MathAi

Overview

MathAi is a project dedicated to enhancing the reasoning capabilities of Large Language Models (LLMs) on complex mathematical word problems. This repository contains a pipeline designed to fine-tune and evaluate LLMs using the MATH dataset by Hendrycks, which includes a diverse set of challenging math problems ranging from algebra to calculus.

Table of Contents

• Installation
• Dataset
• Pipeline
• Usage
• Contributing
• License
• Acknowledgements

Installation

To set up the MathAi pipeline, follow these steps:

1. Clone the repository:

   git clone https://github.com/Husnain-Amjad/MathAi.git
   cd MathAi
   

2. Install dependencies: Ensure you have Python 3.8+ installed. Then, install the required packages:

   pip install -r requirements.txt
   

3. MATH dataset: The MATH dataset by Hendrycks is placed in the AUG_MATH/ directory.

Dataset

The MATH dataset, created by Hendrycks et al., consists of 12500 math word problems across various difficulty levels and topics, including:

• Algebra
• Geometry
• Number Theory
• Calculus
• Probability

Each problem includes a question, a step-by-step solution, and the final answer, making it ideal for training and evaluating LLMs on mathematical reasoning.

Pipeline

The MathAi pipeline is designed to:

1. Preprocess the MATH dataset: Clean and format the dataset for model training.
2. Fine-tune LLMs: Adapt pre-trained LLMs to improve their performance on math word problems.
3. Evaluate performance: Use metrics such as accuracy and step-by-step reasoning correctness to assess model improvements.
4. Generate solutions: Allow the fine-tuned model to generate solutions for new math problems.

The pipeline supports popular LLM frameworks like PyTorch and Hugging Face Transformers.

Directory Structure

MathAi/
├── AUG_MATH/              # MATH dataset and related files
├── __pycache__/           # Compiled Python files
├── README.md              # Project documentation
├── base_data_loader.py    # Data loading utilities
├── bnb_config.py          # Configuration for training
├── data_processing.py     # Data preprocessing script (updated for padding tokens)
├── inference.py           # Inference script
├── lora_config.py         # LoRA configuration
├── main.py                
├── no_trainer_based_training.py  # Trainer-less training with comments cleaned
├── requirements.txt       # Dependency list (updated configuration)
├── tokenization.py        # Tokenization script (cleaned of useless code)
├── train.py               # Main Training script (updated for padding tokens)
├── trainer_based_training.py  # Trainer-based training script
├── training_args_config.py    # Updated argument passing for no-trainer
├── training_with_trainer.py   # Training with trainer script
└── training_without_trainer.py  # Updated stratified sampling

Usage

To use the MathAi complete pipeline:

Train the model:

python train.py --model_name <model_path>  --sample_ratio 0.30 --save_steps 300 --logging_steps 1500 --epochs 5 --use_quantization --use_lora --lora_rank 16 --lora_dropout 0.1 --output_dir <directory_path>

Replace <model_path> with the desired model and <directory_path> with output directory path.

Contributing

Currently repository is under maintainance. Contributions will be welcomed soon.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Acknowledgements

• The MATH dataset by Hendrycks et al. for providing a comprehensive set of math problems.

The open-source community for tools like PyTorch and Hugging Face Transformers.