Classifier-Bias-MR Model Card

1. Introduction

This model is designed to classify text data into 'biased' and 'non-biased' categories, utilizing the power of the BERT architecture.

2. Model Details

Architecture: BERT (base, uncased)

Fine-tuning Objective: Binary classification (Bias / Non-Bias)

3. Dataset Details

3.1 Source

The model is trained on the BABE dataset, which contains news articles from various sources.

3.2 Categories

Biased: 1863 articles

Non-Biased: 1810 articles

4. Training Procedure

Optimizer: AdamW

Fine tuned: Optuna

5. Model Performance

On a random split validation dataset:

Accuracy: 81%

F1 Score (Biased): 81%

F1 Score (Non-Biased): 81%

6. Usage

#Use a pipeline as a high-level helper

from transformers import pipeline

pipe = pipeline("text-classification", model="Social-Media-Fairness/Classifier_bias_mizanur_base_model")

#Load model directly

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("Social-Media-Fairness/Classifier_bias_mizanur_base_model")

model = AutoModelForSequenceClassification.from_pretrained("Social-Media-Fairness/Classifier_bias_mizanur_base_model")

7. Potential Applications

Content Moderation: To identify and filter out biased content from platforms.

Research: Assisting in bias-related studies and media analysis.

Education: For teaching and understanding media bias.

8. Caveats and Limitations

The model's training data originates from a specific dataset (BABE) which might not represent all kinds of biases or content.

The performance metrics are based on a random validation split, so the model's performance might vary in real-world applications.

9. Maintenance and Updates

For updates on model performance, improvements, or changes, please refer to the repository Social-Media-Fairness/Classifier_bias_mizanur_base_model.

10. Licensing

This model is licensed under OpenRail.