# NVIDIA Outlines Nemotron Fine-Tuning for French Public Services By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/42307 Published: 2026-10-28T03:40:38.000Z Updated: 2026-10-28T03:40:38.000Z Section: Technology > The guide configures 2,000 synthetic resident profiles and 2,000 training records, but reports no aggregate benchmark result. NVIDIA published a developer guide Tuesday, Oct. 27, showing how to adapt Nemotron 3.5 Lightning to French public-service questions with simulated resident profiles and official procedure text. The French walkthrough covers identity documents and vehicle registration. It uses Nemotron-Personas-France, developed by NVIDIA with French AI startup Pleias. A Belgium dataset developed with Pleias and KU Leuven is offered as an alternative, but the guide does not demonstrate a Belgian public-service assistant. The synthetic profiles are designed around regional population, geography, language and culture. Developers also set a behavior contract that defines the assistant’s audience, safety boundaries, response format and approved reference material. Together, those elements are meant to vary how residents phrase questions while keeping answers tied to supplied procedures. The recipe configures a sample of 2,000 profiles and 2,000 generated training records. Those figures describe tutorial settings; they do not establish that a production training run was completed. The workflow then calls for auditing the records, fine-tuning with LoRA, merging the adapter and comparing the adapted model with the base model on the same fixed prompts. One passport example contrasts a base-model answer that suggests applying by mail with an adapted answer describing a home visit and a medical certificate issued less than three months earlier. French rules allow a home visit for applicants unable to travel because of a medical reason or infirmity, with that certificate requirement. The guide recommends human review of model comparisons and says lower training loss alone does not prove better results for the task. It provides no aggregate benchmark score. Persona data can broaden the range of questions represented, but does not by itself ensure fair or accurate representation. NVIDIA has also published a [fine-tuning recipe for telecom workflows](). ## Links in this article - [fine-tuning recipe for telecom workflows](https://www.tokenpost.com/news/technology/27879)