license: creativeml-openrail-m
datasets:
- prithivMLmods/Spam-Text-Detect-Analysis
language:
- en
base_model:
- google-bert/bert-base-uncased
pipeline_tag: text-classification
library_name: transformers
SPAM DETECTION UNCASED [ SPAM / HAM ]
Overview
This project implements a spam detection model using the BERT (Bidirectional Encoder Representations from Transformers) architecture and leverages Weights & Biases (wandb) for experiment tracking. The model is trained and evaluated using the prithivMLmods/Spam-Text-Detect-Analysis dataset from Hugging Face.
🛠️ Requirements
- Python 3.x
- PyTorch
- Transformers
- Datasets
- Weights & Biases
- Scikit-learn
Install Dependencies
You can install the required dependencies with the following:
pip install transformers datasets wandb scikit-learn
📈 Model Training
Model Architecture
The model uses BERT for sequence classification:
- Pre-trained Model:
bert-base-uncased - Task: Binary classification (Spam / Ham)
- Optimization: Cross-entropy loss
Training Arguments
- Learning rate:
2e-5 - Batch size: 16
- Epochs: 3
- Evaluation: Epoch-based.
🔗 Dataset
The model uses the Spam Text Detection Dataset available at Hugging Face Datasets.
You can access the dataset here.
🖥️ Instructions
Clone and Set Up
Clone the repository, if applicable:
git clone <repository-url>
cd <project-directory>
Ensure dependencies are installed with:
pip install -r requirements.txt
Train the Model
After installing dependencies, you can train the model using:
from train import main # Assuming training is implemented in a `train.py`
Replace train.py with your script's entry point.
✨ Weights & Biases Integration
We use Weights & Biases for:
- Real-time logging of training and evaluation metrics.
- Tracking experiments.
- Monitoring evaluation loss, precision, recall, and accuracy.
Set up wandb by initializing this in the script:
import wandb
wandb.init(project="spam-detection")
📊 Metrics
The following metrics were logged:
- Accuracy: Final validation accuracy.
- Precision: Fraction of predicted positive cases that were truly positive.
- Recall: Fraction of actual positive cases predicted.
- F1 Score: Harmonic mean of precision and recall.
- Evaluation Loss: Loss during validation on evaluation splits.
🚀 Results
Using BERT with the provided dataset:
- Validation Accuracy:
0.9937 - Precision:
0.9931 - Recall:
0.9597 - F1 Score:
0.9761
📁 Files and Directories
model/: Contains trained model checkpoints.data/: Scripts for processing datasets.wandb/: All logged artifacts from Weights & Biases runs.results/: Training and evaluation results are saved here.
📜 Acknowledgements
Dataset Source: Spam-Text-Detect-Analysis on Hugging Face
Model: BERT for sequence classification from Hugging Face Transformers.