Model Card: Debiaser Package

Design Principles

1. Accuracy: The Debiaser package prioritizes accurate identification of bias and proper rephrasing of sentences. Ensuring precise discrimination of bias and maintaining the integrity of sentence meaning are primary goals.

2. Affordability: Unlike paid Language Model Models (LLMs) such as GPT-3.5 or Clause, the Debiaser package is designed to be accessible and cost-free for individuals, small organizations, and projects with significant workloads. It will consistently remain free of charge.

3. Performance: While performance is of slightly lesser importance compared to accuracy and affordability, the Debiaser package aims to offer reasonable speed and resource efficiency. It ensures the package is efficient without being excessively resource-intensive.

Usage

You can utilize the Debiaser package in Google Colab by following the demo provided above.

1. Instantiate the Package:

   from debiaser.main import Debiaser
   
   db = Debiaser()
   db.setup()
   

2. Classification:

   Classify a sentence as either biased or non-biased:

   classification_result = db.classify("All Americans are quite racist.")
   print(classification_result)
   

3. Debiasing:

   Debias a sentence using the package:

   debiased_text = db.debias("Men make better leaders than women.")
   print(debiased_text)
   

Implementation

The Debiaser package employs the Falcon-7B model for both bias classification and debiasing tasks. The model is quantized to 4 bits, enabling it to operate in low-compute environments like the free Google Colab. The design ensures efficient execution while maintaining accurate results.

For further information and updates, refer to the documentation provided with the package.

---

Note: This Model Card provides an overview of the design principles, usage, and implementation details of the Debiaser package. It aims to communicate key aspects of the package's functionality and features to users and developers.