Tag der Technik 2024

LLM Research and Paper Writing

---

This repository documents a 1 hour seminar on "Scientific Paper Writing" designed for high school juniors and seniors. In this course, the students are given a hands-on introduction to the scientific workflow; from coming up with a research question to experimenting to writing a paper.

Setting up the Course

Note that no software needs to be installed on the students' PCs. All the software required is run within a Docker Container on a server hosted by the organizers. The containers ship with:

• LaTeX (for writing the paper)
• Python (for experiments)
• The VSCode Server code-server

The VSCode server can be accessed via any web browser to write and execute code (see the screenshot above).

Tip

If you do not want to build the Docker Container yourself (no need to unless you want to change the course material), you can use our prebuild container:

ghcr.io/webis-de/seminar-tag-der-technik:latest

Requirements

• A server to host Docker Containers on (e.g., Kubernetes or simply a server with Docker installed)
• A working Open AI API chat completions endpoint (e.g., an API Key for the OpenAI API or a self-hosted service like ollama)

Caution

If you use a public API using the API Key (e.g., if you use OpenAI's official API), course participants can read the API key as well, and it should be handled as compromised. If you host the chat completions API yourself (e.g., using ollama), it does not need to be publicly accessible but only accessible from the server the containers are run on.

Deploying

Using Docker

docker run -p <port>:8080 ghcr.io/webis-de/seminar-tag-der-technik:latest

Using Kubernetes

See tag-der-technik.k8s.yaml for a deployment template (some values need to be set before it can be used).

Setup

1. Deploy a container for each student (see the hints above) and then perform the following steps on each student PC. In this guide we will assume that the students can reach the service at tdtXX.example.com. That is, student 1 opens tdt01.example.com, student 2 opens tdt02.example.com, ...
2. Open tdtXX.example.com on any (up to date) web browser
3. Type in the password when prompted (see the configuration in Dockerfile; per default we set c7b0a)
4. (Optional) When asked if you want to trust the authors, click "yes"
5. The bottom right corner should ask if you want to install the recommended plugins; click "yes"
6. Wait for extension installations to complete
7. Build the paper once (open paper/paper.tex and press Strg+S; paper/paper.pdf should pop up)
8. (Recommended) Open paper/paper.tex, paper.paper.pdf, and experiments/experiments.ipynb as shown in the screenshot above
9. In experiments.ipynb click Select Kernel > Python Environmentsand choose the kernel marked with a star
10. (Optional) Collapse the larger code section in Step 2 (after "Evaluation") by clicking on the gray or blue bar to the left of it

To Test if Everything Works

Run all Cells in the Jupyter Notebook. Output similar to what is shown in the screenshot above should appear.

Useful Commands

Building the Docker Container

docker build . -t tag-der-technik

Deploying the Docker Container

docker run -p <port>:8080 tag-der-technik

Replace <port> with the port you want the application to be deployed on.