NVIDIA published a walkthrough on 27 October of post-training Nemotron open models with its Nemotron-Personas datasets, built with French AI company Pleias; the Belgium set also involved KU Leuven. Both belong to an open data programme that now covers locale-grounded persona datasets for 12 countries. Each persona is synthetic but grounded in regional demographic, geographic, linguistic and cultural distributions - age, commune, job and household following census data - and is paired with official procedure text for the facts in an answer. The worked example is a French passport query from someone who cannot travel: the unadapted model advises sending the application by registered mail to the tax authority, which contradicts the actual rule that an official must visit the applicant's home. After post-training on persona-generated conversations, the reply matches the documented procedure and asks for a medical certificate under three months old and the tax stamp. Nvidia notes the stamp costs 43 euros in French Guiana against 86 euros in mainland France. Pleias uses three persona datasets, including France and Belgium, for a spine-care assistant called Lamina.
Our takeBe precise about what fails: the base model's French is fluent and its procedure is wrong. That is a grounding gap, not a language gap, so persona data helps mainly by generating the varied questions that decide which procedure applies to whom. The walkthrough shows one before-and-after example and no aggregate evaluation, so it is a template rather than evidence of general improvement. The risk to watch is that census-grounded personas reproduce documented demographics well and rare edge cases poorly - exactly the cases where a wrong procedure does the most damage.