| | --- |
| | language: |
| | - en |
| | license: apache-2.0 |
| | tags: |
| | - text-classification |
| | - distilbert |
| | - clickbait |
| | - moderation |
| | datasets: |
| | - marksverdhei/clickbait_title_classification |
| | metrics: |
| | - accuracy |
| | - f1 |
| | - precision |
| | - recall |
| | --- |
| | |
| | # Clickbait Classifier 🎣 |
| |
|
| | This model is a fine-tuned version of `distilbert-base-uncased` trained to classify text (news headlines, article titles, video names) into two categories: **Clickbait** and **Non-Clickbait**. |
| |
|
| | It is optimized for filtering out sensationalist headlines and improving content recommendation algorithms. |
| |
|
| | ## Intended Use |
| |
|
| | The primary goal of this model is to automatically detect clickbait titles to help users and platforms prioritize high-quality informative content over misleading or exaggerated headlines. |
| |
|
| | - **Input:** Raw English text (headlines, titles, tweets). |
| | - **Return:** A binary classification label (`Clickbait` or `Non-Clickbait`) with a confidence score. |
| |
|
| | ## Training Data |
| |
|
| | The model was fine-tuned using the `bhargavasthet/clickbait_dataset`, which contains a balanced collection of headlines explicitly labeled as clickbait (e.g., from Buzzfeed, Upworthy) and non-clickbait (e.g., from Reuters, The New York Times). |
| |
|
| | ## Performance Metrics |
| |
|
| | The model achieved excellent performance on the `marksverdhei/clickbait_title_classification` validation set: |
| |
|
| | - **Accuracy:** `0.9864` (98.6%) |
| | - **F1 Score:** `0.9862` (98.6%) |
| | - **Precision:** `0.9867` (98.6%) |
| | - **Recall:** `0.9857` (98.5%) |
| | - **Evaluation Loss:** `0.0488` |
| |
|
| | ## Training Constraints & Hyperparameters |
| |
|
| | The model was trained under the following conditions: |
| | - **Base Architecture:** `distilbert-base-uncased` (chosen for speed and efficiency) |
| | - **Maximum Sequence Length:** 128 |
| | - **Learning Rate:** 2e-05 |
| | - **Batch Size:** 64 |
| | - **Precision:** Mixed Precision (fp16) |
| | - **Optimizer Strategy:** Early Stopping (patience=3) |
| | - **Epochs:** 3 |
| |
|
| | ## Usage 🚀 |
| |
|
| | You can easily integrate this model into your applications using the Hugging Face `transformers` library pipeline: |
| |
|
| | ```python |
| | from transformers import pipeline |
| | |
| | # Load the clickbait classifier |
| | classifier = pipeline("text-classification", model="ENTUM-AI/distilbert-clickbait-classifier") |
| | |
| | # Test with a sensational headline |
| | text_1 = "10 Bizarre Facts About Apples That Will BLOW YOUR MIND! 🍎🤯" |
| | result_1 = classifier(text_1) |
| | print(f"Text: '{text_1}'\nPrediction: {result_1}\n") |
| | |
| | # Test with a normal news headline |
| | text_2 = "Apple releases new quarterly earnings report showing 5% growth." |
| | result_2 = classifier(text_2) |
| | print(f"Text: '{text_2}'\nPrediction: {result_2}") |
| | ``` |
| |
|
| | ## Expected Output format: |
| | ```json |
| | [{'label': 'Clickbait', 'score': 0.9921}] |
| | ``` |
| |
|
| | ## Potential Applications |
| | - 📰 **News Aggregators:** Filter out low-quality clickbait articles. |
| | - 📱 **Social Media Feeds:** Demote clickbait posts in recommendation algorithms. |
| | - ✉️ **Email Spam Filters:** Detect clickbait-style subject lines in promotional emails. |
| |
|
