Optimizing LLM-Powered Agents for Tabular Data Analytics 🚀

Overview 📋

This repository contains the code and resources for my diploma thesis. The project explores the use of Large Language Models (LLMs) in analyzing tabular data using natural language by generation and execution of Python code. The thesis includes a comprehensive literature review, development of an LLM-based Agent program, and performance evaluations using fine-tuned and state-of-the-art models.

Project Structure 🗂️

The project is organized as follows:

TableQA-LLMAgent/       # Root directory
│
├── .github/            # CI/CD workflows
│   └── workflows/
│
├── README.md           # This README file
├── main.py             # Agent usage example
├── poetry.lock         # Poetry dependency management
├── pyproject.toml      # Readable dependencies
│
├── tableqallmagent/    # Source code of the package
│   ├── __init__.py
│   ├── agent.py        # Constructor and the main interface
│   ├── code_manipulation.py # Processing generated code
│   ├── coder_llms.py   # Forward passes for coding LLMs
│   ├── llms.py         # Higher level methods for LLMs
│   ├── logger.py       # Color constants for readability
│   ├── prompts.py      # Prompting strategies and formatting
│
├── dataset/            # Multiple datasets and preprocessing
├── dist/               # PyPI versions
├── evaluation/         # LLM-as-evaluator
├── finetuning/         # LoRA training scripts and configs
├── plots/              # Directory to store generated images
└── tests/              # pytest

Installation 🔧

To get started with the project, follow these steps:

1. Clone the repository:

   git clone https://github.com/poludmik/TableQA-LLMAgent.git
   cd TableQA-LLMAgent

2. Install the dependencies:

   poetry install

Usage 🚀

You can run the main script to see the basic example functionalities of the agent:

python main.py

Features ✨

• Fine-Tuning: Fine-tuning LLMs using LoRA and QLoRA techniques.
• Code Generation: Generating Python code to analyze tabular data.
• Model Evaluation: Rigorous benchmarks for evaluating LLM Agents.
• MLOps: Tracking experiments using MLOps tools to ensure reproducibility.

Results 📊

The fine-tuning experiments significantly improved the performance of the Code Llama 7B Python model from 35.3% to 60.3% on the proposed evaluation benchmark.

Contact 📫

For any questions or feedback, please reach out to me:

Mikhail Poludin
michael.poludin@gmail.com

License 📜

This project is licensed under the MIT License.