---
dataset_info:
  features:
  - name: sms
    dtype: string
  - name: label
    dtype: int64
  - name: char_len
    dtype: int64
  - name: word_count
    dtype: int64
  - name: punct_score
    dtype: int64
  - name: spam_keywords
    dtype: int64
  - name: lexical_diversity
    dtype: float64
  - name: readability
    dtype: float64
  - name: caps_ratio
    dtype: float64
  - name: digit_ratio
    dtype: float64
  - name: exclaim_ratio
    dtype: float64
  - name: url_flag
    dtype: int64
  - name: spammy_words
    dtype: int64
  - name: entropy
    dtype: float64
  splits:
  - name: train
    num_bytes: 974514
    num_examples: 5171
  download_size: 446788
  dataset_size: 974514
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# SMS Spam Enriched Dataset

An enriched version of the classic **SMS Spam Collection Dataset from UC Irvine** with additional engineered features and semantic embeddings.  
This dataset is designed for **spam detection, feature engineering experiments, and model interpretability research**.

---

## Dataset Overview

- **Total samples**: 5,171  
- **Classes**: 
  - `0`: Ham (non-spam)  
  - `1`: Spam  

---

## Enrichments Added

Alongside the raw SMS text (`sms`) and labels (`label`), we engineered multiple new features:

1. **char_len** → Total characters in the message  
2. **word_count** → Total words in the message  
3. **punct_score** → Weighted score for punctuation usage (`!`, `?`, `...`)  
4. **spam_keywords** → Count of known spammy tokens (`free`, `win`, `urgent`, etc.)  
5. **lexical_diversity** → Ratio of unique words to total words  
6. **readability** → Flesch Reading Ease score  
7. **caps_ratio** → Proportion of uppercase characters  
8. **digit_ratio** → Proportion of numeric digits  
9. **exclaim_ratio** → Ratio of exclamation marks to total characters  
10. **url_flag** → Binary indicator for presence of URLs  
11. **spammy_words** → Count of flagged high-signal words  
12. **entropy** → Shannon entropy of characters  
13. **embeddings** → Sentence-transformer vector representations (for downstream tasks like clustering, semantic similarity, visualization)

---

## Example Row

| sms                                            | label | char_len | word_count | punct_score | spam_keywords | lexical_diversity | readability | caps_ratio | digit_ratio | url_flag | entropy |
|------------------------------------------------|-------|----------|------------|-------------|---------------|-------------------|-------------|------------|-------------|----------|---------|
| "Free entry in 2 a wkly comp to win FA Cup..." |   1   |   156    |    28      |     0       |      3        |       0.857       |    80.83    |   0.064    |   0.16      |    0     |  4.69   |

---

## Visualizations

Below is a **PCA projection** of SMS embeddings, showing clear separation between spam (red) and ham (blue):  

![PCA Plot](https://huggingface.co/datasets/GenAIDevTOProd/sms-spam-enriched/blob/main/PCA%20project%20of%20sms%20embeddings.png)

---

## Benchmark Models

We trained baseline classifiers using the enriched dataset:

- **Logistic Regression (with combined features)**  
  - Accuracy: ~99%  
  - F1 (spam): ~0.95  

- **Random Forest (with combined features)**  
  - Accuracy: ~98%  
  - F1 (spam): ~0.92

  Logistic Regression Report (Combined Features):

  
              precision    recall  f1-score   support

           0       0.99      1.00      0.99       904
           1       0.98      0.92      0.95       131

    accuracy                           0.99      1035

  
   macro avg       0.99      0.96      0.97      1035

  
weighted avg       0.99      0.99      0.99      1035


Random Forest Report (Combined Features):
              precision    recall  f1-score   support

           0       0.98      1.00      0.99       904
           1       0.99      0.86      0.92       131

    accuracy                           0.98      1035
    
   macro avg       0.99      0.93      0.96      1035
   
weighted avg       0.98      0.98      0.98      1035

---

## Use Cases

- Spam detection model training  
- Feature engineering demonstration  
- Embedding-based similarity and clustering tasks  
- Educational material for NLP + ML pipelines  

---

## Citation

If you use this dataset, please cite the original SMS Spam Collection dataset:  

> @inproceedings{Almeida2011SpamFiltering, title={Contributions to the Study of SMS Spam Filtering: New Collection and Results}, author={Tiago A. Almeida and Jose Maria Gomez Hidalgo and Akebo Yamakami}, year={2011}, booktitle = "Proceedings of the 2011 ACM Symposium on Document Engineering (DOCENG'11)", }.  

> Dataset Enrichment and Feature Engineering contributions by Naga Adithya Kaushik (GenAIDevTOProd). 
---

## License

This dataset is distributed under the same terms as the original SMS Spam dataset (publicly available for research).  

---