A Model-Agnostic Framework for Natural Language Explanations of Predictive Models

This repository contains the source code for the Master's Thesis: "A Model-Agnostic Framework for Natural Language Explanations of Predictive Models".

The framework introduces a robust, Object-Oriented architecture that integrates Feature Importance Methods (SHAP, LIME) and Logic-based Approaches (Araucana-XAI) with Large Language Models (Llama-3 via Unsloth). The system generates coherent, truthful, and role-adapted clinical reports, specifically designed to mitigate LLM hallucinations through structural logic injection.

📂 Repository Contents

• A_Model_Agnostic_Framework_for_NLE_of_Predictive_Models.ipynb: The complete, self-contained Jupyter Notebook implementing the OOP framework and experimental pipeline.

🚀 Quick Start (Google Colab)

The easiest way to replicate the experiments is utilizing Google Colab's free GPU tier.

1. Download the .ipynb file from this repository.
2. Upload it to Google Colab.
3. Runtime Setup:
   • Go to Runtime > Change runtime type.
   • Select T4 GPU as the hardware accelerator.
4. Execution:
   • Run the installation cells to set up the environment.
   • The framework automatically handles the download and caching of the Llama-3 model.

🛠️ System Architecture & Requirements

The project is built on Python 3.10 and utilizes a Singleton Pattern to manage GPU memory efficiently.

• Core Class: NLEFramework (Orchestrates XAI extraction and Prompt Generation).
• Hardware: NVIDIA Tesla T4 (Required for 4-bit quantization).
• Key Libraries:
  • unsloth (Efficient LLM Inference)
  • araucanaxai (Symbolic Decision Tree Surrogates)
  • shap & lime (Feature Attribution)
  • scikit-learn (Black-box modeling)

🔄 Extensibility & Domain Adaptation

This framework is designed with modularity in mind, using an Adapter Pattern to switch between different domains (e.g., Oncology, Finance, Law) without modifying the core codebase.

How to Adapt to a New Domain:

1. Data Layer: Replace the data loading function in Block 2 with your custom tabular dataset. Retrain the black_box_model (e.g., Random Forest) on your new data.

2. Initialization: Instantiate the framework with your specific class labels:

   nle = NLEFramework(class_names={0: "Loan Denied", 1: "Loan Approved"})
   

3. Adapter Layer (Prompt Engineering): Do not edit the source code. Instead, register a new Adapter using the register_adapter() method. Define a new System Role and use the provided UNIVERSAL_USER_TEMPLATE.

   # Example: Defining a Financial Analyst Persona
   finance_system = """
   SYSTEM ROLE: You are a Senior Financial Analyst.
   OBJECTIVE: Explain the credit score decision based on debt-to-income ratio.
   ...
   """
   
   # Registering the new capability
   nle.register_adapter("finance_expert", finance_system, UNIVERSAL_USER_TEMPLATE)
   

🔬 Experimental Reproducibility

The notebook includes a comparative experiment block comparing:

• Baseline Approach: Explanations based solely on SHAP/LIME weights.
• Hybrid Approach (Proposed): Explanations grounded in Araucana's symbolic logic trees.

Run Block 5 in the notebook to generate the side-by-side comparison reports for specific test instances.