Focus of the D3 Innovation Lab
Language models related to generative AI are a central part of the digital landscape today and are used in many areas of natural language processing, such as text generation, revision, and translation. General-purpose models such as GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder Representations from Transformers) are particularly versatile and can be used for a wide variety of tasks without adaptation. However, they do not always offer optimal performance for specific use cases. Through adaptation methods such as fine-tuning, prompt engineering, or the integration of new knowledge sources, these models can be specifically optimized for specific requirements to increase their performance for specific use cases.
Such adapted instances enable the versatility of general-purpose models to be used even more effectively. By fine-tuning them for specific data or tasks, better results can be achieved and the models can be adapted to specific needs. This project aims to develop a guide to help users select the language model best suited to their individual requirements and fine-tune it accordingly. In this way, the potential of generative AI systems is to be exploited to the best possible extent.
Review and results
The project achieved several key results that have a lasting impact both in terms of content and methodology. Of particular note is the fact that 105 students were reached over the course of six workshops, who intensively explored the functionality, possibilities, and limitations of generative language models. This not only imparted technical knowledge but also promoted a reflective approach to tools supported by generative AI.
The students acquired numerous skills: They learned to recognize differences, for example, in performance, between general and fine-tuned language models and to critically evaluate their potential applications. Practical exercises introduced them to topics such as data preparation, model training, data-driven decision-making for language models, as well as fine-tuning approaches and evaluation of results. Furthermore, important interdisciplinary skills were also developed, including critical reflection on ethical issues (bias, fairness, data protection) and the ability to assess the potential and risks of AI-supported applications in different application areas.
Another key outcome was the development and piloting of a guide for selecting and adapting language models. Although this guide has not yet been made publicly available due to necessary quality assurance, it has already been tested and iterated in the workshops. Students provided valuable feedback: They emphasized that the guide provides orientation in an otherwise confusing model landscape and helped them make informed decisions about a) the choice of a base language model, b) whether fine-tuning is necessary, and c) the most suitable fine-tuning approach for an underlying use case.
Tips from lecturers for lecturers
New didactic formats were tested, particularly workshop-based approaches with a strong practical focus (hands-on learning). One practical experience demonstrated the importance of engaging heterogeneous groups from students with prior technical knowledge to those outside the field equally – ideally based on a common starting point – through differentiated tasks, clear structuring, and accompanying reflection phases. This helped sharpen skills in moderating interdisciplinary discussions and dealing with diverse learning requirements. This led to particularly engaging conversations with and among the students, who contributed and shared their perspectives.
Furthermore, the project raised awareness of the importance of continuous feedback loops, even in ongoing workshop series: The direct integration of student feedback into the further development of the concept proved particularly valuable.