SWE-Smith-DataScience

This project fine-tunes an open-source AI model to specialize in Data Science tasks using a portion of the SWE-Smith dataset.

Table of Contents

• Overview
• Features
• Prerequisites
• Installation
• Project Structure

Overview

SWE-Smith-DataScience provides a modular pipeline to fine-tune large language models on specialized data science problems extracted from the SWE-Smith dataset and evaluate them via the SWE-agent framework.

Features

• Model Fine-Tuning: Leverage finetune.py to train on data subsets.
• Automated Evaluation: Use the SWE-agent module to benchmark performance.
• Reproducible Workflows: Modular scripts for data processing and training.

Prerequisites

• Python 3.8 or higher
• Git
• (Optional) Modal CLI for scalable remote training
• Required Python packages:

  pip install torch transformers datasets modal
  

Installation

git clone https://github.com/HugoGoHe/SWE-Smith-DataScience.git
cd SWE-Smith-DataScience

Project Structure

SWE-Smith-DataScience/
├── SWE-agent/           # Evaluation agent and benchmarking scripts
├── finetune.py          # Fine-tuning script for the model
├── requirements.txt     # Python dependencies (optional)
├── LICENSE              # MIT License
└── README.md            # Project documentation