| --- |
| library_name: transformers |
| language: |
| - en |
| metrics: |
| - f1 |
| base_model: |
| - distilbert/distilbert-base-uncased |
| datasets: |
| - AbdulHadi806/mail_spam_ham_dataset |
| pipeline_tag: text-classification |
| --- |
| |
| # Model Card for Model ID |
|
|
| Text classification Model for Spam Detection (Deep Learning Project). |
|
|
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| Model developped for the "Deep Learning with Python" course Project |
|
|
| - **Developed by:** Diavila Rostaing Engandzi |
| - **Model type:** Binary Text Classification |
| - **Language(s) (NLP):** English |
| - **Finetuned from model:** DistilBERT |
|
|
| ### Model Sources |
|
|
| - **Demo [optional]:** https://huggingface.co/picket-cliff/deepl-project |
|
|
| ## Uses |
|
|
| The model is intended to be used to sort spam in emails. Clone and Run the app.py file in the Demo to see it in action. |
|
|
| ## Training Details |
|
|
| ### Training Data |
|
|
| Subset from the email_data.csv dataset [card]. |
| |
| A benchmark dataset for email classification with around 5000 emailed classified between "ham" and "spam". |
| To evaluate the model, data was separated between training and test datasets (80-20 split). |
| |
| #### Preprocessing |
| |
| Deep learning models cannot process raw text; they require numerical tensors. We utilized the Hugging Face DistilBertTokenizer. |
| |
| 1. Sub-word Tokenization: Instead of splitting by spaces (which struggles with typos and rare words), DistilBERT uses WordPiece tokenization. For example, an out-of-vocabulary word might be broken into known sub-words, preventing the model from encountering "Unknown" tokens. |
| |
| 2. Special Tokens: The tokenizer automatically prepends the [CLS] (Classification) token to the start of the sequence and the [SEP] (Separator) token at the end. The final hidden state corresponding to the [CLS] token is what the model uses for the binary classification decision. |
| |
| 3. Truncation and Padding: Transformer models require fixed-size input matrices for batch processing. Based on our EDA length distribution, we set max_length = 128. |
|
|
| o Sentences longer than 128 tokens were truncated. |
| |
| o Sentences shorter than 128 tokens were padded with the [PAD] token (ID 0). |
|
|
| 4. Attention Masks: To prevent the model from performing Self-Attention on meaningless padding tokens, the tokenizer generates an attention_mask (an array of 1s for real words and 0s for padding). |
| |
| |
| ## Evaluation |
| |
| Results obtained directly from training on the training dataset then evaluating the model on the testing data. |
| Result are compared to a baseline (dummy classifier) for reference. |
| |
| ### Testing Data, Factors & Metrics |
| |
| #### Testing Data |
| |
| |
| |
| #### Metrics |
| |
| Accuracy, f1 score (macro and weighted) |
| |
| ### Results |
| |
| When evaluated on a 80-20 split we obtained: |
| |
| • Accuracy: 99.10% |
| |
| • Macro Average F1-Score: 0.98 |
| |
| • Weighted Average F1-Score: 0.99 |
| |
| Meanwhile the dummy achieved 86.6% accuracy. |
| |
| #### Summary |
| |
| The model performance is |
