CHAIR OF

SOFTWARE ENGINEERING

The CSSE is excited to announce the release of kkirchheim/german-gpt2-medium, a GPT-2-style language model pre-trained on German text, offering improved capabilities for natural language processing in the German language. Existing German-only models often fall short in terms of context window size and parameter count. Our model bridges this gap, making it better equipped for tasks requiring larger context understanding and more efficient language modeling at modest parameter-count. Running the quantized model requires less than 1GB of VRAM. Explore technical details here.

Key Features

• 358M Parameters: Over twice the size of existing German-only GPT-2 models.
• Extended Context Length: A context window of 2048 tokens, double the standard 1024 tokens in similar models.
• High-Quality Dataset: Trained on 300GB of high-quality German text, leveraging the German Colossal, Cleaned Common Crawl Corpus (GC4).

Tim Gonschorek and Konstantin Kirchheim presented their research contributions at this year's SafeComp conference, held in Florence, Italy.

Tim Gonschorek presented his work on "Validating Design-Intent System Specification Models with State-of-the-art Large Language Models," which explores innovative methods for validating system specifications using advanced language models.

Meanwhile, Konstantin Kirchheim received the Best Paper Award at the Workshop on Artificial Intelligence Safety Engineering (WAISE), co-located with SafeComp, for his paper "Language Models as Reasoners for Out-of-Distribution Detection." His work offers new insights into leveraging language models for improving safety-critical systems.

We proudly congratulate both Tim and Konstantin for their achievements.

Use of collaborative robots, that allow shared workspace with human workers, integration of intelligent sensors enabling minor robot path corrections are already widely used in the industry. But although these advances are certainly great milestones, what still remained mostly unchanged and inadvertently hinders better integration, is the approach to robotic programming itself. In most cases the human coworker is treated more like a hindrance and a potential collision object that the robot needs to avoid, instead of an active cooperator for performing a common task.

The new project “Human-centered production through human-robot teaming” aims to address this problem and enable a completely new type of automation, by developing a technology that allows humans and robots to work together flexibly and efficiently.

Marco Filax successfully defended his PhD with the title "Fine-Grained Open-World Recognition: Identifying Retail Products in Supermarkets". Congratulations on the excellent work!

The Chair of Software and Systems Engineering recently hosted a meeting for the Safetrain project consortium in Magdeburg. Over two days, our partners from Siemens, the Fraunhofer Institute, Bridgefield, and others, engaged in focused discussions. The agenda included detailed presentations and workshops, primarily centered on reviewing the project's current progress.

We extend our thanks to all attendees for their participation and contributions. Your presence in Magdeburg was greatly valued. We look forward to the possibility of welcoming you again in the near future.

Es wird im Wintersemester 2024 zwei Prüfungsslots für mündliche Prüfungen im Fach Spezifikationstechnik geben.
Diese sind am Mittwoch, den 30.01.2024 und am Donnerstag, den 31.01.2024. Jeweils von 8.30Uhr bis 11Uhr. Bis zum 7.12.2023 wird für die Anmeldung über das LSF möglich sein..

Der letztendliche Prüfungstermin wird Ihnen dann nach der endgültigen Frist zugeteilt.

Bei Fragen können Sie sich gern an Frau Rulf oder Tim Gonschorek wenden.

We are pleased to announce that the new website of our chair, accessible at cse.ovgu.de, has been launched. This replaces our old website, which was available at cse.cs.ovgu.de.

Our joint paper with ABB on „Measuring the robustness of machinelearning models against data quality issues in industrial time series data“ presented at INDIN 2023 has received the Best Paper Award in the methods category. Congratulations to all authors!