Mustafaege/sms-spam-balanced

SMS Spam Collection (Balanced)

A balanced SMS spam dataset for text classification.

Overview

Property	Value
Total Samples	1,494
Train	1,045 (70%)
Validation	149 (10%)
Test	300 (20%)
Classes	ham (0), spam (1)

Dataset Description

This is a balanced version of the UCI SMS Spam Collection dataset. Originally, the dataset had 5,572 messages with an imbalanced distribution (4,825 ham, 747 spam). We balanced it to 747 ham and 747 spam for training efficiency.

Label Distribution (Balanced)

Label	Count
ham (0)	747
spam (1)	747

Split Distribution

Split	ham	spam	Total
train	~522	~523	1,045
validation	~75	~74	149
test	~150	~150	300

How It Was Created

1. Downloaded from UCI Machine Learning Repository
2. Balanced by undersampling ham messages to match spam count (747 each)
3. Split into train (70%), validation (10%), test (20%) sets
4. Converted labels to integers: ham=0, spam=1

Usage with Hugging Face

from datasets import load_dataset

# Load dataset
dataset = load_dataset("mustafaege/sms-spam-balanced")

# Access splits
train_data = dataset['train']
val_data = dataset['validation']
test_data = dataset['test']

# Example
print(train_data[0])
# {'Text': 'Go until jurong point, crazy..', 'Label': 0}

Usage with Pandas

from datasets import load_dataset

dataset = load_dataset("mustafaege/sms-spam-balanced")

# Convert to pandas
train_df = dataset['train'].to_pandas()
print(train_df.head())

Fields

Field	Type	Description
Text	string	SMS message text
Label	int	0 = ham (not spam), 1 = spam

Example Messages

Ham (Label 0)

• "Go until jurong point, crazy.."
• "Ok lar... Joking wif u oni..."
• "U dun say so early hor..."

Spam (Label 1)

• "Free entry in 2 a wkly comp to win FA Cup final..."
• "Had your contract mobile 11 Mnths? Latest Moto..."
• "This is the 2nd time we have tried to contact u..."

For Fine-tuning

This dataset is designed for fine-tuning LLMs for spam detection.

Training Details Used

• Epochs: 5
• Batch Size: 16
• Learning Rate: 2e-5
• Optimizer: AdamW
• Loss Function: CrossEntropyLoss

Results

Metric	Score
Test Accuracy	~95%
Validation Accuracy	~95%+

Companion Model

Model: mustafaege/spam-detector-gpt2-hf

License

The original UCI SMS Spam Collection dataset is available under the UCI Machine Learning Repository terms.

References

• Original Dataset: UCI SMS Spam Collection