FairSense-AI API Documentation

Introduction

FairSense-AI is an AI-driven platform for detecting and analyzing bias in textual and visual content. This document outlines the key functions and APIs provided by FairSense for integration and usage. We are releasing a multimodal bias detection toolkit

1. Core Components

Device Setup

Sets up the device for model computation (CPU or GPU).

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

Model Initialization

Preloads the required models: - Text Model: unsloth/Llama-3.2-1B-Instruct or meta-llama/Llama-3.2-1B-Instruct or any instruct model - Image Captioning Model: Salesforce/blip-image-captioning-large - Summarizer: sshleifer/distilbart-cnn-12-6

2. Helper Functions

post_process_response(response)

  • Purpose: Cleans and summarizes AI model responses.
  • Parameters:
  • response (str): The raw response from the AI model.
  • Returns: A cleaned and summarized response string.
  • Example: python processed_response = post_process_response("This is the raw response from the model.") print(processed_response)

highlight_bias(text, bias_words)

  • Purpose: Highlights specific biased words in the text.
  • Parameters:
  • text (str): The input text to analyze.
  • bias_words (list): A list of words to highlight as biased.
  • Returns: HTML-formatted text with highlighted bias words.
  • Example: python highlighted_text = highlight_bias("This is a biased statement.", ["biased"]) print(highlighted_text)

3. Text Analysis

generate_response_with_model(prompt, progress=None)

  • Purpose: Generates a response from the AI model for a given prompt.
  • Parameters:
  • prompt (str): The input prompt for the model.
  • progress (callable, optional): Function to track progress.
  • Returns: AI-generated response as a string.
  • Example: python response = generate_response_with_model("Analyze this text for bias.") print(response)

analyze_text_for_bias(text_input, progress=gr.Progress())

  • Purpose: Analyzes a given text for bias and provides a detailed analysis.
  • Parameters:
  • text_input (str): Text to analyze.
  • progress (gr.Progress): Progress tracker.
  • Returns: Highlighted text and detailed analysis.
  • Example: python highlighted, analysis = analyze_text_for_bias("This text may contain bias.") print(highlighted) print(analysis)

4. Image Analysis

preprocess_image(image)

  • Purpose: Converts images to grayscale and applies thresholding for OCR.
  • Parameters:
  • image (PIL.Image): The input image.
  • Returns: A preprocessed image for OCR.
  • Example: python from PIL import Image image = Image.open("example.jpg") preprocessed = preprocess_image(image) preprocessed.show()

analyze_image_for_bias(image, progress=gr.Progress())

  • Purpose: Analyzes an image for bias by extracting text and generating captions.
  • Parameters:
  • image (PIL.Image): The input image.
  • progress (gr.Progress): Progress tracker.
  • Returns: Highlighted captions and detailed analysis.
  • Example: python image = Image.open("example.jpg") highlighted, analysis = analyze_image_for_bias(image) print(highlighted) print(analysis)

5. Batch Processing

analyze_text_csv(file, output_filename="analysis_results.csv")

  • Purpose: Analyzes a CSV file of text entries for bias.
  • Parameters:
  • file (File): CSV file with text data.
  • output_filename (str): Name of the output CSV file.
  • Returns: An HTML table with analysis results.
  • Example: python html_table = analyze_text_csv("data.csv") print(html_table)

analyze_images_batch(images, output_filename="image_analysis_results.csv")

  • Purpose: Analyzes multiple images for bias.
  • Parameters:
  • images (list): List of image paths.
  • output_filename (str): Name of the output file.
  • Returns: HTML table with analysis results and image previews.
  • Example: python results = analyze_images_batch(["image1.jpg", "image2.png"]) print(results)

save_results_to_csv(df, filename="results.csv")

  • Purpose: Saves analysis results to a CSV file.
  • Parameters:
  • df (pandas.DataFrame): DataFrame containing results.
  • filename (str): Name of the output file.
  • Returns: Path to the saved file.
  • Example: python results_df = pd.DataFrame([{"text": "example", "analysis": "unbiased"}]) save_path = save_results_to_csv(results_df, "output.csv") print(save_path)

6. AI Governance

ai_governance_response(prompt, progress=None)

  • Purpose: Provides insights into AI governance and safety.
  • Parameters:
  • prompt (str): Topic or question about AI governance.
  • progress (callable, optional): Progress tracker.
  • Returns: AI-generated insights and recommendations.
  • Example: python insights = ai_governance_response("Discuss AI ethics.") print(insights)

7. AI Safety Dashboard

display_ai_safety_dashboard()

  • Purpose: Visualizes AI safety risks using interactive charts.
  • Returns: Tuple containing bar chart, pie chart, scatter plot, and DataFrame.
  • Example: python fig_bar, fig_pie, fig_scatter, risks_df = display_ai_safety_dashboard() fig_bar.show()

Next Steps

This documentation provides the foundation for integrating FairSense into your workflows. Contact: For inquiries, collaborations, or feedback, connect with Shaina Raza, PhD, at shaina.raza@vectorinstitute.ai. Let me know if you need anything else added! 😊