Ginidu2003/Distilbert-Base-News-classifier

Model Details

Model Name: Ginidu2003/Distilbert-Base-News-classifier
Model Type: Text Classification
Base Model: distilbert/distilbert-base-uncased
Language(s): English
Finetuned from model: distilbert/distilbert-base-uncased

Model Description

This is a fine-tuned DistilBERT model designed to classify English news articles into 5 categories:

• Business
• Opinion
• Political gossip
• Sports
• World news

Uses

Direct Use

• Classify news articles into one of the five predefined categories.
• Suitable for English news (Like Daily Mirror style).

Downstream Use

• Can be integrated into web applications (Streamlit/Gradio) for automated news categorization.
• Can be used for real-time news filtering and topic-based news recommendation systems.

Out-of-Scope Use

• Not intended for other languages.
• Not trained for sentiment analysis, fake news detection, or hate speech detection.
• Not suitable for very short texts.

Bias, Risks, and Limitations

• The model is trained only on Daily Mirror news data, so it may perform poorly on other news sources or different writing styles.
• Potential bias towards Sri Lankan context and English used in Sri Lankan media.
• Performance may degrade on very long or very short articles.

How to Get Started with the Model

from transformers import pipeline

classifier = pipeline(
    "text-classification", 
    model="Ginidu2003/Distilbert-Base-News-classifier"
)

result = classifier("Your news article text here...")
print(result)

Training Details

Training Data

• Dataset: Daily Mirror Sri Lankan English news (2024–2025)
• Total samples: ~1,018 articles (after preprocessing and deduplication)
• Classes: 5 balanced categories (Business, Opinion, Political gossip, Sports, World news)
• Preprocessing: Lowercasing, punctuation removal, lemmatization

Training Procedure

• Framework: Hugging Face Transformers + Trainer API
• Base Model: distilbert/distilbert-base-uncased
• Epochs: 4
• Batch Size: 8
• Learning Rate: 2e-5
• Validation Accuracy: 90.19%

Evaluation

Validation Set Results (20% hold-out):

• Accuracy: 91.18%
• Model shows strong and consistent performance across all 5 classes.

Environmental Impact

• Training was done on a single GPU (T4 GPU on Colab)
• Estimated carbon emissions: Very low (small model + small dataset)