Master Thesis: Cross-Lingual Safety Alignment & System Identity Injection via DPO
AI Sweden
Summary
Ensuring that an open-weight European model maintains a consistent system identity ("OpenEuroLLM"). Models often over-refuse or exhibit language-dependent safety leaks when prompted in low-resource European languages.
Job description
Project Background and Problem Statement
T4.6 has established the openeurollm-model-identity dataset and a multilingual safety refusal suite. This thesis asks:
- Can joint DPO co-training on openeurollm-model-identity and Safe-RLHF pairs eliminate cross-lingual safety leaks and over-refusal without degrading instruction-following performance in Swedish and Icelandic?
Outline
- Literature Study: Review cross-lingual safety gaps, EU AI Act risk taxonomies, and system prompt adherence in DPO.
- Implementation: Fine-tune Prelude 9B using openeurollm-model-identity and target-language Safe-RLHF preference pairs using Hugging Face TRL or Megatron-LM.
- Evaluation: Conduct cross-lingual jailbreak probes (e.g., Do-Not-Answer, X-Alpaca) and evaluate over-refusal rates on benign administrative queries.
Who we're looking for
We are seeking curious, self-driven MSc students eager to work at the frontier of open-weight European AI research (LLMs). You thrive on empirical discovery, design rigorous experiments, and let data challenge your assumptions.
- Ongoing Master's studies in Computer Science, Data Science, Machine Learning, Engineering Physics, or a related quantitative field.
- Proficiency in Python and hands-on experience with modern deep learning frameworks (PyTorch, Hugging Face ecosystem).
- Familiarity with LLM post-training alignment (e.g., SFT, DPO, RLHF/RLVR) or context-extension, alongside comfort running distributed GPU training in Linux/HPC environments.
Requirements
- Ongoing Master's studies in Computer Science, Data Science, Machine Learning, Engineering Physics, or a related quantitative field.
- Proficiency in Python and hands-on experience with modern deep learning frameworks (PyTorch, Hugging Face ecosystem).
- Familiarity with LLM post-training alignment (e.g., SFT, DPO, RLHF/RLVR) or context-extension, alongside comfort running distributed GPU training in Linux/HPC environments.
Company overview
AI Sweden is Sweden's national center for applied artificial intelligence, established in 2019 with a mission to accelerate the adoption and implementation of AI technologies for societal benefit, economic competitiveness, and national advancement. As a collaborative research institute, it brings together over 170 partners from diverse sectors including government agencies, universities, research institutions, startups, and global corporations. The organization operates as a catalyst for AI innovation, focusing on practical applications across healthcare, energy systems, municipalities, and various industries. With a core team of over 80 professionals comprising AI experts, data scientists, research engineers, linguists, mathematicians, policy experts, and change leaders, AI Sweden facilitates knowledge sharing, talent development, and infrastructure building. The institute operates on a not-for-profit basis, funded by Sweden's innovation agency Vinnova, the Swedish Agency for Economic and Regional Growth, European Regional Development Fund, Västra Götalandsregionen, and partner contributions. Through its AI Labs, talent programs, workshops, and collaborative projects, AI Sweden addresses critical societal challenges while bridging the gap between cutting-edge research and real-world implementation, positioning Sweden as a leader in responsible and impactful AI development.