persona-data 0.2.3

Shared dataset loading and prompt formatting for implicit-personalization projects

persona-data

Shared dataset loading, prompt formatting, and environment utilities for the implicit-personalization projects.

Overview

persona-data provides the common dataset and prompt helpers used across the persona projects:

• SynthPersonaDataset for persona profiles plus QA pairs
• PersonaGuessDataset for turn-based persona games
• NemotronPersonasFranceDataset for French persona profiles from NVIDIA
• NemotronPersonasUSADataset for US persona profiles from NVIDIA
• prompt helpers for roleplay and multiple-choice evaluation
• environment helpers for seeds, devices, and artifact paths

Installation

Add as a uv git source in your project's pyproject.toml:

[project]
dependencies = ["persona-data"]

[tool.uv.sources]
persona-data = { git = "ssh://[e-mail]/implicit-personalization/persona-data.git" }

Then run uv sync.

For local development alongside other repos, use an editable path source:

[tool.uv.sources]
persona-data = { path = "../persona-data", editable = true }

Testing

uv run --with pytest pytest tests/test_datasets.py

The release workflow also runs tests/smoke_test.py against the built wheel and source distribution.

Package layout

src/persona_data/
├── __init__.py
├── synth_persona.py       # SynthPersonaDataset, PersonaDataset, PersonaData, QAPair, BiographySection, Statement
├── persona_guess.py       # PersonaGuessDataset, GameRecord, Turn
├── nemotron_personas.py   # NemotronPersonasFranceDataset, NemotronPersonasUSADataset
├── prompts.py             # format_roleplay_prompt, system_prompt_for_variant, format_mc_question, format_messages
└── environment.py         # load_env, set_seed, get_device, get_artifacts_dir

Datasets

Each dataset is a module with its own types and a loader that downloads from Hugging Face, cached via HF_HOME.

SynthPersona

from persona_data.synth_persona import SynthPersonaDataset

dataset = SynthPersonaDataset()

persona = dataset[0]
persona.name              # "Ethan Robinson"
persona.templated_view    # short attribute-based system prompt
persona.biography_view    # full biography text
persona.sections          # list of BiographySection

qa_pairs = dataset.get_qa(persona.id, type="implicit", difficulty=[1, 2])
questions = dataset.questions(persona.id, type="explicit")

PersonaGuess

from persona_data.persona_guess import PersonaGuessDataset

games = PersonaGuessDataset()
game = games[0]
turns = games.get_qa(game.game_id, player="A")
questions = games.questions(game.game_id, player="B")

Prompt formatting

from persona_data.prompts import format_messages, format_roleplay_prompt

system_prompt = format_roleplay_prompt(persona.biography_view)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Where did you grow up?"},
    {"role": "assistant", "content": "I grew up in Little Rock, Arkansas."},
]
full_prompt, response_start_idx = format_messages(messages, tokenizer)

format_roleplay_prompt supports mode="roleplay" (default) and mode="conversational".

Use system_prompt_for_variant(persona, variant) when iterating over persona variants — it returns a persona-less prompt for "baseline" and reads <variant>_view otherwise. Downstream artifact code can use BASELINE_PERSONA_ID and BASELINE_PERSONA_NAME from persona_data.prompts for the shared baseline identity.

For multiple-choice prompts, use format_mc_question(qa) to render the question, choices, and trailing answer-only instruction. Use mc_answer_only_instruction(n_choices) if you need just the instruction text, and mc_correct_letter(qa) to get the gold label.

format_messages handles tokenizers that do not support the "system" role (for example Gemma 2) by merging system content into the first user message.

Environment helpers

from persona_data.environment import load_env, set_seed, get_device, get_artifacts_dir

load_env()            # loads .env from cwd (searches parent dirs)
set_seed(1337)        # sets random, numpy, and torch seeds
device = get_device() # cuda > mps > cpu

Used by

• persona-vectors — activation extraction and steering
• cues_attribution — section-level ablation attribution
• persona-2-lora — LoRA-based persona internalization