Gemma2-STaR-MathHelper

Description

Gemma2-STaR-MathHelper is a project that leverages the Gemma2 language model to efficiently solve mathematical problems. The project integrates a Retrieval-Augmented Generation (RAG) system and the Self-Teaching with Rationale (STaR) method along with mathematical tools to provide precise solutions and well-structured rationales.

Features

• LLM Model: Uses the Gemma2-9b model from Ollama.
• Information Retrieval System: Implemented with FAISS and SentenceTransformer for embeddings.
• REACT Agent: Configured to use mathematical tools for problem-solving.
• Integrated Calculator: Tools for evaluating expressions, integration, differentiation, equation solving, and matrix operations.
• STaR Iterations: Generation and improvement of examples through STaR iterations.

Requirements

• Python 3.8 or higher
• Python libraries:
  • langchain-community
  • sentence-transformers
  • faiss-cpu
  • dotenv
  • huggingface_hub

Installation

1. Clone the repository:

   git clone https://github.com/your-username/Gemma2-STaR-MathHelper.git
   cd Gemma2-STaR-MathHelper

2. Create a virtual environment:

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install the dependencies:

   pip install -r requirements.txt

4. Configure the environment variables:

   Create a .env file in the root directory of the project and add your OpenAI API key:

   OPENAI_API_KEY=your_api_key

Usage

1. Ensure the virtual environment is activated:

   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

2. Run the main script:

   python main.py

3. Modify data/apuntes_matematicas.txt to add your own mathematical notes.

Project Structure

• main.py: Main script that configures and runs the agent.
• utils.py: Utilities for loading notes, setting up the QA chain, and calculator tools.
• calculator.py: Implementation of the Calculator class with mathematical functions.
• star_model.py: Implementation of functions for generating and filtering rationales.
• data/apuntes_matematicas.txt: Text file with mathematical notes for the retrieval system.

Paper Implementation

This project implements the concepts described in the paper "STaR: Self-Teaching with Rationales for Improved Explainability and Performance". Here's a brief overview of the implementation steps:

1. Retrieval-Augmented Generation (RAG):

   • Uses FAISS for indexing and retrieving relevant mathematical notes.
   • Utilizes SentenceTransformer to generate embeddings for the notes.

2. Self-Teaching with Rationales (STaR):

   • Generates initial rationales for a set of example questions and answers.
   • Filters and selects correct rationales to create a refined set of examples.
   • Iteratively improves the examples through multiple iterations of generating and filtering rationales.

3. Integration with Tools:

   • Configures an agent with access to various mathematical tools (calculator, integrator, differentiator, etc.) to solve complex mathematical problems.
   • Uses the REACT framework to describe the steps and tools used in solving problems.

Example Usage

To solve a mathematical problem, you can modify main.py to use the solve_math_problem function with your desired question:

question = "What is the integral of x^2?"
answer = solve_math_problem(question)
print(answer)