Root Signals has announced Root Judge. That is a model for evaluating large language models (LLMs). The training took place on the European LUMI supercomputer, in which Belgium also participated.
The American company Root Signals is introducing Root Judge. This is an LLM with unique functionality. The model must evaluate the output of other models to extract errors and hallucinations from them.
Root Judge is a modified version of Meta’s Llama-3.3-70B. The training was run on 384 AMD Radeon Instinct MI250X GPUs inside the LUMI supercomputer. That system is located in Finland and is among the fastest supercomputers in the world. Belgium also helped build LUMI.
Review
Designed as an assessment model for LLMs, Root Judge helps organizations detect hallucinatory responses in AI applications. The model can also be used to compare different model implementations and supports implementations where privacy is an important issue.
According to Root Signals, Root Judge provides transparent and explainable assessments, allowing companies to deploy AI models more reliably. The model is accessible under an open license.
The LLM further stands out because it was trained on a European system with AMD hardware. Indeed, all too many LLMs are the product of training on American HPC clusters, usually equipped with Nvidia accelerators. This innovation shows that Europe also has systems powerful enough for AI training, as well as alternatives to Nvidia.