TrustLLM Project Started

The EU project TrustLLM was officially launched on 20-21 November. In the next three years, the project will create a new generation of large language models with improved accessibility, reliability, and mitigated negative effects such as hallucinations and bias. The consortium has a focus on northern Europe. For example, AI Sweden, located in Stockholm, is a project partner that has created the LLM GPT-SWE, the strongest model in the Swedish language to date. The company Mi∂eind has worked with OpenAI on improving their models’ capacity in Icelandic, a low-resource language with only approximately 300,000 speakers for which the internet is flooded with low-quality automatic translations.

Within the project, JSC’s role will be twofold: Firstly, HPC experts will ensure that resource-intensive GPU-based LLM training will take place at an ideal level of efficiency on all systems. Secondly, the team from the Simulation and Data Lab Applied Machine Learning will work on new features of LLMs for digesting longer inputs, develop methods for reducing the parameter count, and work on scaling laws that allow for predictions of LLM performance without even having to train them.

More information can be found on the website of Linköping University, the project coordinator.

Contact: Dr. Stefan Kesselheim

from JSC News No. 300, 12 December 2023