LlamaMathVerifier

A fine-tuned Llama3-8B model for mathematical answer verification using Supervised Fine-Tuning (SFT). This project implements a binary classification system to determine the correctness of mathematical solutions across various topics.

Project Overview

This implementation uses Parameter-Efficient Fine-Tuning (PEFT) with LoRA to adapt Llama3-8B for mathematical reasoning verification. The model evaluates mathematical solutions and provides binary (True/False) verification of answer correctness.

Dataset

Overview

The dataset consists of mathematical problems and their solutions, specifically structured for answer verification:

Structure

• Questions: Mathematical problems across various topics
• Answers: Provided solutions to be verified
• Solutions: Detailed explanations and reasoning
• Labels: Binary classification (True/False) indicating answer correctness

Specifications

• Training Set Size: 900000
• Validation Set Size: 100000
• Test Set Size: 10000
• Format: CSV
• Label Distribution: [Distribution of True/False labels if known]

Key Features

• Diverse mathematical topics
• Detailed solution explanations
• Binary verification labels
• No external datasets permitted
• Test partition reserved for evaluation only

Data Processing

• Text cleaning and normalization
• Prompt formatting
• Token length optimization
• Special character handling for mathematical symbols
• Solution explanation integration

Usage Restrictions

• Limited to competition-provided dataset
• No external data augmentation
• Test set strictly for evaluation
• Compliance with NYU academic rules

Key Features

• 🚀 Supervised Fine-Tuning of Llama3-8B
• 🔧 LoRA-based parameter-efficient training
• 📊 Binary classification for answer verification
• 🧮 Specialized mathematical reasoning
• 🔍 Solution explanation integration

Technical Specifications

• Model: Llama3-8B
• Training Method: SFT with LoRA
• Task: Binary Classification
• Input: Mathematical questions with answers and explanations
• Output: Boolean verification (True/False)

Implementation Details

• Parameter-Efficient Fine-Tuning using LoRA
• 4-bit quantization for memory efficiency
• Optimized prompt engineering
• Comprehensive evaluation metrics
• Inference pipeline for solution verification

Performance

• Baseline Accuracy (on running on Colab): 0.56025
• Finetuned Acccuracy: 0.82438
• Target Metric: Binary Classification Accuracy
• Evaluation: Validation Set Accuracy

Usage

Detailed implementation and usage instructions are provided in the Jupyter notebook. The repository includes:

• Training pipeline
• Inference code
• Evaluation metrics
• Hyperparameter configurations
• Example usage scenarios

Requirements

• Python 3.8+
• pyTorch
• transformers
• unsloth
• datasets
• random
• gc
• trl
• tqdm

Citation

If you use this implementation in your work, please cite:

@software{mathverifyllm2024,
  author = {[Abhishek Mahajan, Neeha Rathna Janjanam, Nimit Kanani]},
  title = {MathVerifyLLM: Mathematical Answer Verification using Llama3-8B},
  year = {2024},
  publisher = {GitHub},
  url = {https://github.com/nrjanjanam/MathVerifyLLM}
}