nedimazar/thesisPublic

Anonymized version of the repo for my MSc thesis: "Interfacing Operating Systems with Large Language Models"

Dependencies

1. Python: 3.11.2
2. python3.11 -m venv venv
3. source venv/bin/activate
4. pip install -r requirements.txt
5. Make sure OPENAI_API_KEY environment variable is set.

roger (CLI agent):

Installation

1. Alias the script in your .bashrc/.zshrc like this:
   • alias roger="noglob [path_to_venv]/venv/bin/python [path_to_repo]/thesis/src/agent/roger.py"
2. Also add the repository to your PYTHONPATH
   • export PYTHONPATH=$PYTHONPATH:[path_to_repo]

Usage

• roger [Natural Language Pompt]

src/agent

roger's implementation lives here, you can edit the setup and fewshot promptsin utils.py. Feel free to change model="gpt-3.5-turbo" to model="gpt-4" in roger.py to use GPT-4.

src/finetune

The NL2Bash (Lin et al. 2018) dataset is used to fine-tune various GPT-3 models here. Data preparation, fine-tuning, and shell scripts to get predictions from the fine-tuned models can also be found here.

src/meta

Contains an inference cost estimation script.

src/userStudy

This directory contains the user study material. There are also scripts to annotate the experiments. The dockerStuff directory was used to set up a container for participants to use in the second task of the user study.