Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tailoring Chatbots for Higher Education: Some Insights and Experiences

Gerd Kortemeyer

TL;DR

ETH Zurich analyzes how higher education institutions can tailor LLMs through three main customization routes: training from scratch, fine-tuning, and augmentation (RAG). It discusses public open-weight projects like Apertus, commercial APIs, on-premises inference, and governance concerns such as data provenance and EU residency. The paper argues that RAG is often a practical starting point, while fine-tuning and selective infrastructure investments can provide deeper domain alignment. The findings offer actionable guidance for universities planning scaleable, responsible LLM-enabled education.

Abstract

The general availability of general-purpose Large Language Models continues to impact on higher education, yet they may not always be useful for specialized tasks. When using these models, oftentimes the need for particular domain knowledge becomes quickly apparent, and the desire for customized bots arises. Customization holds the promise of leading to more accurate and contextually relevant responses, enhancing the educational experience. This report relates insights and experiences from one particular technical university in Switzerland, ETH Zurich, to describe what "customizing" Large Language Models means in practical terms for higher education institutions.

Tailoring Chatbots for Higher Education: Some Insights and Experiences

TL;DR

ETH Zurich analyzes how higher education institutions can tailor LLMs through three main customization routes: training from scratch, fine-tuning, and augmentation (RAG). It discusses public open-weight projects like Apertus, commercial APIs, on-premises inference, and governance concerns such as data provenance and EU residency. The paper argues that RAG is often a practical starting point, while fine-tuning and selective infrastructure investments can provide deeper domain alignment. The findings offer actionable guidance for universities planning scaleable, responsible LLM-enabled education.

Abstract

The general availability of general-purpose Large Language Models continues to impact on higher education, yet they may not always be useful for specialized tasks. When using these models, oftentimes the need for particular domain knowledge becomes quickly apparent, and the desire for customized bots arises. Customization holds the promise of leading to more accurate and contextually relevant responses, enhancing the educational experience. This report relates insights and experiences from one particular technical university in Switzerland, ETH Zurich, to describe what "customizing" Large Language Models means in practical terms for higher education institutions.
Paper Structure (9 sections, 2 figures)

This paper contains 9 sections, 2 figures.

Table of Contents

  1. Large Language Models
  2. Customized Large Language Models
  3. Training from Scratch
  4. Public Systems
  5. Commercial Systems
  6. Fine-Tuning
  7. Augmentation
  8. Disclaimer
  9. Conclusion

Figures (2)

  • Figure 1: Overview of the customization methods. The "'Effort"-axis can be considered logarithmic.
  • Figure 2: Example of the chatbot component of Ethel. Shown on the left is a short excerpt of the lecture script, on the right a dialogue with the system.