Companion repository for Stammzelltisch on May 21!

In the Stammzelltisch on May 21, 2024, I want to present the current state of our project to develop an assistance system for the design of clinical studies using artificial intelligence techniques, running title of the app is "trialGPT". In order to make the topic more tangible, I intend to include a short hands-on session. For this part, the participants ideally install some software in advance (see section Setup). In this session, the participants are invited to take part in a short tutorial for the use of large language models (LLMs) on their local computer.

The examples are in Python, using mainly the python library langchain and the Llama2 LLM, installed on the local machine via the software Ollama. Furthermore, the langchain library is used to perform retrieval augmented generation (RAG), a technique which allow to feed the knowledge of additional documents into the LLM. Last, the tutorial contains a quick walk-through through the prototype of trialGPT.

In this readme-file, the steps to install the required software are described. Furthermore, instructions how to (hopefully) completely remove the installed software after the tutorial are provided.

Setup - to be done ideally in advance

What you need:

• Laptop with at least 8 Gb of Ram (tutorial tested mainly on Apple silicon)
• Ideally macOS with Homebrew
  • For other operating systems a little more action is required on the user side, but in most cases everything should work fine as well
• Terminal
• Browser

Setup python and python environment

Python and its environment can easily be set up via conda. conda is a software that allows to install scientific software in an easy and easily removable way. miniconda is a minimalistic version of conda.

You already have conda installed:

Open a terminal and type:

conda update conda

You want to install conda de novo (approx. 3 minutes):

• On macOS:

brew install --cask miniconda

• Else: Follow instructions here

Setup Ollama (approx 4 GB will be downloaded)

• On macOS:

brew install ollama

• Else: Follow instructions here
• To download the Llama2 LLM, you need to start a local Ollama server. To do so, open a terminal and type (leave this terminal open for the rest of this tutorial):

ollama serve

• Now open a second terminal, from which you can steer the ollama model:

ollama pull llama2

Download tutorial files

The tutorial files will be created in your home folder and can be removed at the end of the session. It might not work from the IMB-Intranet - so please connect via Eduroam.

cd ~
git clone https://github.com/sebastian-gerdes/stammzelltisch_2024_04_16.git
cd stammzelltisch_2024_04_16

Setup python environment (a few hundred MBs will be downloaded)

conda env create --name stammzelltisch -f stammzelltisch.yml
conda activate stammzelltisch

Open the jupyter file

• Launch jupyter notebook server:

jupyter notebook

• Open server in the browser
• Open the notebook hands_on/llm_tutorial.ipynb
• Execute the code cells from top to bottom by selecting them and then pressing Shift + Enter
• Be patient, — interaction with the LLM can take some moments
• When finished, feel free to kill all terminals you have used for the tutorial

Removing the installed files via terminal:

• Miniconda including the created environments

rm -r ~/miniconda/

• Ollama

brew uninstall ollama

• Tutorial files

rm -r ~/stammzelltisch_2024_04_16