About the Project

When the time came to write my thesis, the topic of Explainable AI (XAI) was suggested to me by one of my professors at the university. I decided to specifically look into ChatBots, as they are pivotal in the field of LLM applications. The official topic of my thesis was:

Building an Interpretable Natural Language AI Tool based on Transformer Models and approaches of Explainable AI.

The result of this work is a 70-page thesis that explores how techniques of XAI can be applied to LLMs and ChatBots. I examined several different approaches and open-source implementations after outlining all the important fundamentals.

Additionally, I developed a working prototype aimed at interpreting and explaining the decisions of a Mistral 7B Instruct Model within a chat interface. This prototype is designed to improve explainability for common users.

Features

The thesis project includes several key features:

• Comprehensive Research: Detailed exploration of XAI techniques applied to LLMs and ChatBots.
• Prototype Development: Creation of a prototype to interpret and explain LLM decisions.
• User-friendly Interface: Designed to make AI decisions understandable to non-experts.
• Extensive Documentation: A 70-page thesis documenting the research and development process.

Technology

The implementation of my thesis relied on several open-source technologies due to their availability and robustness. The key technologies used include:

• Huggingface Transformers: Utilized for model implementation and handling.
• Captum and SHAP: Employed for XAI implementation, with KernelSHAP used for interpreting the Mistral model.
• Gradio: Used for building the user interface and interaction layer.
• FastAPI: Used for developing the backend and API services.

Architecture

The architecture of the project is designed to facilitate research and prototyping in a flexible and efficient manner:

• Programming Language: Python, chosen for its extensive libraries and community support.
• Model Framework: Huggingface Transformers for handling transformer models.
• XAI Tools: Captum and SHAP for explainability, providing insights into model decisions.
• UI Framework: Gradio for building interactive user interfaces.
• Backend: FastAPI for creating robust backend services.
• Development Environment: Jupyter Notebooks for iterative development and documentation.