NVIDIA releases Nemotron-Personas for France and Belgium, with a fine-tuning workflow for Nemotron 3.5 Lightning

NVIDIA has released Nemotron-Personas datasets for France and Belgium, along with a post-training workflow that fine-tunes Nemotron 3.5 Lightning on local procedures and cultural context. NVIDIA published the details on its developer blog on October 27, 2026 (authors: Sophia Abbassi and Shashank Verma).

The datasets

NVIDIA built Nemotron-Personas-France with French AI startup Pleias. Nemotron-Personas-Belgium was built with Pleias and Belgian university KU Leuven. The Belgium dataset spans Dutch, French, German and English. Both belong to the open NVIDIA Nemotron data program, which includes locale-grounded persona datasets for 12 countries.

Each persona is a synthetic resident grounded in regional demographic, geographic, linguistic and cultural distributions. According to NVIDIA, a grounded persona has an age, a commune, a job and a household that follow French census data, plus written profiles of the person's work and daily life. A generic persona, by contrast, reads like "a French adult who needs a passport."

The example: a passport query

NVIDIA's illustration is a French passport-agency agent. A user writes that they can no longer travel because of their health and asks how to apply. According to NVIDIA, a foundation model without post-training suggested applying by post, which conflicts with official documentation: a government official must make a home visit, and mail submission is not an option.

The fine-tuned model, in NVIDIA's example, says an agent can come to the home, requires a medical certificate less than 3 months old, and lists the documents to bring. Local detail can also change the correct answer: for a resident of French Guiana, the passport stamp costs €43, against €86 in mainland France.

The workflow

NVIDIA describes a pipeline that pairs each persona with official procedure text, which supplies the facts in each answer. The steps:

• Define a behavior contract and a holdout set.
• Use the personas, the contract and the procedure text to generate synthetic conversations.
• Fine-tune Nemotron 3.5 Lightning with LoRA on NVIDIA NeMo.
• Merge the adapter.
• Evaluate against a frozen test set.

The Nemotron repository supplies reusable steps for data generation, training and evaluation, plus agent skills that help coding agents assemble the workflow. NVIDIA says the same loop applies to any Nemotron-Personas market: swap in another persona dataset, update the behavior contract and reviewers, and the rest stays the same. NVIDIA also points to a sizing calculator for estimating GPU needs, training time and costs, and to open Nemotron models and datasets on Hugging Face.

NVIDIA recommends the approach for models that already have some capability in the target language and need application-specific behavior, such as terminology, formality, response structure or escalation rules.

Pleias and Lamina

Pleias uses three Nemotron-Personas datasets, including France and Belgium, in its work on Lamina, a spine-care assistant from a Pleias partner. The personas turn a few real patient cases into varied training conversations and let Pleias test the assistant across patient profiles. Pleias says: "Better grounded data is what makes a small efficient regional-specific model viable."

NVIDIA states that persona grounding improves coverage and reduces blind spots in training data, but is not a guarantee of fairness or proper representation.

What this means

The workflow separates two problems: who is asking, which the personas cover, and what the correct answer is, which comes from official procedure text. The passport example shows why that matters: a fluent French answer can still be wrong on procedure. For teams building agents for government or healthcare-adjacent use, the evaluation step against a frozen test set is where that kind of error would surface. NVIDIA's own caveat is worth keeping in view: persona grounding widens coverage but does not certify fairness.