Nvidia shows how to teach Nemotron French passport rules with synthetic residents

Nvidia has published a hands-on guide to adapting its open Nemotron 3.5 Lightning model to French public-service questions, using synthetic residents from the new Nemotron-Personas-France and Belgium datasets.

A passport question that exposes the gap

The post opens with a test case: a French user who can no longer travel for health reasons asks how to apply for a passport. According to Nvidia, a base model with no post-training suggested applying by registered mail, a route the official documentation does not support. The authors say a government official must instead make a home visit. The fine-tuned model, in Nvidia’s example, gives that exception and asks for a medical certificate less than three months old.

Nvidia also notes that local facts can change the right answer. A passport stamp costs 43 euros in French Guiana against 86 euros in mainland France.

The workflow, step by step

The datasets were built with French startup Pleias, and the Belgian one also with the university KU Leuven. Each persona is a synthetic resident whose age, commune, job and household follow regional statistics. They are part of a Nemotron data program that Nvidia says covers locale-grounded personas for 12 countries.

The recipe has five stages:

• Write a behavior contract (audience, tone, allowed facts, safety rules, answer structure) and freeze a holdout set. Nvidia suggests 100 to 300 human-written or de-identified cases for production work.
• Sample 2,000 personas and use the sdg/data_designer step to generate a customer message and an agent reply for each, grounded in official procedure text. Nvidia says to set enable_thinking: false under providers, extra_body, chat_template_kwargs. Otherwise reasoning text lands in the training data.
• Audit the output with scripts/audit_and_split.py, which drops malformed rows, leftover reasoning and leaked persona names, then splits 90/10 by persona.
• Fine-tune with LoRA using the NeMo Automodel recipe for Lightning, then merge the adapter into the base checkpoint.
• Serve base and tuned models side by side and run compare_models.py, which writes eval/comparison.md for human review.

The hardware listed is one DGX Station GB300 or one DGX Spark. On the Spark, Nvidia advises a smaller batch size and shorter sequences, and a hosted API for data generation. The code lives in the Nemotron repository at tag v0.2.0, under use-case-examples/french-public-service-assistant. Python 3.11 to 3.13, Docker and the NVIDIA Container Toolkit are required.

Pleias, per the post, uses three persona datasets in Lamina, a spine-care assistant from one of its partners, to stretch a few real patient cases into varied conversations. In Pleias’s words:

Better grounded data is what makes a small efficient regional-specific model viable.

What changes for teams building local assistants

The guide is aimed at developers who already have a model that speaks the target language and need it to follow application rules: terminology, formality, escalation. For them the practical shift is where the facts come from. Instead of hoping a general model remembers local procedure, the training pairs each persona with official text and checks the answers against it.

The cost side is helped by the method itself. In our explainer on LoRA, the original 2021 paper claimed a 10,000-fold cut in trainable parameters and about three times less GPU memory than full fine-tuning of GPT-3 175B, with no added inference delay once the adapter is merged. That is the same merge step Nvidia describes here. The same article notes an Nvidia video model where LoRA touched about 0.16 percent of roughly 373 million parameters.

The data also fits the wider push for national control over AI. Our sovereign AI explainer describes a 2026 Brookings framework that counts data and models among eight layers a country can try to control, and notes that Nvidia, which sells the hardware, has a commercial interest in that framing. Open persona datasets for 12 markets give developers one cheap piece of the data layer.

What the post does not establish

The before and after answers are single examples chosen by Nvidia. The post gives no benchmark scores, no error rates and no count of how many test prompts the tuned model passed, so the size of the improvement is unknown. Its own evaluation table asks for human review and zero critical failures on a frozen safety set, which implies the results still need checking by each team.

Nvidia is also careful about limits. Persona grounding, it says, reduces blind spots in the data but is not a guarantee of fairness or proper representation, and it advises using persona fields only to vary tone and phrasing, not to guess a customer’s needs. It also recommends stripping names and unneeded demographic detail from the final training set.

For next steps, Nvidia points to NeMo RL and NeMo Gym for reinforcement learning and multi-node training, and to Nemotron 3.5 ASR, a 600-million-parameter streaming speech model, for voice assistants. The datasets and models are on Hugging Face.