ExplainableAI-Project2: Social Explainability By Design

Project Overview

This project focuses on generating natural language explanations for AI decision-making processes. It demonstrates how formal, technical explanations can be transformed into conversational, user-friendly explanations that are more accessible to non-technical users.

Group Number: 1

Team Members:

1. Abhinav Atmuri
2. Bar Melinarskiy
3. Konstantinos Zavantias
4. Nikita Aksjonov

Getting Started

Prerequisites

• Python 3.8+
• Docker and Docker Compose
• Required Python packages (install via pip install -r requirements.txt):
  • anytree
  • numpy
  • pandas
  • requests

Project Structure

• Part1/ - Contains all code for the first part of the assignment regarding general goal tree explanations using anytree
• Part2/ - Natural language explanations for decisions made by the agent in the goal tree
  • nl_generation.ipynb: Jupyter notebook for generating natural language explanations
  • ollama_docker/ollama.yaml: Docker configuration for the language model
• generate_explanations_for_test_cases.py: Generates baseline technical explanations
• requirements.txt: List of required Python packages

Running the Project

Step 1: Generate Baseline Explanations

Run the following command to generate the initial technical explanations:

python generate_explanations_for_test_cases.py

Step 2: Set Up the Language Model

Start the Docker container with the language model (Mistral):

docker-compose -f ./Part2/ollama_docker/ollama.yaml up --build -d

Step 3: Generate Natural Language Explanations

Open and run the Jupyter notebook:

jupyter notebook Part2/nl_generation.ipynb

Explanation Output

The generated explanations will be saved in:

• explanations/baseline/: Technical explanations
• explanations/llama3/ or explanations/mistral/: Natural language explanations

Shutting Down

To stop the Docker container when finished:

docker-compose -f ./Part2/ollama_docker/ollama.yaml down

Project Description

This project demonstrates how AI systems can provide more human-friendly explanations of their decision-making processes. By transforming formal reasoning codes and technical explanations into natural language, we make AI decisions more transparent and accessible to non-technical users.