NVIDIA publishes recipe for fine-tuning Nemotron 3.5 Lightning on French public-service procedures with synthetic personas

The workflow samples about 2,000 personas from the 1M-record Nemotron-Personas-France set. Each persona is paired with official procedure text so NeMo Data Designer can generate customer and agent exchanges. A second model judges the replies, and an audit step removes leftover reasoning traces and leaked names. The data is split 90/10 by persona so no persona appears in both training and validation. A pre-tuned LoRA recipe then trains the model, and the authors advise changing only the data paths and training schedule. [1] [2] [8] [9]

The persona release is presented as one step in a wider series. Earlier releases covered the US and Korea, Pleias built a Belgian set, and NVIDIA says the program now spans 12 countries. The linked video says the same loop carries over to another locale by swapping the personas, the behavior contract and the reviewers. [1] [2] [4] [5] [6]

The headline before-and-after case shows the base model telling a person too ill to travel to apply for a passport by registered post. The official Service-Public page lists no postal route and requires an in-person visit for fingerprints. That page does not state the home-visit exception the tuned model gives. [1] [2] [3]

The repository calls the example a pattern rather than a production assistant. It warns that judge-based filtering can err in both directions and urges review by native speakers. For evaluation it recommends a frozen holdout of 100 to 300 human-written cases, scored on correctness, French quality, locale fit, safety, robustness and retention. [1] [9]