The rapid adoption of Large Language Models (LLMs) in civilian sectors necessitates robust safety mechanisms to prevent the generation of harmful, illegal, or unethical content. While current safety tools are predominantly English-centric, there is a critical need for guardrails that operate effectively across diverse linguistic contexts. This project aims to train and release a state-of-the-art, instruction-tuned safety model designed to classify and mitigate risks in both user prompts and model responses across multiple European and global languages.
Leveraging EuroHPC computing resources, the researchers will fine-tune a foundation model on a massive, curated dataset of safety taxonomies, covering categories such as hate speech, violence, self-harm, personally identifiable information (PII), and sexual content. The resulting model will function as a programmable "safety shield," capable of identifying policy violations and facilitating responsible AI behavior in mixed- language environments. By producing an open, accessible multilingual safety guard, this project will lower the barrier for developers to deploy Trustworthy AI systems, ensuring compliance with safety standards while preserving model utility.
Principal Investigator, Company and Country
Orestis Papakyriakopoulos, Alinia AI Europe SL, Spain