AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Fine-Tuning NVIDIA Nemotron to Know Local Rules, Not Just Local Languages

Collected Oct 27, 2026

NVIDIA has published a walkthrough for adapting its open Nemotron models to regional rules and customs rather than just regional languages, built around two new locale-grounded datasets and a complete example project. The company created Nemotron-Personas-France with the French startup Pleias, and Nemotron-Personas-Belgium with Pleias and KU Leuven. Both sit inside the open Nemotron data program, which NVIDIA says now covers persona datasets for 12 countries. The worked example ships in the Nemotron code repository, tagged v0.2.0, as a use case called french-public-service-assistant.

The motivating case is a passport question in French from someone who can no longer travel for health reasons. NVIDIA shows a base model answering that the applicant should mail the file by registered letter to the public finances department — advice the company says is unsupported and contradicts government documentation, since an official must visit the home instead. In the fine-tuned version of the same interaction, the answer explains that a home visit can be requested with a medical certificate less than three months old, points to the ANTS website or a town hall with a registration station, lists the original documents to bring, and notes the tax stamp fee.

The mechanism is synthetic data generation steered by personas. Each persona is a synthetic resident sampled to match a region's demographic, geographic, linguistic and cultural distributions: a grounded persona carries an age, a commune, a job and a household that follow census data, plus written profiles of work and daily life. A general instruction is replaced by someone specific, which changes both how a question is phrased and sometimes what the correct answer is. NVIDIA's own illustration: the passport tax stamp costs €43 for a resident of French Guiana versus €86 in mainland France.

The example workflow runs in five stages. First, write a behavior contract describing what a good reply must do; the tutorial recommends freezing a holdout set of 100 to 300 human-written or de-identified production cases at this point, though it is optional for the exercise. Sample test cases include local facts (stating the €43 Guyane stamp and not inventing a phone or kiosk route), out-of-scope questions (refusing to confirm town hall Saturday hours and pointing to service-public.gouv.fr), and safety scenarios (a caller claiming to be from the administration asking the user to install software).

Second, generate conversations. A script samples 2,000 personas spread across departments and occupations, and a data-designer step turns each into a customer message and an agent reply using a generator model. Persona fields are meant to vary voice, not to infer needs or competence from sensitive attributes, and names and unnecessary demographic details are stripped from the final training data. Each answer is checked by a separate judge model, which calls the NVIDIA-hosted API by default. The guide advises previewing eight records against the behavior contract before a full run, and starting with 500 to 1,000 records on a DGX Spark or any hosted endpoint.

A configuration pitfall is called out explicitly: Nemotron 3.5 Lightning emits reasoning traces by default, and the enable_thinking: false switch must live under providers, extra_body, chat_template_kwargs. Placed under inference_parameters the run fails; omitted entirely, reasoning text contaminates the training data. Other listed settings are temperature 0.8, max_tokens 2048 and 64 parallel requests — tightened to 1024 tokens and 5 parallel requests on a Spark — with the generator served through vLLM (image v0.26.0) using the model nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16.

Third, an audit script drops rows with formatting problems, leftover reasoning or leaked persona names, then splits the remainder 90/10 by source_uuid so no persona appears in both train and validation sets. Fourth, fine-tune with LoRA starting from the validated NeMo Automodel recipe for Lightning, changing data paths and the training schedule while leaving architecture-tuned model and LoRA settings alone; on a Spark the guide suggests local_batch_size 1, seq_length 8192 and a 16 GB shared-memory allocation in the AutoModel container (nvcr.io/nvidia/nemo-automodel:26.08). Flat or rising validation loss from the start is a signal to inspect the data, not the model.

Fifth, merge the adapter into the base checkpoint with the convert/merge_lora step, serve the result as lightning-fr on port 8001 alongside the base model on 8000, and evaluate. NVIDIA stresses that lower loss is not proof: the same frozen prompts go to both models and answers are compared side by side, with human review recommended over an automated judge. Reviewers look for hallucinations such as invented fees, deadlines or documents, and for replies that break character by announcing they are a language model. On a DGX Station both servers can run together at 0.45 GPU memory utilization each; on a Spark only one model is served at a time and the comparison script is run twice.

Prerequisites are Python 3.11 to 3.13, uv, Git, Docker, the NVIDIA Container Toolkit, and either a DGX Station GB300 or a DGX Spark, though NVIDIA says the workflow adapts to other GPU systems. Separately, for voice front ends the post points to Nemotron 3.5 ASR, a 600-million-parameter streaming recogniser with 40 language locales — 32 transcribable out of the box and eight requiring fine-tuning — plus a speech fine-tuning recipe.

This suggests the personas are mainly a coverage tool: they multiply scarce real examples into varied conversations and expose a model to edge cases a small hand-built set would miss. NVIDIA is careful to say persona grounding improves coverage and reduces blind spots rather than guaranteeing fairness or representation, and the quality of every answer still depends on the official procedure text paired with each persona — garbage procedures in, confident wrong answers out. A likely trade-off is that heavy LoRA tuning on generated data can sharpen tone and format while making the model more brittle outside its target domain, which is why the out-of-scope and safety test categories matter as much as the local-fact ones.

Why it matters: teams building public-service, healthcare or support agents in a specific country can copy this loop — behavior contract, personas, synthetic generation, LoRA, human comparison — instead of waiting for a bigger general model to absorb local detail. The persona datasets for 12 markets lower the data-collection cost, though the approach still needs authoritative source text and human reviewers per locale, so the real budget shifts from model training to domain curation. Who is affected: product teams shipping regional assistants, and developers who now have a concrete recipe and tooling to adapt an open model rather than prompt around its gaps.

Read at NVIDIA Developer Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

As AI adoption expands globally, people across languages and cultures are using it in more aspects of their daily lives. General-purpose language models serve...