Title:	DL4NLP: Deep Learning for Natural Language Processing
DNr:	Berzelius-2026-268
Project Type:	LiU Berzelius
Principal Investigator:	Marco Kuhlmann <marco.kuhlmann@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-10-02 – 2027-05-01
Classification:	10208
Homepage:	https://liu-nlp.ai/
Keywords:

Abstract

The Natural Language Processing (NLP) group at Linköping University develops techniques for fine-tuning, adapting, and analyzing large language models (LLMs), with a particular focus on multilingual modeling and medium- and low-resource languages. Several projects that were exploratory at the time of our previous application have since moved into full implementation and scaling, and our resource consumption reflects this: demand for GPU hours has grown substantially with the number of concurrent projects and with the broader shift in NLP research toward larger models, larger datasets, and more compute-intensive training and evaluation protocols. Since gaining access to Berzelius-Hopper, we have moved our most demanding workloads there, with clear gains in feasibility and throughput. These workloads are (i) continued pre-training and training from scratch, (ii) fine-tuning and adaptation of medium- and large-scale models, and (iii) multi-GPU inference and synthetic data generation for evaluation, translation, and data curation. In parallel, our needs remain mixed across projects: smaller-scale experiments, ablations, evaluation runs, and method development continue to run well on Berzelius Ampere. While we have adjusted our experiments to the current allocation of 5,000 hours on Hopper and 10,000 hours on Berzelius, there is an increasing expectation to test approaches on larger models, and report variance on multiple seeds even for resoruce-intensive activities such as continued pre-training, which is why our computational demand continues to grow. We therefore request an allocation of 8,000 hours on Hopper and 15,800 hours on Ampere.