Title:	DL4NLP: Deep Learning for Natural Language Processing
DNr:	Berzelius-2026-79
Project Type:	LiU Berzelius
Principal Investigator:	Marco Kuhlmann <marco.kuhlmann@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-03-31 – 2026-10-01
Classification:	10208
Homepage:	https://liu-nlp.ai/
Keywords:

Abstract

The Natural Language Processing (NLP) group at Linköping University is working on developing techniques for fine-tuning, adapting, and analyzing large language models (LLMs), with a particular focus on multilingual modeling and medium- and low-resource languages. Given our recent utilization patterns and the fact that several projects have progressed from exploratory phases to full implementation and scaling, it is important to address our resource needs for the foreseeable future. Our demand for GPU resources has increased substantially due to a growing number of concurrent projects and the broader shift in NLP research toward larger models, larger datasets, and more compute-intensive training and evaluation protocols. With the introduction of **Berzelius-Hopper**, we can significantly improve the feasibility and efficiency of our core workloads, which increasingly include (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 workflows for evaluation, translation, and data curation. At the same time, our group maintains mixed needs across projects, where smaller-scale experiments, ablations, evaluation runs, and method development remain well-served by the existing **Berzelius-Ampere** partition. We therefore request an updated allocation of 20,000 GPU hours per month, distributed as 16,000 GPU hours/month on Berzelius-Hopper (H200) and 4,000 GPU hours/month on Berzelius-Ampere (A100). This split reflects both the computational intensity of our medium-model training and high-throughput inference workloads, and the continued need for Ampere resources for smaller-scale experiments, baselines, and evaluation pipelines.