Skip to content

Nemotron Personas

Persona-only loaders for the NVIDIA Nemotron Personas datasets (no QA pairs).

Class Source
NemotronPersonasFranceDataset nvidia/Nemotron-Personas-France
NemotronPersonasUSADataset nvidia/Nemotron-Personas-USA 
from persona_data.nemotron_personas import (
    NemotronPersonasFranceDataset,
    NemotronPersonasUSADataset,
)

dataset = NemotronPersonasFranceDataset(sample_size=200)
usa_dataset = NemotronPersonasUSADataset(sample_size=200)

The loader discovers every data/train-*.parquet shard in sorted order and reads rows until it has collected sample_size personas (a leading slice, not a random sample). first_name and last_name are derived from the persona text when possible.

Records use the shared PersonaData type, exposing id, persona, templated_view, biography_view, and the derived name.

persona = dataset[0]

persona.id              # source uuid
persona.name            # derived display name, or uuid as fallback
persona.biography_view  # original persona text
persona.templated_view  # normalized name, demographics, location, and bio fields
dataset.get_persona(persona.id)
dataset.supports_qa     # False

These datasets are persona-only; they do not expose get_qa() or QA splits.

Templated view differences

  • France: location formatted as commune, departement, country; trailing Household type field.
  • USA: location formatted as city, state, zipcode, country; Bachelors field inserted before Marital status.