Governments publish high-quality institutional data and content (service portals, statistical definitions, numerical indicators), but Large Language Models (LLMs) training corpora are largely opaque. Consequently, it is unclear whether and to what extent these sources actually contribute to the behavior of the models, nor through what observable signals their impact can be rigorously assessed. In this thesis we study the role of French institutional sources as data providers for open-weight LLM by replicating and adapting a two-component evaluation framework to the French context. The first test is leakage-oriented and measures behavioral recall signals of (i) correct numerical values and (ii) verifiable provenance/attribution, in the absence of external retrieval and under controlled prompting.