I am giving two tutorials at the CEA-EDF-INRIA Computer Science School 2026, as part of the program on LLMs, software engineering, and formal methods.

The slides are available here:

Monday: Data Cleaning and Pretraining

This session follows the life of a model before it is “ready”: collecting data at scale, cleaning and filtering it, building diverse training mixtures, and using self-supervised learning for pretraining. We will also look at why source code is such a useful data modality for strengthening logical reasoning and structured problem solving in LLMs.

Tuesday: LLMs and AI Safety

This session focuses on the ethical and technical guardrails needed to deploy LLMs responsibly. We will discuss alignment, bias mitigation, red teaming, and practical ways to evaluate the reliability and safety of model outputs, especially in sensitive settings such as software development.