this snippet in your training script: python import wandb wandb.init(project="spam-detection")

πŸ“ Directory Structure

The directory is organized to ensure scalability and clear separation of components:

project-directory/
β”‚
β”œβ”€β”€ data/                # Dataset processing scripts
β”œβ”€β”€ wandb/              # Logged artifacts from wandb runs
β”œβ”€β”€ results/            # Save training and evaluation results
β”œβ”€β”€ model/              # Trained model checkpoints
β”œβ”€β”€ requirements.txt    # List of dependencies
└── train.py            # Main script for training the model

🌐 Gradio Interface

A Gradio interface is provided to test the model interactively. The interface allows users to input text and get predictions on whether the text is Spam or Ham.

Example Usage

import gradio as gr
import torch
from transformers import BertTokenizer, BertForSequenceClassification

# Load the pre-trained BERT model and tokenizer
MODEL_PATH = "prithivMLmods/Spam-Bert-Uncased"
tokenizer = BertTokenizer.from_pretrained(MODEL_PATH)
model = BertForSequenceClassification.from_pretrained(MODEL_PATH)

# Function to predict if a given text is Spam or Ham
def predict_spam(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        prediction = torch.argmax(logits, axis=-1).item()
    return "Spam" if prediction == 1 else "Ham"

# Gradio UI
inputs = gr.Textbox(label="Enter Text", placeholder="Type a message to check if it's Spam or Ham...")
outputs = gr.Label(label="Prediction")

examples = [
    ["Win $1000 gift cards now by clicking here!"],
    ["You have been selected for a lottery."],
    ["Hello, how was your day?"],
    ["Earn money without any effort. Click here."],
    ["Meeting tomorrow at 10 AM. Don't be late."],
    ["Claim your free prize now!"],
    ["Are we still on for dinner tonight?"],
    ["Exclusive offer just for you, act now!"],
    ["Let's catch up over coffee soon."],
    ["Congratulations, you've won a new car!"]
]

gr_interface = gr.Interface(
    fn=predict_spam,
    inputs=inputs,
    outputs=outputs,
    examples=examples,
    title="Spam Detection with BERT",
    description="Type a message in the text box to check if it's Spam or Ham using a pre-trained BERT model."
)

gr_interface.launch()

Downloads last month
17
Safetensors
Model size
109M params
Tensor type
F32
Β·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for prithivMLmods/Spam-Bert-Uncased

Finetuned
(2281)
this model

Dataset used to train prithivMLmods/Spam-Bert-Uncased

Space using prithivMLmods/Spam-Bert-Uncased 1

Collection including prithivMLmods/Spam-Bert-Uncased